umbral_core/migrate.rs
1//! The migration engine — the north star.
2//!
3//! Implements the **declare → migrate → change → migrate** cycle from
4//! `arch.md §0`. Users declare or change a model, run `migrate`, and the
5//! framework either generates the missing migration file (via `make`)
6//! or applies pending migration files (via `run`).
7//!
8//! At M5 (this milestone) the surface ships:
9//!
10//! - A process-wide [`ModelRegistry`] populated by
11//! `App::builder().model::<T>()`.
12//! - A [`Snapshot`] of every registered model's metadata, JSON-
13//! serialisable so it can live inside a migration file's
14//! `snapshot_after`.
15//! - An [`Operation`] enum with the two minimum-viable ops:
16//! [`Operation::CreateTable`] and [`Operation::DropTable`]. Column-
17//! level ops (`AddColumn`, `DropColumn`, `AlterColumn`) land at M8
18//! alongside rename detection (per `arch.md §7` and
19//! `docs/specs/06-migration-engine.md`). The "M5.1" label in the
20//! `UnsupportedChange` error message is shorthand for the same slot.
21//! - A [`MigrationFile`] format (one JSON file per migration) carrying
22//! `id`, `operations`, and `snapshot_after`.
23//! - The `umbral_migrations` tracking table (one row per applied
24//! migration, keyed by `(plugin, name)`).
25//! - High-level entry points: [`make`], [`run`], [`show`].
26//!
27//! Reserved for M5.1+:
28//!
29//! - Column-level ops (`AddColumn`, `DropColumn`, `AlterColumn`).
30//! - Rename-detection vs drop+add disambiguation (spec 06 §M8).
31//! - `RunSql` / `RunCode` data-migration ops.
32//! - Squashing, `--fake`, `--fake-initial` (PRD F-MIG-6 P2).
33//! - Cross-plugin migration dependencies (needs M7 Plugin contract).
34//!
35//! See `docs/specs/06-migration-engine.md` for the full target shape.
36
37use std::path::{Path, PathBuf};
38use std::sync::OnceLock;
39
40use serde::{Deserialize, Serialize};
41
42use crate::backend::DatabaseBackend;
43use crate::orm::{FieldSpec, Model, SqlType};
44
45/// Per-process model registry. Published by `AppBuilder::build()`
46/// after `.model::<T>()` calls and `.plugin(...)` registrations
47/// collected metadata into the builder.
48///
49/// Stored as a flat vector of `(plugin_name, model)` pairs so M5's
50/// existing `registered_models()` keeps working (drop the plugin
51/// names) and the M7 plugin-aware walks (`registered_plugins`,
52/// `models_for_plugin`) can read the same source of truth without a
53/// second registry. The plugin name `"app"` covers models registered
54/// via `.model::<T>()`; every other name is a real Plugin's.
55static REGISTRY: OnceLock<Vec<(String, ModelMeta)>> = OnceLock::new();
56
57/// Initialize the registry with one entry per plugin.
58///
59/// `App::build()` calls this after collecting `.model::<T>()` into the
60/// implicit `"app"` plugin and walking every registered plugin's
61/// `Plugin::models()`. Plugins missing from the map contribute zero
62/// models (default-noop `models()` returns an empty vec; the entry
63/// can be omitted).
64pub(crate) fn init_plugins(per_plugin: std::collections::HashMap<String, Vec<ModelMeta>>) {
65 let mut flat: Vec<(String, ModelMeta)> = Vec::new();
66 let mut plugin_names: Vec<String> = per_plugin.keys().cloned().collect();
67 plugin_names.sort();
68 for plugin in plugin_names {
69 for m in per_plugin.get(&plugin).cloned().unwrap_or_default() {
70 flat.push((plugin.clone(), m));
71 }
72 }
73 REGISTRY
74 .set(flat)
75 .expect("umbral::migrate::init_plugins called more than once");
76}
77
78/// Return every registered model, flat. Drops the per-plugin grouping;
79/// useful when the caller only needs the model set (e.g. M5's `make`
80/// when the codebase only had a single `"app"` plugin).
81///
82/// # Panics
83///
84/// Panics if `App::build()` hasn't run.
85pub fn registered_models() -> Vec<ModelMeta> {
86 registered_models_opt()
87 .expect("umbral: model registry not initialised — did you call App::build()?")
88}
89
90/// [`registered_models`] without the panic: `None` when the registry hasn't been
91/// initialised yet.
92///
93/// For callers that legitimately run before (or without) `App::build()` — the
94/// boot system checks are the case in point, and a check that panics is a check
95/// that can never report anything.
96pub fn registered_models_opt() -> Option<Vec<ModelMeta>> {
97 Some(REGISTRY.get()?.iter().map(|(_, m)| m.clone()).collect())
98}
99
100/// Whether the model registry has been initialised. False before
101/// `App::build()` has run; true after the phase-3 `init_plugins`
102/// call publishes the per-plugin map. Used by system checks that
103/// walk the registry — they return an empty result when the
104/// registry isn't ready rather than panicking (so low-level tests
105/// that drive `check::run_all` without booting an App keep working).
106pub fn is_initialised() -> bool {
107 REGISTRY.get().is_some()
108}
109
110/// PK lift Pass E — cached `(pk_column_name, pk_sql_type)` lookup
111/// keyed by table name. Used by the FK decode path
112/// (`fk_target_pk_sql_type` in `orm/dynamic.rs`) and the
113/// select_related hydrators, both of which previously cloned the
114/// full `Vec<ModelMeta>` per call and linear-scanned for the
115/// target's PK column.
116///
117/// REGISTRY is a `OnceLock` set once during `App::build`; this cache
118/// reads from it the first time anyone asks for a PK lookup AFTER
119/// initialisation, then serves from a `HashMap` for every
120/// subsequent call. Eliminates the per-row `registered_models()`
121/// clone in hot decode loops.
122///
123/// Returns `None` when the registry isn't initialised (the cache
124/// stays uninstantiated so a follow-up call after `App::build`
125/// gets the real table set), OR when the named table isn't in the
126/// registry (orphan / system / typo).
127pub fn pk_meta_for_table(table: &str) -> Option<(String, crate::orm::SqlType)> {
128 if !is_initialised() {
129 // Defer cache init until App::build has populated REGISTRY.
130 // The cache MUST NOT memoize an empty map; otherwise
131 // post-init callers would see no PK metadata forever.
132 return None;
133 }
134 static CACHE: std::sync::OnceLock<
135 std::collections::HashMap<String, (String, crate::orm::SqlType)>,
136 > = std::sync::OnceLock::new();
137 let map = CACHE.get_or_init(|| {
138 let mut out = std::collections::HashMap::new();
139 for m in registered_models() {
140 if let Some(pk) = m.pk_column() {
141 out.insert(m.table.clone(), (pk.name.clone(), pk.ty));
142 }
143 }
144 out
145 });
146 map.get(table).cloned()
147}
148
149/// Cached model lookup by SQL table name.
150///
151/// Unlike [`registered_models`], this does not deep-clone the full
152/// registry on every call. It clones only the matched [`ModelMeta`],
153/// which keeps row-by-row dynamic serializers from paying
154/// O(registry-size) per row.
155pub fn model_meta_for_table(table: &str) -> Option<ModelMeta> {
156 if !is_initialised() {
157 return None;
158 }
159 static CACHE: std::sync::OnceLock<std::collections::HashMap<String, ModelMeta>> =
160 std::sync::OnceLock::new();
161 let map = CACHE.get_or_init(|| {
162 registered_models()
163 .into_iter()
164 .map(|m| (m.table.clone(), m))
165 .collect()
166 });
167 map.get(table).cloned()
168}
169
170/// The SQL type a column's value actually binds / decodes as (PK lift).
171/// Equals `col.ty` for everything except a `ForeignKey`, where it resolves
172/// to the referenced model's PK type via [`pk_meta_for_table`] — so an FK
173/// pointing at a `String`-slug- or `Uuid`-PK target is handled as text /
174/// uuid instead of being forced through i64. Falls back to `BigInt` (the
175/// historical default) when the target can't be resolved (registry not yet
176/// initialised, or an unregistered target table).
177///
178/// The single source of truth for "what shape is this FK really?", used by
179/// `backup` (dump/load) and the dynamic filter helpers.
180pub fn fk_effective_type(col: &Column) -> crate::orm::SqlType {
181 if matches!(col.ty, crate::orm::SqlType::ForeignKey) {
182 col.fk_target
183 .as_deref()
184 .and_then(pk_meta_for_table)
185 .map(|(_, ty)| ty)
186 .unwrap_or(crate::orm::SqlType::BigInt)
187 } else {
188 col.ty
189 }
190}
191
192/// Return the registered plugin names that contributed at least one
193/// model. Sorted deterministically. Used as a fallback when no
194/// topological order is published; the M7 walk used this directly,
195/// and M8 prefers [`plugin_order`] when it's been set.
196pub fn registered_plugins() -> Vec<String> {
197 let mut names: Vec<String> = REGISTRY
198 .get()
199 .expect("umbral: model registry not initialised — did you call App::build()?")
200 .iter()
201 .map(|(p, _)| p.clone())
202 .collect();
203 names.sort();
204 names.dedup();
205 names
206}
207
208/// The topological plugin order published by `App::build()` after its
209/// phase 1.5 sort. `None` until that runs; the CLI subcommands
210/// (`makemigrations`, `migrate`, `showmigrations`) call `App::build()`
211/// via `boot_for_management` before reaching the migration engine.
212static PLUGIN_ORDER: OnceLock<Vec<String>> = OnceLock::new();
213
214/// Per-model database alias (`Model::NAME -> alias`) published by
215/// `App::build()` after walking each registered plugin's
216/// `Plugin::database()`. Models whose plugin returned `None` are
217/// absent from the map; QuerySet's `resolve_pool` falls back to the
218/// `"default"` alias for those. Lookup is `O(1)` on a `HashMap`.
219static MODEL_ALIASES: OnceLock<std::collections::HashMap<String, String>> = OnceLock::new();
220
221/// Publish the topological plugin order. Called by `App::build()` once
222/// the phase 1.5 sort has produced the order. Must include the
223/// implicit `"app"` plugin even when no real plugins are registered.
224pub(crate) fn init_plugin_order(order: Vec<String>) {
225 PLUGIN_ORDER
226 .set(order)
227 .expect("umbral::migrate::init_plugin_order called more than once");
228}
229
230/// Return the topological plugin order if `App::build()` published
231/// one; otherwise fall back to [`registered_plugins`] (sorted by
232/// name). The fallback keeps existing M5 / M6 tests working without
233/// requiring them to wire a full plugin sort.
234pub fn plugin_order() -> Vec<String> {
235 PLUGIN_ORDER
236 .get()
237 .cloned()
238 .unwrap_or_else(registered_plugins)
239}
240
241/// The client-facing API endpoints every registered plugin advertised
242/// via `Plugin::api_endpoints()`, collected by `App::build()`. `None`
243/// until that runs; an app with no advertising plugins publishes an
244/// empty vec.
245static API_ENDPOINTS: OnceLock<Vec<crate::plugin::ApiEndpoint>> = OnceLock::new();
246
247/// Publish the collected `Plugin::api_endpoints()`. Called once by
248/// `App::build()` after walking every registered plugin.
249pub(crate) fn init_api_endpoints(endpoints: Vec<crate::plugin::ApiEndpoint>) {
250 let _ = API_ENDPOINTS.set(endpoints);
251}
252
253/// Every callable endpoint registered plugins advertised for service
254/// discovery, in plugin-registration order. Empty until `App::build()`
255/// has run. A REST API root (or any discovery surface) reads this to
256/// list plugin endpoints without depending on those plugins' crates.
257pub fn registered_api_endpoints() -> Vec<crate::plugin::ApiEndpoint> {
258 API_ENDPOINTS.get().cloned().unwrap_or_default()
259}
260
261/// Publish the per-model alias routing. Called by `App::build()`
262/// during phase 3 after walking every plugin's `Plugin::database()`.
263/// Plugins that returned `None` contribute no entries; only the
264/// explicit overrides land here.
265pub(crate) fn init_model_aliases(map: std::collections::HashMap<String, String>) {
266 MODEL_ALIASES
267 .set(map)
268 .expect("umbral::migrate::init_model_aliases called more than once");
269}
270
271/// Look up the database alias for a SQL table name — the reverse of
272/// the `Model::NAME → alias` lookup that [`model_alias`] does. Walks
273/// the registered model metas to find the one whose `table` matches
274/// (snake_case of the struct name + any `#[umbral(table = "...")]`
275/// override) and returns its alias if set. Falls back to `"default"`
276/// when no model owns the table (e.g. orphan schema, the
277/// `umbral_migrations` table itself) — those land on the main pool.
278///
279/// Used by the migration engine's per-DB dispatch in [`run_in`] to
280/// route each operation to the right pool.
281pub fn table_alias(table_name: &str) -> String {
282 for meta in registered_models() {
283 if meta.table == table_name {
284 return model_alias(&meta.name).unwrap_or_else(|| "default".to_string());
285 }
286 }
287 "default".to_string()
288}
289
290/// Look up the database alias for one model. Returns `None` if the
291/// model isn't routed explicitly (the caller falls back to the
292/// `"default"` pool); returns `None` even when the alias map hasn't
293/// been initialised so low-level tests that drive `init_plugins`
294/// directly don't have to wire a second call.
295pub fn model_alias(model_name: &str) -> Option<String> {
296 MODEL_ALIASES.get()?.get(model_name).cloned()
297}
298
299static MODEL_META_BY_NAME: OnceLock<std::collections::HashMap<String, ModelMeta>> = OnceLock::new();
300
301/// Cached `&ModelMeta` lookup by model name. Returns `None` before
302/// `App::build` populates the registry (low-level tests), which the routing
303/// seam treats as "fall back to legacy static routing".
304pub fn model_meta_ref(name: &str) -> Option<&'static ModelMeta> {
305 if !is_initialised() {
306 return None;
307 }
308 MODEL_META_BY_NAME
309 .get_or_init(|| {
310 registered_models()
311 .into_iter()
312 .map(|m| (m.name.clone(), m))
313 .collect()
314 })
315 .get(name)
316}
317
318/// Return the models registered against a specific plugin. Empty if
319/// no plugin by that name registered models.
320pub fn models_for_plugin(plugin: &str) -> Vec<ModelMeta> {
321 REGISTRY
322 .get()
323 .expect("umbral: model registry not initialised — did you call App::build()?")
324 .iter()
325 .filter(|(p, _)| p == plugin)
326 .map(|(_, m)| m.clone())
327 .collect()
328}
329
330/// Static metadata for one registered model, copied off the `Model`
331/// trait's `const`s when the user calls `App::builder().model::<T>()`.
332///
333/// Owned (no lifetimes) so the registry can hold an arbitrary number
334/// without the lifetime contortions a slice of trait references would
335/// need. The cost is one Vec at `App::build` time; the win is
336/// `registered_models()` having a plain `&'static [ModelMeta]` signature.
337#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
338pub struct ModelMeta {
339 /// The struct name (`Model::NAME`). Identifies the model across
340 /// snapshot diffs even if the table is renamed.
341 pub name: String,
342 /// The SQL table name (`Model::TABLE`).
343 pub table: String,
344 /// One owned column descriptor per field, in declaration order.
345 /// Owned (`Column`, not the underlying static `FieldSpec`) so the
346 /// snapshot round-trips cleanly through serde.
347 pub fields: Vec<Column>,
348 /// Human-readable display name from `Model::DISPLAY`. Defaults to
349 /// `Model::NAME` when no `#[umbral(display = "...")]` is present.
350 #[serde(default)]
351 pub display: String,
352 /// Lucide icon slug from `Model::ICON`. Defaults to `"database"`.
353 #[serde(default = "default_icon")]
354 pub icon: String,
355 /// Database alias from `Model::DATABASE`, when set. `None` means
356 /// "fall back to the owning plugin's `Plugin::database()`, then
357 /// the `default` pool." Captured here so `App::build`'s alias
358 /// routing can read it without re-reaching into the trait at a
359 /// later phase.
360 #[serde(default, skip_serializing_if = "Option::is_none")]
361 pub database: Option<String>,
362 /// Mirrors `Model::SINGLETON`. Closes BUG-9 in
363 /// `bugs/tests/testBugs.md`. Default `false`; admin renderers
364 /// read it to auto-redirect list-view to the edit form.
365 #[serde(default, skip_serializing_if = "is_false")]
366 pub singleton: bool,
367 /// Mirrors `Model::UNIQUE_TOGETHER`. Composite-UNIQUE constraints,
368 /// each inner `Vec<String>` listing the columns of one constraint.
369 /// Closes BUG-6.
370 #[serde(default, skip_serializing_if = "Vec::is_empty")]
371 pub unique_together: Vec<Vec<String>>,
372 /// Mirrors `Model::INDEXES`. Each inner `Vec<String>` lists the
373 /// columns of one multi-column index. Closes BUG-7.
374 #[serde(default, skip_serializing_if = "Vec::is_empty")]
375 pub indexes: Vec<Vec<String>>,
376 /// Mirrors `Model::ORDERING`. Each tuple is `(column, descending)`
377 /// — `descending == true` lowers to `ORDER BY col DESC`. Closes
378 /// BUG-8.
379 #[serde(default, skip_serializing_if = "Vec::is_empty")]
380 pub ordering: Vec<(String, bool)>,
381 /// Mirrors `Model::M2M_RELATIONS`. Many-to-many relations declared
382 /// on this model. The migration engine uses this to auto-generate
383 /// junction tables. Closes BUG-16.
384 #[serde(default, skip_serializing_if = "Vec::is_empty")]
385 pub m2m_relations: Vec<M2MRelation>,
386 /// Mirrors `Model::SOFT_DELETE` (`#[umbral(soft_delete)]`). The
387 /// dynamic / annotate paths read this to auto-exclude
388 /// `deleted_at IS NULL` children from correlated counts and to
389 /// drive trash-aware admin views without re-reaching into the
390 /// typed trait. Shared enabler for gaps2 #35 + #39a.
391 #[serde(default, skip_serializing_if = "is_false")]
392 pub soft_delete: bool,
393 /// `#[umbral(audited)]`. Lives on `ModelMeta` (not re-derived per path) so
394 /// the DYNAMIC write path — which is what admin and REST run on — reads the
395 /// same source of truth the typed one does.
396 #[serde(default, skip_serializing_if = "std::ops::Not::not")]
397 pub audited: bool,
398 /// features #73 — `#[umbral(view = "...")]`. `Some(sql)` means this model is
399 /// backed by a database VIEW whose body is `sql`; the migration engine emits
400 /// `CREATE VIEW` for it and never `CREATE TABLE`. Carried on `ModelMeta` (not
401 /// just the trait) because the *snapshot* is what `makemigrations` diffs — a
402 /// view whose SQL changed has to be detectable without the model being in scope.
403 #[serde(default, skip_serializing_if = "Option::is_none")]
404 pub view: Option<String>,
405 /// features #73 — the view is MATERIALIZED (Postgres-only). Meaningless unless
406 /// `view` is `Some`.
407 #[serde(default, skip_serializing_if = "is_false")]
408 pub materialized: bool,
409 /// The app label (the owning plugin's name), mirrors `Model::APP_LABEL`.
410 /// Sourced from `#[umbral(plugin = "...")]`; `"app"` when absent.
411 /// Authoritative for permission codenames (gaps2 #80g): replaces the
412 /// old table-name-split heuristic that collided distinct models. The
413 /// `#[serde(default)]` keeps pre-#80g snapshot JSON round-tripping.
414 #[serde(default = "default_app_label")]
415 pub app_label: String,
416}
417
418fn default_app_label() -> String {
419 "app".to_string()
420}
421
422impl Default for ModelMeta {
423 fn default() -> Self {
424 Self {
425 name: String::new(),
426 table: String::new(),
427 fields: Vec::new(),
428 display: String::new(),
429 icon: default_icon(),
430 database: None,
431 singleton: false,
432 unique_together: Vec::new(),
433 indexes: Vec::new(),
434 ordering: Vec::new(),
435 m2m_relations: Vec::new(),
436 soft_delete: false,
437 audited: false,
438 view: None,
439 materialized: false,
440 app_label: default_app_label(),
441 }
442 }
443}
444
445impl Default for Column {
446 /// A nullable-free `BigInt` column with every optional marker off, so a
447 /// hand-built fixture can name the two or three fields it cares about and
448 /// `..Column::default()` the rest.
449 ///
450 /// `db_constraint` defaults to `true`, matching the serde default: an FK
451 /// emits a physical `REFERENCES` clause unless the field opts out. Getting
452 /// this backwards would silently drop constraints AND the plugin ordering
453 /// edges derived from them (see `app::fk_plugin_edges`).
454 fn default() -> Self {
455 Self {
456 name: String::new(),
457 ty: crate::orm::SqlType::BigInt,
458 primary_key: false,
459 nullable: false,
460 fk_target: None,
461 noform: false,
462 privileged: false,
463 private: false,
464 secret: false,
465 db_constraint: true,
466 noedit: false,
467 is_string_repr: false,
468 max_length: 0,
469 choices: Vec::new(),
470 choice_labels: Vec::new(),
471 default: String::new(),
472 is_multichoice: false,
473 unique: false,
474 on_delete: crate::orm::FkAction::NoAction,
475 on_update: crate::orm::FkAction::NoAction,
476 index: false,
477 auto_now_add: false,
478 auto_uuid: false,
479 auto_now: false,
480 auto_user_add: false,
481 auto_user: false,
482 trim: false,
483 lowercase: false,
484 case_insensitive: false,
485 help: String::new(),
486 example: String::new(),
487 widget: None,
488 supported_backends: Vec::new(),
489 min: None,
490 max: None,
491 text_format: None,
492 slug_from: None,
493 }
494 }
495}
496
497/// Owned mirror of `orm::M2MRelationSpec` so `ModelMeta` can be
498/// serialised into migration JSON without lifetimes.
499#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
500pub struct M2MRelation {
501 pub field_name: String,
502 pub target_table: String,
503 pub target_name: String,
504}
505
506fn default_icon() -> String {
507 "database".to_string()
508}
509
510/// Serde default for [`Operation::CreateM2MTable`]'s `parent_ty` /
511/// `child_ty` fields. Older snapshot files (pre-phase-2) had no
512/// per-side PK type and assumed `BigInt` on both ends — this keeps
513/// them round-tripping without rewrites.
514fn default_bigint() -> crate::orm::SqlType {
515 crate::orm::SqlType::BigInt
516}
517
518impl ModelMeta {
519 /// The primary-key column on this model. Every umbral model
520 /// has exactly one PK by construction (the derive enforces
521 /// it), but the lookup is `Option`-shaped because nothing
522 /// stops a hand-written `ModelMeta` (test fixtures, etc.)
523 /// from omitting it.
524 pub fn pk_column(&self) -> Option<&Column> {
525 self.fields.iter().find(|c| c.primary_key)
526 }
527
528 /// Read static metadata off `T: Model` into an owned `ModelMeta`.
529 /// Called from `AppBuilder::model::<T>()`.
530 pub fn for_<T: Model>() -> Self {
531 Self {
532 name: T::NAME.to_string(),
533 table: T::TABLE.to_string(),
534 fields: T::FIELDS.iter().map(Column::from).collect(),
535 display: T::DISPLAY.to_string(),
536 icon: T::ICON.to_string(),
537 database: T::DATABASE.map(|s| s.to_string()),
538 singleton: T::SINGLETON,
539 unique_together: T::UNIQUE_TOGETHER
540 .iter()
541 .map(|group| group.iter().map(|s| s.to_string()).collect())
542 .collect(),
543 indexes: T::INDEXES
544 .iter()
545 .map(|group| group.iter().map(|s| s.to_string()).collect())
546 .collect(),
547 ordering: T::ORDERING
548 .iter()
549 .map(|(col, desc)| (col.to_string(), *desc))
550 .collect(),
551 m2m_relations: T::M2M_RELATIONS
552 .iter()
553 .map(|r| M2MRelation {
554 field_name: r.field_name.to_string(),
555 target_table: r.target_table.to_string(),
556 target_name: r.target_name.to_string(),
557 })
558 .collect(),
559 soft_delete: T::SOFT_DELETE,
560 audited: T::AUDITED,
561 view: T::VIEW.map(|s| s.to_string()),
562 materialized: T::MATERIALIZED,
563 app_label: T::APP_LABEL.to_string(),
564 }
565 }
566}
567
568/// A snapshot of every registered model at a point in time.
569///
570/// Serialised into the `snapshot_after` field of a migration file so
571/// future `makemigrations` runs can diff against it without replaying
572/// every prior migration's operations.
573#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
574pub struct Snapshot {
575 /// Models sorted by name so the JSON is deterministic and the
576 /// snapshot_hash is stable across runs that produce equivalent
577 /// content.
578 pub models: Vec<ModelMeta>,
579}
580
581impl Snapshot {
582 /// Build a snapshot from the live registry (the current state of
583 /// the application's models, post-`App::build`).
584 pub fn current() -> Self {
585 let mut models = registered_models().to_vec();
586 models.sort_by(|a, b| a.name.cmp(&b.name));
587 Self { models }
588 }
589
590 /// Build a snapshot containing only the models registered
591 /// against the given plugin. Used by `make_in` to diff each
592 /// plugin's migrations independently against its own prior
593 /// snapshot, so cross-plugin model sets don't bleed into one
594 /// migration file.
595 pub fn current_for(plugin: &str) -> Self {
596 let mut models = models_for_plugin(plugin);
597 models.sort_by(|a, b| a.name.cmp(&b.name));
598 Self { models }
599 }
600
601 /// Compute the snapshot's SHA-256 hash, hex-encoded. Stored in the
602 /// `umbral_migrations.snapshot_hash` column for drift detection.
603 pub fn hash(&self) -> String {
604 use sha2::{Digest, Sha256};
605 let json = serde_json::to_string(self).expect("Snapshot serializes");
606 let digest = Sha256::digest(json.as_bytes());
607 hex(&digest[..])
608 }
609}
610
611fn hex(bytes: &[u8]) -> String {
612 const HEX: &[u8; 16] = b"0123456789abcdef";
613 let mut s = String::with_capacity(bytes.len() * 2);
614 for b in bytes {
615 s.push(HEX[(b >> 4) as usize] as char);
616 s.push(HEX[(b & 0x0f) as usize] as char);
617 }
618 s
619}
620
621/// One operation inside a migration. The migration engine renders each
622/// operation to SQL via the active backend (M4 `DatabaseBackend::
623/// map_type`) and runs them in declaration order inside one
624/// transaction per migration file.
625///
626/// Shipped: table-level ops (`CreateTable` / `DropTable`), `AddColumn` /
627/// `DropColumn`, `AlterColumn` (nullable flips), index / constraint ops
628/// (`AddIndex` / `DropIndex`), rename detection (`RenameTable` /
629/// `RenameColumn`), database views (`CreateView` / `DropView`), M2M junction
630/// tables (`CreateM2MTable` / `DropM2MTable`), column comments
631/// (`SetColumnComment`), and the `RunSql` data-migration escape hatch.
632/// Still deferred: `AlterColumn` **type changes**, primary-key flips, and the
633/// `RunCode` Rust data-migration op. See `docs/specs/06-migration-engine.md`.
634#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
635#[serde(tag = "kind")]
636pub enum Operation {
637 /// Create a new table. `columns` is in declaration order; the
638 /// engine builds a sea-query `Table::create()` over them and runs
639 /// the rendered DDL. `unique_together` lowers to inline
640 /// `UNIQUE (col1, col2)` clauses; `indexes` lowers to follow-up
641 /// `CREATE INDEX` statements after the table is created. Both
642 /// default to empty for backward-compat with older snapshots.
643 CreateTable {
644 table: String,
645 columns: Vec<Column>,
646 #[serde(default, skip_serializing_if = "Vec::is_empty")]
647 unique_together: Vec<Vec<String>>,
648 #[serde(default, skip_serializing_if = "Vec::is_empty")]
649 indexes: Vec<Vec<String>>,
650 },
651 /// Drop an existing table.
652 DropTable { table: String },
653 /// features #73 — create a database VIEW from `#[umbral(view = "...")]`.
654 ///
655 /// There is no `AlterView`, and there never will be: a view holds no rows, so
656 /// changing one is a drop and a recreate. That is why views escape every hard
657 /// problem the table side of this engine spent its life on — no data-preserving
658 /// ALTER, no rename-vs-drop disambiguation, nothing to back up.
659 CreateView {
660 /// The view's name in the database (the model's `TABLE`).
661 name: String,
662 /// The SELECT body, verbatim from the attribute. Opaque to the engine.
663 sql: String,
664 /// `CREATE MATERIALIZED VIEW` (Postgres-only) rather than `CREATE VIEW`.
665 #[serde(default, skip_serializing_if = "is_false")]
666 materialized: bool,
667 },
668 /// features #73 — drop a database VIEW.
669 DropView {
670 /// The view's name in the database.
671 name: String,
672 /// Whether it was created as MATERIALIZED — the DROP statement differs.
673 #[serde(default, skip_serializing_if = "is_false")]
674 materialized: bool,
675 },
676 /// Add a new column to an existing table. Rendered as
677 /// `ALTER TABLE x ADD COLUMN y TYPE [NOT NULL]`. SQLite refuses a
678 /// non-nullable add against a populated table without a default;
679 /// the engine surfaces that as a sqlx error at apply time (M8 v1).
680 /// A future op `AddColumnWithDefault` lifts the restriction once
681 /// the `#[umbral(default = ...)]` attribute lands.
682 AddColumn { table: String, column: Column },
683 /// Drop a column from an existing table. Rendered as
684 /// `ALTER TABLE x DROP COLUMN y`. SQLite 3.35+ and Postgres
685 /// support this natively; older SQLite would need a table-
686 /// recreation dance the engine doesn't implement.
687 DropColumn { table: String, column: String },
688 /// Alter a column's nullable flag (the only safe in-place change
689 /// the engine ships at M5.1). Self-contained: carries the full
690 /// new column list so the SQLite table-recreation dance can
691 /// rebuild the schema without re-reading the snapshot. The
692 /// `column` field names the specific column that triggered the
693 /// alter (used for the filename suffix and diagnostics); the
694 /// `new_columns` list is the post-change schema.
695 AlterColumn {
696 table: String,
697 column: String,
698 new_columns: Vec<Column>,
699 /// The table's composite UNIQUE groups (audit_2 core-migrate #10). The
700 /// SQLite recreation dance rebuilds the table from scratch, so it must
701 /// re-emit these — otherwise a nullable-flip / safe-cast alter silently
702 /// DROPS every composite UNIQUE constraint (duplicates become
703 /// insertable — integrity loss). Postgres alters in place and ignores
704 /// this. `serde(default)` keeps older on-disk migrations deserialising.
705 #[serde(default, skip_serializing_if = "Vec::is_empty")]
706 unique_together: Vec<Vec<String>>,
707 /// The table's composite (multi-column) index groups (audit_2
708 /// core-migrate #10). Re-created by the SQLite dance after the rebuild,
709 /// for the same reason as `unique_together`. Single-column / FK /
710 /// soft-delete indexes are re-derived from `new_columns`.
711 #[serde(default, skip_serializing_if = "Vec::is_empty")]
712 indexes: Vec<Vec<String>>,
713 /// Snapshot of the table's columns *before* this alter. Carried
714 /// so the Postgres renderer can decide per-column whether it
715 /// needs a TYPE/USING clause vs a SET/DROP NOT NULL — without
716 /// re-walking the snapshot file. `Option` + `serde(default)`
717 /// keeps older on-disk migrations deserialising cleanly; ops
718 /// produced before this field existed get `None` and fall back
719 /// to the legacy nullable-only Postgres path.
720 #[serde(default, skip_serializing_if = "Option::is_none")]
721 prev_columns: Option<Vec<Column>>,
722 },
723 /// Rename an existing table. Emitted by `diff` when a model's table
724 /// name changes but its `Model::NAME` (the Rust struct name) stays
725 /// the same (first-pass detection), or when the column shapes are
726 /// bit-identical and the struct name changed too (second-pass
727 /// heuristic detection). Both SQLite and Postgres render as
728 /// `ALTER TABLE "<from>" RENAME TO "<to>"`.
729 ///
730 /// The migration tracking table records `(plugin, name)` of each
731 /// applied migration — it is not affected by a table rename inside
732 /// the migration.
733 RenameTable { from: String, to: String },
734 /// Create a many-to-many junction table. Auto-emitted when a model
735 /// gains an `M2M<T>` field. Closes BUG-16 phase 2.
736 ///
737 /// The junction table name is `parent_table_field_name`. Columns:
738 /// `parent_id` (FK to parent), `child_id` (FK to target), both with
739 /// `ON DELETE CASCADE`. Composite PK `(parent_id, child_id)`.
740 ///
741 /// `parent_ty` and `child_ty` carry the SQL types of the
742 /// referenced PK columns — `BigInt` for an `i64` PK, `Text` for a
743 /// `String` slug, `Uuid` for a `uuid::Uuid`. The renderer maps
744 /// these to the right column type per backend; without them the
745 /// junction's `child_id INTEGER` would reject a string codename
746 /// at insert time. `#[serde(default)]` keeps older snapshot files
747 /// (pre-phase-2) round-tripping — they default to `BigInt`,
748 /// matching the original i64-only behaviour.
749 CreateM2MTable {
750 junction_table: String,
751 parent_table: String,
752 parent_col: String,
753 child_table: String,
754 child_col: String,
755 #[serde(default = "default_bigint")]
756 parent_ty: crate::orm::SqlType,
757 #[serde(default = "default_bigint")]
758 child_ty: crate::orm::SqlType,
759 },
760 /// Drop a many-to-many junction table. Auto-emitted when an `M2M<T>`
761 /// field is removed from a model.
762 DropM2MTable { junction_table: String },
763 /// Gap 88: rename a column on an existing table. Emitted by the
764 /// diff engine when a single column with one shape was dropped
765 /// and one with the same shape was added in the same diff —
766 /// the heuristic match for "the user renamed `title` to
767 /// `headline`." Both SQLite (3.25+) and Postgres render as
768 /// `ALTER TABLE "<t>" RENAME COLUMN "<from>" TO "<to>"`.
769 ///
770 /// `column` carries the post-rename column shape so the
771 /// snapshot stays in sync. The migration only renames; never
772 /// alters other column attributes — a rename combined with a
773 /// type change emits a RenameColumn AND a follow-on AlterColumn
774 /// against the new name.
775 RenameColumn {
776 table: String,
777 from: String,
778 to: String,
779 #[serde(default, skip_serializing_if = "Option::is_none")]
780 column: Option<Column>,
781 },
782 /// gaps3 #43: set (or clear) a column's database comment from
783 /// `#[umbral(help = "...")]`. Renders as `COMMENT ON COLUMN "<t>"."<c>" IS
784 /// '<comment>'` on Postgres, and to *nothing* on SQLite, which has no
785 /// comment facility. An empty `comment` clears it with `IS NULL` — Postgres
786 /// distinguishes "no comment" from "the empty comment".
787 ///
788 /// `CreateTable` and `AddColumn` emit their own comments inline, so this op
789 /// exists for the case they can't cover: the help text on an *existing*
790 /// column changed. It touches no rows and takes no lock worth naming, so it
791 /// classifies as `OpSafety::Safe`.
792 SetColumnComment {
793 table: String,
794 column: String,
795 /// Empty string means "remove the comment".
796 comment: String,
797 },
798 /// Gap #69: a raw-SQL **data** migration. Unlike every other
799 /// variant it changes *rows*, not the schema model — so the
800 /// autodetector NEVER emits it (it has no model-state effect), and
801 /// a migration carrying only `RunSql` ops has
802 /// `snapshot_after == snapshot_before`. It is always hand-authored:
803 /// generate an empty migration with `makemigrations --empty
804 /// <plugin>`, then add the `RunSql` op by editing the file.
805 ///
806 /// `sql` is the forward statement(s), executed verbatim on the
807 /// per-migration transaction — same string on both backends (raw
808 /// SQL the renderer passes through untouched), so the author owns
809 /// portability. `reverse_sql` is the optional un-apply statement
810 /// (used by a future `migrate --reverse`); `None` means
811 /// irreversible.
812 ///
813 /// Under schema-per-tenant the op runs **per tenant schema** (the
814 /// schema-migrate loop applies every op under the
815 /// `<schema>, public` search_path), so a tenant-app `RunSql` writes
816 /// tenant rows while reading shared `public` lookup tables — the
817 /// boundary-spanning data migration. A shared-app `RunSql` runs once
818 /// in `public` via the normal `migrate`.
819 RunSql {
820 sql: String,
821 #[serde(default, skip_serializing_if = "Option::is_none")]
822 reverse_sql: Option<String>,
823 },
824 /// Create a composite index, or a composite UNIQUE constraint, on an
825 /// EXISTING table. Emitted by `diff` when a model gains a
826 /// `unique_together` group or a multi-column `indexes` entry with no
827 /// accompanying column change — a case that previously produced NO
828 /// migration at all, so the constraint was silently never created.
829 ///
830 /// `unique` selects `CREATE UNIQUE INDEX` (a `unique_together` group)
831 /// vs a plain `CREATE INDEX` (an `indexes` group). The index NAME is
832 /// deterministic — `uniq_<table>_<cols>` when unique, `idx_<table>_<cols>`
833 /// otherwise — so the matching [`DropIndex`](Operation::DropIndex) can
834 /// name it. Rendered `IF NOT EXISTS` on both backends, so it is a safe
835 /// no-op when a same-migration `AlterColumn` already rebuilt the table
836 /// with the constraint (the SQLite dance) — the two never conflict.
837 AddIndex {
838 table: String,
839 columns: Vec<String>,
840 #[serde(default)]
841 unique: bool,
842 },
843 /// Drop a composite index / UNIQUE constraint previously created by an
844 /// [`AddIndex`](Operation::AddIndex), or by a `CreateTable` that renders
845 /// its `unique_together`/`indexes` as the same deterministically-named
846 /// indexes. Emitted by `diff` when a model LOSES a `unique_together`
847 /// group or `indexes` entry. Rendered `DROP INDEX IF EXISTS <name>` on
848 /// both backends (the name is recomputed from `table` + `columns` +
849 /// `unique`, matching `AddIndex`).
850 DropIndex {
851 table: String,
852 columns: Vec<String>,
853 #[serde(default)]
854 unique: bool,
855 },
856}
857
858impl Operation {
859 /// The primary table this operation targets. For `RenameTable`,
860 /// returns the source name (the post-rename `to` lives in the new
861 /// snapshot, but routing decisions look up the model meta by its
862 /// pre-rename `from`).
863 ///
864 /// Used by `run_in`'s per-DB dispatch loop to route each op to the
865 /// pool where its table actually lives.
866 pub fn table_name(&self) -> &str {
867 match self {
868 Operation::CreateView { name, .. } | Operation::DropView { name, .. } => name,
869 Operation::CreateTable { table, .. }
870 | Operation::DropTable { table }
871 | Operation::AddColumn { table, .. }
872 | Operation::DropColumn { table, .. }
873 | Operation::AlterColumn { table, .. }
874 | Operation::RenameColumn { table, .. }
875 | Operation::SetColumnComment { table, .. }
876 | Operation::AddIndex { table, .. }
877 | Operation::DropIndex { table, .. } => table,
878 Operation::RenameTable { from, .. } => from,
879 Operation::CreateM2MTable { junction_table, .. }
880 | Operation::DropM2MTable { junction_table } => junction_table,
881 // A data migration targets no single table. The empty name
882 // routes it to the `"default"` alias via `table_alias`'s
883 // fallback (see `op_targets_alias`).
884 Operation::RunSql { .. } => "",
885 }
886 }
887}
888
889/// One column inside a [`Operation::CreateTable`].
890///
891/// Mirrors the structure of [`FieldSpec`] but is fully owned for
892/// serialisation. Reconstructed from a `FieldSpec` at diff time.
893#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
894pub struct Column {
895 pub name: String,
896 pub ty: SqlType,
897 pub primary_key: bool,
898 pub nullable: bool,
899 /// For `SqlType::ForeignKey` columns: the SQL table name of the
900 /// referenced model. `None` for all non-FK columns.
901 #[serde(default, skip_serializing_if = "Option::is_none")]
902 pub fk_target: Option<String>,
903 /// When `true`, this field is never shown on any admin form (create or
904 /// edit). Propagated from `FieldSpec::noform`.
905 #[serde(default)]
906 pub noform: bool,
907 /// When `true`, this field is a privileged/server-managed column that the
908 /// untrusted JSON write path (`insert_json`/`update_json`, i.e. REST
909 /// create/update + admin form-submit) strips UNLESS the caller explicitly
910 /// authorizes it via [`DynQuerySet::allow_privileged`]. Propagated from
911 /// `FieldSpec::privileged`. This is the default-DENY guard for mass
912 /// assignment of fields like `is_superuser`/`is_staff`/ownership FKs
913 /// (audit_2 H3): they can't be set by an unprivileged writer even if the
914 /// model exposes them on its create/update surface. Purely a write-auth
915 /// concern — never part of the schema shape, so it is skipped from
916 /// migration snapshots when at its `false` default.
917 #[serde(default, skip_serializing_if = "is_false")]
918 pub privileged: bool,
919 /// Confidential: stripped from serialized output unless the read unlocks it via
920 /// `DynQuerySet::allow_private`. Propagated from `FieldSpec::private`. Purely a
921 /// read-auth concern — never part of the schema shape, so it is skipped from migration
922 /// snapshots at its `false` default, exactly like `privileged`.
923 #[serde(default, skip_serializing_if = "is_false")]
924 pub private: bool,
925 /// Never serialized to a client, with no unlock. Propagated from `FieldSpec::secret`
926 /// (and set automatically for every `Masked<T>` field). Also not schema shape.
927 #[serde(default, skip_serializing_if = "is_false")]
928 pub secret: bool,
929 /// For FK columns: whether to emit a physical `FOREIGN KEY ...
930 /// REFERENCES` constraint. Propagated from `FieldSpec::db_constraint`.
931 /// `false` (set via `#[umbral(db_constraint = false)]`) keeps the
932 /// logical FK (column + `fk_target`) but renders no `REFERENCES`
933 /// clause — the only valid shape for a cross-database FK. Closes
934 /// gaps2 #22. Defaults to `true` so existing migration JSON
935 /// round-trips unchanged (omitted from JSON when at its default).
936 #[serde(default = "default_true", skip_serializing_if = "is_true")]
937 pub db_constraint: bool,
938 /// When `true`, this field appears on the edit form as read-only.
939 /// Propagated from `FieldSpec::noedit`.
940 #[serde(default)]
941 pub noedit: bool,
942 /// Display-string marker — propagated from
943 /// `FieldSpec::is_string_repr`. The admin uses the first column
944 /// with this flag as the default `list_display` label when no
945 /// explicit one is configured.
946 #[serde(default)]
947 pub is_string_repr: bool,
948 /// Display truncation cap — propagated from `FieldSpec::max_length`.
949 /// `0` means no truncation.
950 #[serde(default)]
951 pub max_length: u32,
952 /// Closed-set DB values for a choices column. Propagated from
953 /// `FieldSpec::choices`. Non-empty when the model field carries
954 /// `#[umbral(choices)]`; the migration engine emits a Postgres
955 /// `CHECK (col IN (...))` constraint when this slice is non-empty.
956 /// Empty for every non-choices column.
957 #[serde(default, skip_serializing_if = "Vec::is_empty")]
958 pub choices: Vec<String>,
959 /// Human labels matching `choices` position-for-position. Carried
960 /// alongside `choices` so the admin's `<select>` widget has labels
961 /// without the runtime needing to reflect on the model type.
962 #[serde(default, skip_serializing_if = "Vec::is_empty")]
963 pub choice_labels: Vec<String>,
964 /// SQL `DEFAULT` value — propagated from `FieldSpec::default`.
965 /// Empty string means no default. The migration engine reads this
966 /// at DDL-emit time for both `CREATE TABLE` and `ALTER TABLE ADD
967 /// COLUMN`. Set via `#[umbral(default = "...")]` on the model field.
968 #[serde(default, skip_serializing_if = "String::is_empty")]
969 pub default: String,
970 /// Distinguishes a multi-valued [`MultiChoice<E>`] column from a
971 /// single-valued choices column. Both share `ty: Text` plus the same
972 /// `choices` / `choice_labels` metadata; this flag is the only
973 /// signal that the value is a CSV. Empty / false for every other
974 /// column.
975 ///
976 /// [`MultiChoice<E>`]: crate::orm::MultiChoice
977 #[serde(default, skip_serializing_if = "is_false")]
978 pub is_multichoice: bool,
979
980 /// Carries `FieldSpec::unique` into the migration snapshot. The
981 /// DDL builders emit a `UNIQUE` clause on this column at
982 /// `CREATE TABLE` time when set. Default `false` keeps existing
983 /// migration JSON files round-tripping unchanged (the field is
984 /// omitted on serialise when default).
985 #[serde(default, skip_serializing_if = "is_false")]
986 pub unique: bool,
987
988 /// Carries `FieldSpec::on_delete` into the migration snapshot.
989 /// FK columns only — the DDL builders emit
990 /// `ON DELETE <action>` when this is anything other than
991 /// `NoAction`. Default `NoAction` is omitted from JSON so
992 /// existing migration files round-trip without churn.
993 #[serde(default, skip_serializing_if = "is_no_action")]
994 pub on_delete: crate::orm::FkAction,
995
996 /// Carries `FieldSpec::on_update` into the migration snapshot.
997 /// Same shape as `on_delete`; emits `ON UPDATE <action>`.
998 #[serde(default, skip_serializing_if = "is_no_action")]
999 pub on_update: crate::orm::FkAction,
1000
1001 /// Carries `FieldSpec::index` into the migration snapshot. The
1002 /// CreateTable + AddColumn render paths emit a matching
1003 /// `CREATE INDEX idx_<table>_<col>` for every column whose
1004 /// flag is set. Default `false` keeps existing migration JSON
1005 /// round-tripping unchanged.
1006 #[serde(default, skip_serializing_if = "is_false")]
1007 pub index: bool,
1008
1009 /// Carries `FieldSpec::auto_now_add` into the migration
1010 /// snapshot. The dynamic write path (`DynQuerySet::insert_json`)
1011 /// auto-populates the column with `Utc::now()` when the body
1012 /// omits it. Default `false` so existing migration JSON
1013 /// round-trips unchanged.
1014 #[serde(default, skip_serializing_if = "is_false")]
1015 pub auto_now_add: bool,
1016 /// Carries `FieldSpec::auto_user_add` to the runtime (dyn) write path, so
1017 /// admin and REST stamp the author too — they run on `DynQuerySet`, not the
1018 /// typed one. Not a schema property: skipped in snapshots.
1019 #[serde(default, skip_serializing_if = "std::ops::Not::not")]
1020 pub auto_user_add: bool,
1021 /// Carries `FieldSpec::auto_user`.
1022 #[serde(default, skip_serializing_if = "std::ops::Not::not")]
1023 pub auto_user: bool,
1024
1025 /// Carries `FieldSpec::auto_now` into the migration snapshot.
1026 /// Same shape as `auto_now_add` but fires on update too.
1027 #[serde(default, skip_serializing_if = "is_false")]
1028 pub auto_now: bool,
1029
1030 /// Carries `FieldSpec::auto_uuid` to the runtime write paths, which
1031 /// generate a fresh `Uuid::new_v4()` when the column is omitted/nil on
1032 /// create. A write-behaviour, not a schema property, so it's skipped in
1033 /// snapshots (like `auto_user_add`).
1034 #[serde(default, skip_serializing_if = "std::ops::Not::not")]
1035 pub auto_uuid: bool,
1036
1037 /// Carries `FieldSpec::trim` into the snapshot. Behavioral (dynamic write
1038 /// path strips surrounding whitespace), not schema-affecting — excluded
1039 /// from the schema diff like `auto_now`. Default `false` so existing
1040 /// migration JSON round-trips unchanged.
1041 #[serde(default, skip_serializing_if = "is_false")]
1042 pub trim: bool,
1043
1044 /// Carries `FieldSpec::lowercase` into the snapshot. Behavioral (dynamic
1045 /// write path lowercases the value), not schema-affecting. Default `false`
1046 /// so existing migration JSON round-trips unchanged.
1047 #[serde(default, skip_serializing_if = "is_false")]
1048 pub lowercase: bool,
1049
1050 /// Carries `FieldSpec::case_insensitive` into the snapshot. Schema-affecting
1051 /// (Postgres `citext` / SQLite `COLLATE NOCASE`), but — like `unique` — the
1052 /// diff comparator (`column_shape`) does NOT watch it, so it applies at
1053 /// CREATE TABLE and toggling it on a live column needs a hand-written
1054 /// migration. Default `false` so existing migration JSON round-trips.
1055 #[serde(default, skip_serializing_if = "is_false")]
1056 pub case_insensitive: bool,
1057
1058 /// Carries `FieldSpec::help` into the migration snapshot.
1059 /// Default empty string is omitted from JSON so existing
1060 /// migration files round-trip unchanged.
1061 #[serde(default, skip_serializing_if = "String::is_empty")]
1062 pub help: String,
1063
1064 /// Carries `FieldSpec::example` into the migration snapshot.
1065 /// Same shape as `help`.
1066 #[serde(default, skip_serializing_if = "String::is_empty")]
1067 pub example: String,
1068
1069 /// Carries `FieldSpec::widget` into the migration snapshot — the
1070 /// form-renderer presentation hint (features.md #4). Presentation
1071 /// only, no DB effect, so it's excluded from the schema diff the
1072 /// same way `help` / `example` are. `None` is omitted from JSON so
1073 /// existing migration files round-trip unchanged.
1074 #[serde(default, skip_serializing_if = "Option::is_none")]
1075 pub widget: Option<String>,
1076
1077 /// Carries `FieldSpec::supported_backends` into the migration
1078 /// snapshot. When non-empty, the boot system check rejects the
1079 /// model on any backend not listed. Closes IMP-5 from
1080 /// `bugs/tests/testBugs.md`. Default empty (works on every
1081 /// backend); JSON skip-when-empty so existing migration files
1082 /// don't churn.
1083 #[serde(default, skip_serializing_if = "Vec::is_empty")]
1084 pub supported_backends: Vec<String>,
1085
1086 /// IMP-3: numeric lower bound. `None` means "no minimum"; the
1087 /// DDL emits a `CHECK (col >= N)` constraint when set.
1088 #[serde(default, skip_serializing_if = "Option::is_none")]
1089 pub min: Option<i64>,
1090
1091 /// IMP-3: numeric upper bound. Same shape as `min`.
1092 #[serde(default, skip_serializing_if = "Option::is_none")]
1093 pub max: Option<i64>,
1094
1095 /// BUG-11/12/13: constrained-text marker. `None` is plain text;
1096 /// `Some("slug" | "email" | "url")` flags the column as a
1097 /// `Slug` / `Email` / `Url` wrapper. OpenAPI emits the
1098 /// corresponding `format` / `pattern`; the REST plugin
1099 /// pre-validates the body via `validate_text_format`.
1100 #[serde(default, skip_serializing_if = "Option::is_none")]
1101 pub text_format: Option<String>,
1102
1103 /// Gap 109: auto-derive source. When `Some("title")`, the slug is
1104 /// computed from the row's `title` column at write time if the
1105 /// slug column itself is empty / missing on the body. Pure
1106 /// runtime behaviour — has no DDL effect, so the diff engine
1107 /// ignores changes to this field. `#[serde(default)]` keeps
1108 /// older snapshots round-tripping.
1109 #[serde(default, skip_serializing_if = "Option::is_none")]
1110 pub slug_from: Option<String>,
1111}
1112
1113fn is_no_action(a: &crate::orm::FkAction) -> bool {
1114 matches!(a, crate::orm::FkAction::NoAction)
1115}
1116
1117/// Postgres `COMMENT ON COLUMN` for one column (gaps3 #43).
1118///
1119/// An empty `comment` renders `IS NULL`, which is Postgres's "this column has no
1120/// comment" — distinct from `IS ''`, an empty comment that `\d+` would print as
1121/// a blank line.
1122///
1123/// Help text is prose, and prose has apostrophes, so the literal is escaped by
1124/// doubling single quotes. The value originates in a `#[umbral(help = "...")]`
1125/// attribute (a compile-time literal, not user input), but a migration file is
1126/// hand-editable and an unescaped quote would produce a syntax error at apply
1127/// time rather than a diagnostic at generate time.
1128fn comment_on_column_stmt(table: &str, column: &str, comment: &str) -> String {
1129 let t = table.replace('"', "\"\"");
1130 let c = column.replace('"', "\"\"");
1131 if comment.is_empty() {
1132 return format!("COMMENT ON COLUMN \"{t}\".\"{c}\" IS NULL");
1133 }
1134 let body = comment.replace('\'', "''");
1135 format!("COMMENT ON COLUMN \"{t}\".\"{c}\" IS '{body}'")
1136}
1137
1138/// Build a portable `CREATE INDEX IF NOT EXISTS idx_<table>_<col>
1139/// ON "<table>" ("<col>")` statement. Same DDL on SQLite and
1140/// Postgres — both accept `CREATE INDEX IF NOT EXISTS` and the
1141/// `idx_<table>_<col>` name convention is unique enough that the
1142/// migration engine can re-emit it idempotently on subsequent
1143/// applies. Used by [`render_operation_sqlite`] / `_postgres`
1144/// after a `CreateTable` or `AddColumn` op whose column carries
1145/// the `#[umbral(index)]` flag. Closes BUG-4.
1146fn create_index_stmt(table: &str, column: &str) -> String {
1147 let t = table.replace('"', "\"\"");
1148 let c = column.replace('"', "\"\"");
1149 format!(
1150 "CREATE INDEX IF NOT EXISTS \"idx_{table}_{column}\" ON \"{t}\" (\"{c}\")",
1151 table = table.replace('"', ""),
1152 column = column.replace('"', ""),
1153 )
1154}
1155
1156/// `CREATE UNIQUE INDEX` for a column. Used when a new `#[umbral(unique)]`
1157/// column is ADDed on SQLite: SQLite rejects `ALTER TABLE ADD COLUMN …
1158/// UNIQUE` outright, so the uniqueness is applied as a separate unique
1159/// index instead (semantically identical, and it applies cleanly).
1160fn create_unique_index_stmt(table: &str, column: &str) -> String {
1161 let t = table.replace('"', "\"\"");
1162 let c = column.replace('"', "\"\"");
1163 format!(
1164 "CREATE UNIQUE INDEX IF NOT EXISTS \"uniq_{table}_{column}\" ON \"{t}\" (\"{c}\")",
1165 table = table.replace('"', ""),
1166 column = column.replace('"', ""),
1167 )
1168}
1169
1170/// Build a Postgres `CREATE INDEX ... USING GIN` for a `tsvector`
1171/// (`SqlType::FullText`) column (#33). A tsvector column is useless for
1172/// search without a GIN index, so the migration engine emits one
1173/// automatically for every full-text column — the caller never has to
1174/// hand-write it. **Postgres-only**: GIN is Postgres syntax and FullText
1175/// columns are system-check-gated to Postgres, so this only ever renders
1176/// from `render_operation_postgres`. The `_gin` name suffix keeps it
1177/// distinct from any plain index on the same column.
1178fn create_gin_index_stmt(table: &str, column: &str) -> String {
1179 let t = table.replace('"', "\"\"");
1180 let c = column.replace('"', "\"\"");
1181 format!(
1182 "CREATE INDEX IF NOT EXISTS \"idx_{table}_{column}_gin\" ON \"{t}\" USING GIN (\"{c}\")",
1183 table = table.replace('"', ""),
1184 column = column.replace('"', ""),
1185 )
1186}
1187
1188/// Build a Postgres `CREATE INDEX ... USING GIST` for a PostGIS spatial
1189/// (`SqlType::Geometry` / `SqlType::Geography`) column. Spatial predicates
1190/// (`ST_DWithin`, `&&`, …) are only index-accelerated by a GiST index; a plain
1191/// B-tree does nothing for them. So an `#[umbral(index)]` on a geometry column
1192/// emits GiST rather than the default B-tree. **Postgres-only** — spatial
1193/// columns are system-check-gated to Postgres, so this only renders from
1194/// `render_operation_postgres`.
1195fn create_gist_index_stmt(table: &str, column: &str) -> String {
1196 let t = table.replace('"', "\"\"");
1197 let c = column.replace('"', "\"\"");
1198 format!(
1199 "CREATE INDEX IF NOT EXISTS \"idx_{table}_{column}_gist\" ON \"{t}\" USING GIST (\"{c}\")",
1200 table = table.replace('"', ""),
1201 column = column.replace('"', ""),
1202 )
1203}
1204
1205/// Multi-column variant of [`create_index_stmt`]. Closes BUG-7.
1206/// Renders `CREATE INDEX IF NOT EXISTS idx_<table>_<col1>_<col2>
1207/// ON "<table>" ("<col1>", "<col2>")`. Both backends accept the
1208/// same form. Empty groups render no statement (defensive — the
1209/// macro layer rejects them before the engine sees them, but the
1210/// helper still returns a no-op SQL string to keep the caller
1211/// simple).
1212fn create_multi_index_stmt(table: &str, columns: &[String]) -> String {
1213 // A plain composite index IS an `AddIndex { unique: false }` render —
1214 // delegate so the NAME (`idx_<table>_<cols>`) is defined in exactly one
1215 // place and a `CreateTable`'s composite index and a later `DropIndex`
1216 // always agree on it.
1217 add_index_stmt(table, columns, false)
1218}
1219
1220/// Deterministic name for a composite index. `unique` selects the `uniq_`
1221/// prefix (a `unique_together` group), otherwise `idx_`. Derived purely
1222/// from the quote-stripped table + column list so an [`Operation::AddIndex`]
1223/// and the later [`Operation::DropIndex`] that reverses it always compute
1224/// the same name. The `uniq_`/`idx_` split means a UNIQUE and a plain index
1225/// on the SAME columns never collide.
1226fn index_name(table: &str, columns: &[String], unique: bool) -> String {
1227 let t = table.replace('"', "");
1228 let suffix = columns
1229 .iter()
1230 .map(|c| c.replace('"', ""))
1231 .collect::<Vec<_>>()
1232 .join("_");
1233 let prefix = if unique { "uniq" } else { "idx" };
1234 format!("{prefix}_{t}_{suffix}")
1235}
1236
1237/// `CREATE [UNIQUE] INDEX IF NOT EXISTS "<name>" ON "<table>" (cols)` —
1238/// identical syntax on SQLite and Postgres. The index NAME is a bare
1239/// identifier (via [`index_name`]); the ON-clause table reference is a
1240/// *quoted* identifier with inner quotes doubled. An empty column list
1241/// renders an empty string (defensive no-op; the macro layer rejects
1242/// empty groups upstream).
1243fn add_index_stmt(table: &str, columns: &[String], unique: bool) -> String {
1244 if columns.is_empty() {
1245 return String::new();
1246 }
1247 let name = index_name(table, columns, unique);
1248 let t_esc = table.replace('"', "\"\"");
1249 let col_list = columns
1250 .iter()
1251 .map(|c| format!("\"{}\"", c.replace('"', "\"\"")))
1252 .collect::<Vec<_>>()
1253 .join(", ");
1254 let unique_kw = if unique { "UNIQUE " } else { "" };
1255 format!("CREATE {unique_kw}INDEX IF NOT EXISTS \"{name}\" ON \"{t_esc}\" ({col_list})")
1256}
1257
1258/// `DROP INDEX IF EXISTS "<name>"` — same on both backends. Postgres
1259/// resolves the unqualified name via the search_path (so a schema-per-tenant
1260/// migrate drops the index inside the active schema).
1261fn drop_index_stmt(name: &str) -> String {
1262 let n = name.replace('"', "\"\"");
1263 format!("DROP INDEX IF EXISTS \"{n}\"")
1264}
1265
1266/// Lower an M2M junction column's PK type into the SQLite column
1267/// declaration string used inside the raw `CREATE TABLE` template.
1268/// SQLite has affinity types: every integer width stores as `INTEGER`
1269/// (one ROWID-aliased column), and TEXT covers `String` / `Uuid`.
1270/// Closes BUG-16 phase 2.
1271fn m2m_pk_sql_type_sqlite(ty: crate::orm::SqlType) -> &'static str {
1272 use crate::orm::SqlType;
1273 match ty {
1274 SqlType::SmallInt | SqlType::Integer | SqlType::BigInt | SqlType::ForeignKey => "INTEGER",
1275 // A junction's `child_id` must be declared as whatever the CHILD's primary key
1276 // actually is, or the `REFERENCES` clause points at a column of a different type.
1277 // sqlx stores a `Uuid` as its 16 raw bytes on SQLite, so that is BLOB (gaps3 #80).
1278 SqlType::Uuid => "BLOB",
1279 SqlType::Text => "TEXT",
1280 // The macro-side classifier only sets these for PK columns
1281 // when the user wrote a non-standard PK type. If we ever
1282 // see one here that doesn't make sense as a junction column
1283 // (Boolean, Date, Real, …), TEXT is the safest catch-all
1284 // affinity — SQLite will accept it and the rest of the
1285 // ORM will surface the deeper "this can't be a PK" error
1286 // through the system check.
1287 _ => "TEXT",
1288 }
1289}
1290
1291/// Lower an M2M junction column's PK type into the Postgres column
1292/// declaration string. Postgres is strict about types — `BIGINT` for
1293/// 64-bit integers, `INTEGER` for 32-bit, `SMALLINT` for 16-bit,
1294/// `TEXT` for `String`, `UUID` for `uuid::Uuid`. Mirrors the choices
1295/// `build_column_def_postgres` makes for the same `SqlType` variants.
1296fn m2m_pk_sql_type_postgres(ty: crate::orm::SqlType) -> &'static str {
1297 use crate::orm::SqlType;
1298 match ty {
1299 SqlType::SmallInt => "SMALLINT",
1300 SqlType::Integer => "INTEGER",
1301 SqlType::BigInt | SqlType::ForeignKey => "BIGINT",
1302 SqlType::Text => "TEXT",
1303 SqlType::Uuid => "UUID",
1304 _ => "TEXT",
1305 }
1306}
1307
1308/// Build the ` ON DELETE <action> ON UPDATE <action>` suffix for a
1309/// FK column. Each half is emitted only when its action is anything
1310/// other than `NoAction` — keeps the generated DDL minimal and
1311/// matches the SQL standard's default (NO ACTION when the clause is
1312/// omitted).
1313///
1314/// Closes gap #68. Shared between the SQLite and Postgres builders
1315/// because the REFERENCES tail syntax is identical on both.
1316fn fk_action_suffix(col: &Column) -> String {
1317 let mut s = String::new();
1318 if let Some(kw) = col.on_delete.sql_keyword() {
1319 s.push_str(" ON DELETE ");
1320 s.push_str(kw);
1321 }
1322 if let Some(kw) = col.on_update.sql_keyword() {
1323 s.push_str(" ON UPDATE ");
1324 s.push_str(kw);
1325 }
1326 s
1327}
1328
1329/// Suffix that makes a Postgres FK constraint `DEFERRABLE INITIALLY IMMEDIATE`:
1330/// checked per-statement by default (identical behaviour to a plain FK), but
1331/// allowing `SET CONSTRAINTS ALL DEFERRED` inside a transaction to hold the
1332/// check until commit. The data-transfer engine relies on this to load
1333/// mutually-referential (cyclic) tables on Postgres. SQLite needs no equivalent
1334/// — its `PRAGMA defer_foreign_keys` defers any FK regardless of declaration.
1335const PG_FK_DEFERRABLE: &str = " DEFERRABLE INITIALLY IMMEDIATE";
1336
1337fn is_false(b: &bool) -> bool {
1338 !*b
1339}
1340
1341/// serde default for `Column::db_constraint`: a FK emits its physical
1342/// `REFERENCES` constraint unless the model opts out. Older migration
1343/// JSON predating gaps2 #22 has no `db_constraint` key, so it must
1344/// deserialize as `true` to preserve the historical "always emit"
1345/// behaviour.
1346fn default_true() -> bool {
1347 true
1348}
1349
1350fn is_true(b: &bool) -> bool {
1351 *b
1352}
1353
1354impl From<&FieldSpec> for Column {
1355 fn from(f: &FieldSpec) -> Self {
1356 Self {
1357 name: f.name.to_string(),
1358 ty: f.ty,
1359 primary_key: f.primary_key,
1360 nullable: f.nullable,
1361 fk_target: f.fk_target.map(|s| s.to_string()),
1362 noform: f.noform,
1363 privileged: f.privileged,
1364 private: f.private,
1365 secret: f.secret,
1366 db_constraint: f.db_constraint,
1367 noedit: f.noedit,
1368 is_string_repr: f.is_string_repr,
1369 max_length: f.max_length,
1370 choices: f.choices.iter().map(|s| s.to_string()).collect(),
1371 choice_labels: f.choice_labels.iter().map(|s| s.to_string()).collect(),
1372 default: f.default.to_string(),
1373 is_multichoice: f.is_multichoice,
1374 unique: f.unique,
1375 on_delete: f.on_delete,
1376 on_update: f.on_update,
1377 index: f.index,
1378 auto_now_add: f.auto_now_add,
1379 auto_uuid: f.auto_uuid,
1380 auto_user_add: f.auto_user_add,
1381 auto_user: f.auto_user,
1382 auto_now: f.auto_now,
1383 trim: f.trim,
1384 lowercase: f.lowercase,
1385 case_insensitive: f.case_insensitive,
1386 help: f.help.to_string(),
1387 example: f.example.to_string(),
1388 widget: f.widget.map(|s| s.to_string()),
1389 supported_backends: f.supported_backends.iter().map(|s| s.to_string()).collect(),
1390 min: f.min,
1391 max: f.max,
1392 text_format: f.text_format.map(|s| s.to_string()),
1393 slug_from: f.slug_from.map(|s| s.to_string()),
1394 }
1395 }
1396}
1397
1398/// The on-disk shape of one migration. Files in `migrations/<plugin>/`
1399/// deserialize into this struct.
1400#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
1401pub struct MigrationFile {
1402 /// Stable id, matches the filename minus `.json`.
1403 pub id: String,
1404 /// The plugin that owns this migration. M5 hardcodes `"app"` for
1405 /// the user's binary; M7 generalises to one directory per plugin.
1406 pub plugin: String,
1407 /// Predecessor migrations, in `(plugin, id)` form. Within-plugin
1408 /// predecessors are implicit (the prior numeric file); cross-
1409 /// plugin predecessors land at M7.
1410 #[serde(default)]
1411 pub depends_on: Vec<MigrationRef>,
1412 /// Ordered operations applied when this migration runs.
1413 pub operations: Vec<Operation>,
1414 /// The full snapshot of every model after this migration has run.
1415 /// Source of truth for the next `make` to diff against.
1416 pub snapshot_after: Snapshot,
1417 /// gaps2 #100: when non-empty, this is a *squash* — it collapses the
1418 /// listed predecessor migrations into one optimized file. The originals
1419 /// are kept on disk (non-destructive) so older deploys still migrate; the
1420 /// runner treats the squash and its replaced set as mutually exclusive
1421 /// (see [`squash_plan`]). Empty for an ordinary migration; `#[serde(default)]`
1422 /// so every pre-squash file on disk deserializes unchanged.
1423 #[serde(default, skip_serializing_if = "Vec::is_empty")]
1424 pub replaces: Vec<MigrationRef>,
1425}
1426
1427/// A pointer to one (plugin, migration_id) pair.
1428#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
1429pub struct MigrationRef {
1430 pub plugin: String,
1431 pub migration: String,
1432}
1433
1434/// gaps2 #100 — what the runner does with one on-disk migration file, given the
1435/// tracking-table's applied-set. The decision is pure and backend-agnostic; the
1436/// only DB-specific step downstream is *how* a tracking row is written, never
1437/// *whether* the file's operations run.
1438#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1439pub enum ApplyDecision {
1440 /// Run the file's operations, then record it in the tracking table. The
1441 /// normal "pending migration" path.
1442 Apply,
1443 /// Do nothing: the migration is already applied, or it's an original that a
1444 /// squash the runner is using has shadowed (running it too would
1445 /// double-apply the schema the squash already built).
1446 Skip,
1447 /// A squash whose entire replaced set is already applied. Its operations
1448 /// would rebuild schema that already exists, so DON'T run them — but insert
1449 /// a tracking row so future runs treat the squash as applied and the now
1450 /// redundant original files can be deleted.
1451 RecordOnly,
1452}
1453
1454/// gaps2 #100 — decide, per on-disk migration file for ONE plugin, whether to
1455/// [`Apply`](ApplyDecision::Apply) / [`Skip`](ApplyDecision::Skip) /
1456/// [`RecordOnly`](ApplyDecision::RecordOnly) it, honoring squash `replaces`.
1457///
1458/// `files` is every migration file for the plugin (the returned Vec is in the
1459/// same order). `applied` is the `(plugin, id)` set from the tracking table.
1460/// The function is pure: no DB, no IO — so both the SQLite and Postgres apply
1461/// loops route through it and inherit identical squash semantics.
1462///
1463/// For a squash `S` replacing set `R`:
1464/// - **all of `R` applied** → `S` is `RecordOnly` (or `Skip` if `S` itself is
1465/// already recorded); every member of `R` is `Skip` (already applied).
1466/// - **none of `R` applied** → `S` is `Apply` (or `Skip` if already recorded);
1467/// every member of `R` is `Skip` (the squash builds their schema).
1468/// - **some of `R` applied** (an interrupted transition) → fall back to the
1469/// individual originals if they are ALL still on disk: `S` is `Skip` and each
1470/// unapplied member of `R` applies as usual. If any original is gone, the
1471/// history can't be reconciled → [`MigrateError::SquashInconsistent`].
1472///
1473/// A plain migration (empty `replaces`) is `Skip` if applied or shadowed by an
1474/// active squash, else `Apply`.
1475fn squash_plan(
1476 files: &[MigrationFile],
1477 applied: &std::collections::HashSet<(String, String)>,
1478) -> Result<Vec<ApplyDecision>, MigrateError> {
1479 use std::collections::HashSet;
1480
1481 let is_applied =
1482 |plugin: &str, id: &str| applied.contains(&(plugin.to_string(), id.to_string()));
1483 let on_disk: HashSet<&str> = files.iter().map(|f| f.id.as_str()).collect();
1484
1485 let mut decisions = vec![ApplyDecision::Apply; files.len()];
1486 // Ids covered by a squash the runner is USING (fully-applied or fully-
1487 // unapplied). Their originals must be skipped so they don't double-apply.
1488 let mut shadowed: HashSet<String> = HashSet::new();
1489
1490 // First pass: classify every squash and collect the shadow set.
1491 for (i, f) in files.iter().enumerate() {
1492 if f.replaces.is_empty() {
1493 continue;
1494 }
1495 let applied_in_r = f
1496 .replaces
1497 .iter()
1498 .filter(|m| is_applied(&m.plugin, &m.migration))
1499 .count();
1500 let self_applied = is_applied(&f.plugin, &f.id);
1501
1502 if applied_in_r == f.replaces.len() {
1503 // Whole history already applied: adopt the squash without running it.
1504 decisions[i] = if self_applied {
1505 ApplyDecision::Skip
1506 } else {
1507 ApplyDecision::RecordOnly
1508 };
1509 for m in &f.replaces {
1510 shadowed.insert(m.migration.clone());
1511 }
1512 } else if applied_in_r == 0 {
1513 // Fresh: the squash builds the whole schema in one shot.
1514 decisions[i] = if self_applied {
1515 ApplyDecision::Skip
1516 } else {
1517 ApplyDecision::Apply
1518 };
1519 for m in &f.replaces {
1520 shadowed.insert(m.migration.clone());
1521 }
1522 } else {
1523 // Partial transition: prefer the originals if they all survive.
1524 let missing: Vec<String> = f
1525 .replaces
1526 .iter()
1527 .filter(|m| !on_disk.contains(m.migration.as_str()))
1528 .map(|m| format!("{}/{}", m.plugin, m.migration))
1529 .collect();
1530 if missing.is_empty() {
1531 // Inactive squash: skip it, let the originals run individually.
1532 decisions[i] = ApplyDecision::Skip;
1533 } else {
1534 return Err(MigrateError::SquashInconsistent {
1535 plugin: f.plugin.clone(),
1536 squash: f.id.clone(),
1537 missing,
1538 });
1539 }
1540 }
1541 }
1542
1543 // Second pass: plain migrations (applied or shadowed → Skip, else Apply).
1544 for (i, f) in files.iter().enumerate() {
1545 if !f.replaces.is_empty() {
1546 continue;
1547 }
1548 if shadowed.contains(&f.id) || is_applied(&f.plugin, &f.id) {
1549 decisions[i] = ApplyDecision::Skip;
1550 }
1551 }
1552
1553 // A squash can itself be shadowed by a larger squash (squash-of-squashes):
1554 // an active outer squash covers it, so it must not run either.
1555 for (i, f) in files.iter().enumerate() {
1556 if shadowed.contains(&f.id) {
1557 decisions[i] = ApplyDecision::Skip;
1558 }
1559 }
1560
1561 Ok(decisions)
1562}
1563
1564/// At M5 every migration belongs to a single placeholder plugin. M7's
1565/// Plugin contract replaces this with `Plugin::name()`.
1566pub const APP_PLUGIN_NAME: &str = "app";
1567
1568/// Default directory for migration files. `make` writes into
1569/// `migrations/<plugin>/`; `run` reads from the same place. Override
1570/// with `--migrations-dir` once the CLI grows real arg parsing (M5+).
1571pub const MIGRATIONS_DIR: &str = "migrations";
1572
1573/// The state of a single migration from the perspective of drift detection.
1574/// Returned inside [`DriftReport`] so callers can decide how to handle each
1575/// state independently.
1576#[derive(Debug, Clone, PartialEq, Eq)]
1577pub enum MigrationStatus {
1578 /// The migration is recorded in the tracking table AND the file
1579 /// exists on disk. Normal applied state.
1580 Applied,
1581 /// The migration is recorded in the tracking table BUT the
1582 /// corresponding file is missing from disk. The database is ahead
1583 /// of what version control has; recovering requires restoring the
1584 /// file or running with `--allow-drift`.
1585 AppliedButMissing,
1586 /// The migration file exists on disk AND its sequence number is
1587 /// lower than the highest applied migration for this plugin, but it
1588 /// is not recorded in the tracking table. Looks like someone dropped
1589 /// a migration file back into a directory after a teammate already
1590 /// applied later ones. Should warn, not error.
1591 OutOfOrder,
1592 /// Normal pending state: the file is on disk and its sequence number
1593 /// is higher than anything applied. Ready to apply.
1594 Pending,
1595}
1596
1597/// Per-migration entry inside a [`DriftReport`].
1598#[derive(Debug, Clone, PartialEq, Eq)]
1599pub struct MigrationEntry {
1600 pub plugin: String,
1601 pub name: String,
1602 pub status: MigrationStatus,
1603}
1604
1605/// The output of [`detect_drift`]: one entry per migration (applied or
1606/// on-disk), categorised into the four states above.
1607///
1608/// The caller inspects `has_critical_drift()` to decide whether to abort
1609/// before applying migrations. Surfaced by `show_in_with_drift` for
1610/// `showmigrations` and checked by `run_in_with_drift_check` before
1611/// executing any SQL.
1612#[derive(Debug, Clone, Default)]
1613pub struct DriftReport {
1614 pub entries: Vec<MigrationEntry>,
1615}
1616
1617impl DriftReport {
1618 /// Returns true when at least one migration is `AppliedButMissing`.
1619 /// This state means the tracking table references a file that no
1620 /// longer exists on disk — the operator needs to act before it is
1621 /// safe to continue applying new migrations.
1622 pub fn has_critical_drift(&self) -> bool {
1623 self.entries
1624 .iter()
1625 .any(|e| e.status == MigrationStatus::AppliedButMissing)
1626 }
1627
1628 /// All migrations with `AppliedButMissing` status. Convenience
1629 /// accessor for building the error message.
1630 pub fn missing_on_disk(&self) -> Vec<&MigrationEntry> {
1631 self.entries
1632 .iter()
1633 .filter(|e| e.status == MigrationStatus::AppliedButMissing)
1634 .collect()
1635 }
1636
1637 /// Migrations on disk this binary knows about that the database has NOT
1638 /// applied yet — the readiness blocker (Kikosi #5 / gaps3 #38). Empty means
1639 /// the schema is at least as new as this binary's code.
1640 ///
1641 /// Deliberately excludes `AppliedButMissing` (the database is *ahead*: a
1642 /// valid rollback / backward-compatible-migration state where an older
1643 /// binary should keep serving) and `OutOfOrder` (a stray file the migrate
1644 /// engine only warns about). Only `Pending` — "the schema is behind the
1645 /// code" — should hold traffic off a freshly-booted pod.
1646 pub fn pending(&self) -> Vec<&MigrationEntry> {
1647 self.entries
1648 .iter()
1649 .filter(|e| e.status == MigrationStatus::Pending)
1650 .collect()
1651 }
1652}
1653
1654/// Errors the migration engine can produce.
1655#[derive(Debug)]
1656pub enum MigrateError {
1657 /// IO error reading or writing a migration file or directory.
1658 Io(std::io::Error),
1659 /// JSON parse error on a migration file.
1660 Json(serde_json::Error),
1661 /// sqlx error executing a migration's SQL or touching the
1662 /// tracking table.
1663 Sqlx(sqlx::Error),
1664 /// `make` ran but found no differences against the latest snapshot,
1665 /// so there's nothing to write. Surfaced so the CLI can print
1666 /// "no changes detected" instead of an empty migration file.
1667 NoChanges,
1668 /// The current models diverge from the snapshot in a way M5 v1
1669 /// can't represent yet (anything other than create/drop table).
1670 /// M5.1 lifts this when column-level ops land.
1671 UnsupportedChange(String),
1672 /// A column-level change the engine can't apply automatically:
1673 /// type change, or a nullable flip on a populated SQLite table.
1674 /// Surfaces from `diff` so the build stops before producing a
1675 /// migration that would lose data or fail to apply. The user
1676 /// resolves by hand-writing the migration with the appropriate
1677 /// data-preserving steps. Carries the model / column / reason.
1678 UnsafeAlter {
1679 model: String,
1680 column: String,
1681 reason: String,
1682 },
1683 /// The tracking table records migrations that no longer have
1684 /// corresponding files on disk. Carries the list of missing names.
1685 /// The operator must either restore the files from VCS or run with
1686 /// `--allow-drift` to proceed despite the inconsistency.
1687 DriftDetected { missing: Vec<(String, String)> },
1688 /// A schema-scoped migration ([`run_for_schema`]) was requested against a
1689 /// SQLite pool. SQLite has no schemas, so schema-per-tenant is Postgres-only
1690 /// (mirrors how `Inet`/`Cidr` gate on backend). Carries the schema name.
1691 SchemaUnsupportedOnSqlite { schema: String },
1692 /// `makemigrations --empty <plugin>` named a plugin that isn't
1693 /// registered. Carries the requested name and the registered set so
1694 /// the CLI can list the valid choices.
1695 UnknownPlugin {
1696 requested: String,
1697 known: Vec<String>,
1698 },
1699 /// audit_2 H23 — the column-shape rename heuristic found an unpaired
1700 /// dropped model and an unpaired created model with **identical** column
1701 /// shapes. That's genuinely ambiguous: it's either a model rename (move the
1702 /// old table's rows to the new name) or two unrelated models that happen to
1703 /// share a shape (drop the old, create the new empty). Auto-applying either
1704 /// silently loses or mis-associates data, so `diff` refuses to guess and
1705 /// fails closed. The operator resolves it explicitly via
1706 /// `UMBRAL_MIGRATIONS_ASSUME_RENAMES` (`assume` → rename, `independent` →
1707 /// drop+create) or by hand-writing the op. Carries the two table names.
1708 AmbiguousRename {
1709 from_table: String,
1710 to_table: String,
1711 },
1712 /// audit_2 core-migrate #7 — couldn't acquire the Postgres migration
1713 /// advisory lock within the timeout: another process has been holding it
1714 /// (running a long migration, or wedged). Carries the alias/schema the lock
1715 /// was keyed on and the seconds waited. The operator retries once the other
1716 /// migrator finishes, or investigates a stuck migration.
1717 MigrationLockTimeout {
1718 discriminator: String,
1719 waited_secs: u64,
1720 },
1721 /// gaps2 #100 — a squash migration is stuck in a partially-applied state:
1722 /// SOME of the migrations it replaces are recorded as applied and others
1723 /// aren't, and the unapplied originals are no longer on disk to fall back
1724 /// to. The engine can't safely apply the squash (it would re-run schema the
1725 /// applied originals already built) nor complete the originals (their files
1726 /// are gone). Carries the plugin and squash id. The operator restores the
1727 /// missing original files from VCS (finish the transition), or resets the
1728 /// tracking rows for this plugin to a consistent point.
1729 SquashInconsistent {
1730 plugin: String,
1731 squash: String,
1732 missing: Vec<String>,
1733 },
1734 /// gaps2 #100 — `squashmigrations <plugin>` couldn't produce a squash:
1735 /// fewer than two squashable migrations, an unknown plugin, or the history
1736 /// already contains a squash (nested squashing isn't supported yet). Carries
1737 /// the plugin and a human-readable reason.
1738 CannotSquash { plugin: String, reason: String },
1739}
1740
1741impl std::fmt::Display for MigrateError {
1742 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1743 match self {
1744 MigrateError::Io(e) => write!(f, "umbral migrate: io: {e}"),
1745 MigrateError::Json(e) => write!(f, "umbral migrate: json: {e}"),
1746 MigrateError::Sqlx(e) => write!(f, "umbral migrate: sqlx: {e}"),
1747 MigrateError::NoChanges => write!(
1748 f,
1749 "umbral migrate: no changes detected; declare or change a model first"
1750 ),
1751 MigrateError::UnsupportedChange(msg) => {
1752 write!(f, "umbral migrate: unsupported change at M5 v1: {msg}")
1753 }
1754 MigrateError::UnsafeAlter {
1755 model,
1756 column,
1757 reason,
1758 } => write!(
1759 f,
1760 "umbral migrate: unsafe column change on `{model}.{column}`: {reason}; \
1761 hand-write the migration with a data-preserving step"
1762 ),
1763 MigrateError::DriftDetected { missing } => {
1764 let names: Vec<String> = missing
1765 .iter()
1766 .map(|(plugin, name)| format!("{plugin}/{name}"))
1767 .collect();
1768 write!(
1769 f,
1770 "umbral migrate: drift detected — the following migrations are recorded in \
1771 the tracking table but their files are missing from disk:\n {}\n\
1772 Restore the files from VCS or run `umbral migrate --allow-drift` to \
1773 proceed despite the inconsistency.",
1774 names.join("\n ")
1775 )
1776 }
1777 MigrateError::SchemaUnsupportedOnSqlite { schema } => write!(
1778 f,
1779 "umbral migrate: schema-per-tenant migration into `{schema}` requires \
1780 Postgres; SQLite has no schemas. Point the app at a Postgres pool."
1781 ),
1782 MigrateError::UnknownPlugin { requested, known } => write!(
1783 f,
1784 "umbral makemigrations --empty: no registered plugin named `{requested}`. \
1785 Known plugins: {}",
1786 known.join(", ")
1787 ),
1788 MigrateError::AmbiguousRename {
1789 from_table,
1790 to_table,
1791 } => write!(
1792 f,
1793 "umbral makemigrations: ambiguous rename — the dropped model `{from_table}` and \
1794 the new model `{to_table}` have identical column shapes, so this is either a \
1795 rename (move `{from_table}`'s rows to `{to_table}`) or two unrelated models. \
1796 Refusing to guess: auto-renaming would hand one model's rows to another and skip \
1797 the intended drop, while auto-dropping would delete `{from_table}`'s rows — both \
1798 silent data bugs. Resolve it: set UMBRAL_MIGRATIONS_ASSUME_RENAMES=assume to \
1799 treat every shape match as a rename, or =independent to treat them as unrelated \
1800 (drop + create), or hand-write the intended op into the migration file."
1801 ),
1802 MigrateError::MigrationLockTimeout {
1803 discriminator,
1804 waited_secs,
1805 } => write!(
1806 f,
1807 "umbral migrate: timed out after {waited_secs}s waiting for the Postgres \
1808 migration lock (alias/schema `{discriminator}`). Another process is holding it — \
1809 a long-running migration on another replica, or a wedged migrator. Retry once it \
1810 finishes; if nothing is migrating, check for a stuck backend holding \
1811 pg_advisory_lock."
1812 ),
1813 MigrateError::SquashInconsistent {
1814 plugin,
1815 squash,
1816 missing,
1817 } => write!(
1818 f,
1819 "umbral migrate: squash `{plugin}/{squash}` is half-applied — some of the \
1820 migrations it replaces are recorded as applied and the unapplied ones are \
1821 missing from disk:\n {}\nRestore those original migration files from VCS so \
1822 the transition can finish, then re-run migrate.",
1823 missing.join("\n ")
1824 ),
1825 MigrateError::CannotSquash { plugin, reason } => write!(
1826 f,
1827 "umbral squashmigrations: can't squash `{plugin}`: {reason}"
1828 ),
1829 }
1830 }
1831}
1832
1833impl std::error::Error for MigrateError {}
1834
1835impl From<std::io::Error> for MigrateError {
1836 fn from(e: std::io::Error) -> Self {
1837 Self::Io(e)
1838 }
1839}
1840
1841impl From<serde_json::Error> for MigrateError {
1842 fn from(e: serde_json::Error) -> Self {
1843 Self::Json(e)
1844 }
1845}
1846
1847impl From<sqlx::Error> for MigrateError {
1848 fn from(e: sqlx::Error) -> Self {
1849 Self::Sqlx(e)
1850 }
1851}
1852
1853// =========================================================================
1854// Top-level entry points.
1855// =========================================================================
1856
1857/// Generate one migration file per registered plugin that has changes,
1858/// diffing each plugin's current model set against the latest snapshot
1859/// in `migrations/<plugin>/`. Each new file lands inside its own
1860/// plugin directory with the next sequence number and a `_<short_name>`
1861/// suffix derived from the dominant operation.
1862///
1863/// Returns the paths of every file written, one per plugin that had a
1864/// non-empty diff. Returns `MigrateError::NoChanges` if no plugin
1865/// produced any changes at all.
1866pub async fn make() -> Result<Vec<PathBuf>, MigrateError> {
1867 make_in(Path::new(MIGRATIONS_DIR)).await
1868}
1869
1870/// Same as [`make`] but takes an explicit base directory. Used by
1871/// tests to avoid touching the cwd.
1872///
1873/// Iterates [`plugin_order`], which is the topological order
1874/// published by `App::build()`'s phase 1.5 sort. Cross-plugin FKs
1875/// land in dependency order this way (a plugin's `CreateTable` for
1876/// the FK target runs before the dependent plugin's `CreateTable`).
1877/// Falls back to [`registered_plugins`] when no order has been
1878/// published (e.g. low-level tests that init the registry directly).
1879pub async fn make_in(dir: &Path) -> Result<Vec<PathBuf>, MigrateError> {
1880 let mut written: Vec<PathBuf> = Vec::new();
1881
1882 for plugin in plugin_order() {
1883 let plugin_dir = dir.join(&plugin);
1884
1885 // The previous snapshot is the `snapshot_after` of the highest-
1886 // numbered migration file (filenames are zero-padded so lexical
1887 // sort matches numeric order). An empty or missing directory
1888 // means "no prior state", the first-run case for this plugin.
1889 let existing = list_migration_files(&plugin_dir)?;
1890 let previous = match existing.last() {
1891 Some(path) => read_migration_file(path)?.snapshot_after,
1892 None => Snapshot::default(),
1893 };
1894
1895 let current = Snapshot::current_for(&plugin);
1896 let operations = diff(&previous, ¤t)?;
1897 if operations.is_empty() {
1898 continue;
1899 }
1900
1901 let seq = (existing.len() + 1) as u32;
1902 let suffix = suffix_for(&operations);
1903 let id = format!("{seq:04}_{suffix}");
1904 let filename = format!("{id}.json");
1905
1906 let file = MigrationFile {
1907 id: id.clone(),
1908 plugin: plugin.clone(),
1909 depends_on: Vec::new(),
1910 operations,
1911 snapshot_after: current,
1912 replaces: Vec::new(),
1913 };
1914
1915 std::fs::create_dir_all(&plugin_dir)?;
1916 let path = plugin_dir.join(filename);
1917 let json = serde_json::to_string_pretty(&file)?;
1918 std::fs::write(&path, json)?;
1919 written.push(path);
1920 }
1921
1922 if written.is_empty() {
1923 return Err(MigrateError::NoChanges);
1924 }
1925 Ok(written)
1926}
1927
1928/// Write an **empty** migration for one plugin: the current snapshot
1929/// with an empty `operations` list, the authoring stub for a
1930/// hand-written data migration (`Operation::RunSql`). The developer
1931/// opens the file and adds a `RunSql { sql, reverse_sql }` op.
1932///
1933/// The empty op-list means `snapshot_after == snapshot_before`, so the
1934/// next `make` diffs against the same state and produces nothing — a
1935/// data migration never disturbs the schema-snapshot chain. Mirror of
1936/// [`make`] for the `--empty <plugin>` CLI path.
1937pub async fn make_empty(plugin: &str) -> Result<PathBuf, MigrateError> {
1938 make_empty_in(Path::new(MIGRATIONS_DIR), plugin).await
1939}
1940
1941/// Same as [`make_empty`] but takes an explicit base directory. The
1942/// seam tests drive.
1943pub async fn make_empty_in(dir: &Path, plugin: &str) -> Result<PathBuf, MigrateError> {
1944 // The plugin must be registered, else the snapshot/sequence would be
1945 // meaningless. Fail loudly with the known set.
1946 let known = plugin_order();
1947 if !known.iter().any(|p| p == plugin) {
1948 return Err(MigrateError::UnknownPlugin {
1949 requested: plugin.to_string(),
1950 known,
1951 });
1952 }
1953
1954 let plugin_dir = dir.join(plugin);
1955
1956 // Carry the latest snapshot forward verbatim: an empty migration has
1957 // NO schema effect, so `snapshot_after` equals the previous one. The
1958 // current model snapshot is the same as the prior file's
1959 // `snapshot_after` (no model changed); use the current registry state
1960 // so the file is self-consistent even on a plugin's very first
1961 // migration.
1962 let existing = list_migration_files(&plugin_dir)?;
1963 let snapshot = match existing.last() {
1964 Some(path) => read_migration_file(path)?.snapshot_after,
1965 None => Snapshot::current_for(plugin),
1966 };
1967
1968 let seq = (existing.len() + 1) as u32;
1969 let id = format!("{seq:04}_empty");
1970 let filename = format!("{id}.json");
1971
1972 let file = MigrationFile {
1973 id: id.clone(),
1974 plugin: plugin.to_string(),
1975 depends_on: Vec::new(),
1976 operations: Vec::new(),
1977 snapshot_after: snapshot,
1978 replaces: Vec::new(),
1979 };
1980
1981 std::fs::create_dir_all(&plugin_dir)?;
1982 let path = plugin_dir.join(filename);
1983 let json = serde_json::to_string_pretty(&file)?;
1984 std::fs::write(&path, json)?;
1985 Ok(path)
1986}
1987
1988/// The numeric sequence prefix of a migration id (`0003_add_x` → `0003`),
1989/// used to name a squash `0001_squashed_0003`.
1990fn seq_prefix(id: &str) -> &str {
1991 id.split('_').next().unwrap_or(id)
1992}
1993
1994/// Outcome of [`squash_in`]: where the squash landed and what it collapsed.
1995pub struct SquashOutcome {
1996 /// Path of the written squash file.
1997 pub path: PathBuf,
1998 /// Id of the new squash migration (e.g. `0001_squashed_0005`).
1999 pub id: String,
2000 /// Ids of the original migrations it replaces — kept on disk, non-destructive.
2001 pub replaced: Vec<String>,
2002}
2003
2004/// gaps2 #100 — collapse a plugin's entire linear migration history into ONE
2005/// optimized squash file, **non-destructively**: the originals stay on disk so
2006/// older deploys still migrate, and the runner ([`squash_plan`]) treats the
2007/// squash and its originals as mutually exclusive.
2008///
2009/// The squash's operations are the diff from an empty schema to the last
2010/// migration's `snapshot_after` — i.e. one `CreateTable` per model with every
2011/// intermediate `AlterColumn` already folded into the final column set. That's
2012/// the minimal replay of the whole history, which is the entire point (a 20-file
2013/// history with repeated alters collapses to N clean creates).
2014///
2015/// Refuses (with [`MigrateError::CannotSquash`]) when it can't produce a safe
2016/// squash: fewer than two migrations, a history that already contains a squash
2017/// (nested squashing is out of scope), or a history containing a `RunSql` data
2018/// migration — a snapshot diff can't see hand-written data steps, so squashing
2019/// across one would silently drop it. The operator squashes the schema-only
2020/// prefix, or leaves that history intact.
2021pub fn squash_in(dir: &Path, plugin: &str) -> Result<SquashOutcome, MigrateError> {
2022 let plugin_dir = dir.join(plugin);
2023 let paths = list_migration_files(&plugin_dir)?;
2024 let files: Vec<MigrationFile> = paths
2025 .iter()
2026 .map(|p| read_migration_file(p))
2027 .collect::<Result<_, _>>()?;
2028
2029 if files.len() < 2 {
2030 return Err(MigrateError::CannotSquash {
2031 plugin: plugin.to_string(),
2032 reason: format!(
2033 "need at least 2 migrations to squash, found {}",
2034 files.len()
2035 ),
2036 });
2037 }
2038 if files.iter().any(|f| !f.replaces.is_empty()) {
2039 return Err(MigrateError::CannotSquash {
2040 plugin: plugin.to_string(),
2041 reason: "this history already contains a squash; nested squashing isn't supported \
2042 yet. Delete the now-redundant original files once every deploy has migrated, \
2043 then squash again"
2044 .to_string(),
2045 });
2046 }
2047 if files.iter().any(|f| {
2048 f.operations
2049 .iter()
2050 .any(|op| matches!(op, Operation::RunSql { .. }))
2051 }) {
2052 return Err(MigrateError::CannotSquash {
2053 plugin: plugin.to_string(),
2054 reason: "this history contains a RunSql data migration, which a snapshot diff can't \
2055 reconstruct — squashing across it would silently drop the data step. Squash \
2056 the schema-only migrations before/after it, or leave this history intact"
2057 .to_string(),
2058 });
2059 }
2060
2061 let first = &files[0];
2062 let last = files.last().expect("len >= 2 checked above");
2063
2064 // Optimal from-scratch replay: empty schema → final snapshot. Every
2065 // intermediate alter is already baked into `snapshot_after`.
2066 let empty = Snapshot { models: Vec::new() };
2067 let operations = diff(&empty, &last.snapshot_after)?;
2068
2069 let id = format!(
2070 "{}_squashed_{}",
2071 seq_prefix(&first.id),
2072 seq_prefix(&last.id)
2073 );
2074 let replaced: Vec<String> = files.iter().map(|f| f.id.clone()).collect();
2075 let squash = MigrationFile {
2076 id: id.clone(),
2077 plugin: plugin.to_string(),
2078 // Preserve the first migration's cross-plugin predecessors: the squash
2079 // stands in for the whole run, so it inherits the run's external deps.
2080 depends_on: first.depends_on.clone(),
2081 operations,
2082 snapshot_after: last.snapshot_after.clone(),
2083 replaces: replaced
2084 .iter()
2085 .map(|mid| MigrationRef {
2086 plugin: plugin.to_string(),
2087 migration: mid.clone(),
2088 })
2089 .collect(),
2090 };
2091
2092 let path = plugin_dir.join(format!("{id}.json"));
2093 let json = serde_json::to_string_pretty(&squash)?;
2094 std::fs::write(&path, json)?;
2095 Ok(SquashOutcome { path, id, replaced })
2096}
2097
2098/// Create a table for every registered model, deriving the schema from the
2099/// models themselves. **For tests.**
2100///
2101/// This is the migration engine pointed at an empty database: diff `Snapshot`
2102/// nothing against [`Snapshot::current`], render the resulting `CreateTable`
2103/// operations for the ambient backend, and run them. Same `ModelMeta` the ORM
2104/// reads, same renderer [`run`] uses — so the schema it produces cannot
2105/// disagree with your models.
2106///
2107/// # Why this exists
2108///
2109/// The alternative, and what most test suites reach for, is a hand-written
2110/// `CREATE TABLE` in the test file. That is a **second source of truth** for the
2111/// schema, and it drifts: add a column to the model and the test's table
2112/// silently lacks it. The ORM then queries a column that does not exist, and any
2113/// error-swallowing on the path (an `unwrap_or(false)`) turns that failure into a
2114/// test that *passes with the wrong answer* rather than failing. That is not a
2115/// hypothetical — a soft-delete column drifting out of a fixture is exactly what
2116/// made an authorization test report that a user was already a moderator when
2117/// they were not.
2118///
2119/// # Not a substitute for `migrate`
2120///
2121/// It writes no ledger row and runs no migration file. It cannot ALTER an
2122/// existing table, so it will not carry a database from one shape to another —
2123/// which is the entire job of [`run`] and the reason `makemigrations` exists.
2124/// Point it at a throwaway database, never a real one.
2125///
2126/// Idempotent: an "already exists" error is skipped, so a shared test boot that
2127/// runs twice is harmless. Returns the number of statements executed.
2128pub async fn create_tables_for_tests() -> Result<u64, MigrateError> {
2129 create_tables_for_tests_on("default").await
2130}
2131
2132/// [`create_tables_for_tests`], against a named database alias.
2133///
2134/// A read/write-split or multi-database test registers more than one pool
2135/// (`.database("default", ..).database("replica", ..)`) and needs the schema on
2136/// each of them — a replica with no tables is not a replica. The ambient-pool
2137/// variant only ever reaches `"default"`.
2138pub async fn create_tables_for_tests_on(alias: &str) -> Result<u64, MigrateError> {
2139 let ops = diff(&Snapshot { models: Vec::new() }, &Snapshot::current())?;
2140 let pool = crate::db::pool_for_dispatched(alias);
2141 let backend = match pool {
2142 crate::db::DbPool::Sqlite(_) => "sqlite",
2143 crate::db::DbPool::Postgres(_) => "postgres",
2144 };
2145
2146 let mut executed = 0u64;
2147 for op in &ops {
2148 for sql in render_operation_for(op, backend) {
2149 let res = match crate::db::pool_for_dispatched(alias) {
2150 crate::db::DbPool::Sqlite(p) => sqlx::query(&sql).execute(p).await.map(|_| ()),
2151 crate::db::DbPool::Postgres(p) => sqlx::query(&sql).execute(p).await.map(|_| ()),
2152 };
2153 match res {
2154 Ok(()) => executed += 1,
2155 // A shared `OnceCell` boot may land here twice; a test that dies
2156 // on its own setup is worse than useless.
2157 Err(e) if e.to_string().to_lowercase().contains("already exists") => {}
2158 Err(e) => return Err(MigrateError::Sqlx(e)),
2159 }
2160 }
2161 }
2162 Ok(executed)
2163}
2164
2165/// Apply every pending migration across every registered plugin's
2166/// `migrations/<plugin>/` directory to the ambient pool. Reads the
2167/// `umbral_migrations` tracking table to determine "pending"; each
2168/// migration runs in its own transaction along with its tracking-table
2169/// insert.
2170///
2171/// Returns the total number of migrations applied (zero if every
2172/// plugin's migrations were already in the tracking table).
2173///
2174/// This variant performs a drift check before executing any SQL. If
2175/// any migration is `AppliedButMissing` (in the DB but not on disk),
2176/// the call returns [`MigrateError::DriftDetected`] listing the
2177/// missing names. Pass `allow_drift = true` (via [`run_checked_in`])
2178/// to suppress the error and proceed anyway (with a warning printed to
2179/// stderr).
2180pub async fn run() -> Result<u64, MigrateError> {
2181 run_checked(false).await
2182}
2183
2184/// Same as [`run`] but controls drift handling.
2185/// `allow_drift = true` corresponds to the `--allow-drift` CLI flag:
2186/// the command logs a warning and proceeds even if some applied
2187/// migrations are missing on disk.
2188pub async fn run_checked(allow_drift: bool) -> Result<u64, MigrateError> {
2189 run_checked_in(Path::new(MIGRATIONS_DIR), allow_drift).await
2190}
2191
2192/// Same as [`run_checked`] but takes an explicit base directory.
2193pub async fn run_checked_in(dir: &Path, allow_drift: bool) -> Result<u64, MigrateError> {
2194 let mut total: u64 = 0;
2195 // Walk every registered DB. Drift-detection on the default pool
2196 // is the dominant flow; secondary pools currently use the same
2197 // tracking-table-vs-disk comparison but only against the
2198 // migration files whose ops actually targeted that DB. A future
2199 // pass can teach `detect_all_drift` to be alias-aware so drift
2200 // warnings name the offending pool — today it warns once per
2201 // checked DB if the issue is present in any.
2202 for alias in crate::db::registered_aliases() {
2203 match crate::db::pool_for_dispatched(&alias) {
2204 crate::db::DbPool::Sqlite(p) => {
2205 total += run_in_sqlite_checked(dir, p, allow_drift, &alias).await?
2206 }
2207 crate::db::DbPool::Postgres(p) => {
2208 total += run_in_postgres_checked(dir, p, allow_drift, &alias).await?
2209 }
2210 }
2211 }
2212 Ok(total)
2213}
2214
2215/// Same as [`run`] but takes an explicit base directory. Used by
2216/// tests to avoid touching the cwd.
2217///
2218/// Iterates `registered_plugins()` in sorted-by-name order. M7 v1
2219/// accepts this as a limitation: cross-plugin FK ordering wants
2220/// topological order across plugins (the FK target's `CreateTable`
2221/// applies before the dependent plugin's `CreateTable`), but the
2222/// engine doesn't see `Plugin::dependencies()` from inside this
2223/// standalone function. M8 lifts the limitation via a registry that
2224/// remembers the toposorted order computed at `App::build()` time.
2225///
2226/// This legacy entry point does NOT perform drift checking so the
2227/// existing tests (which bypass drift by design) keep passing. New
2228/// callers should prefer [`run_checked_in`].
2229pub async fn run_in(dir: &Path) -> Result<u64, MigrateError> {
2230 let mut total: u64 = 0;
2231 // Walk every registered DB so each pool gets its own
2232 // `umbral_migrations` table and runs only the operations targeting
2233 // tables routed to it. Order is alphabetical for determinism;
2234 // the "default" pool is always present.
2235 for alias in crate::db::registered_aliases() {
2236 match crate::db::pool_for_dispatched(&alias) {
2237 crate::db::DbPool::Sqlite(p) => {
2238 total += run_in_sqlite_for_alias(dir, &alias, p, None).await?
2239 }
2240 crate::db::DbPool::Postgres(p) => {
2241 total += run_in_postgres_for_alias(dir, &alias, p, None).await?
2242 }
2243 }
2244 }
2245 Ok(total)
2246}
2247
2248/// Apply only the **SHARED** apps' pending migrations to the default pool —
2249/// the `public`/shared half of schema-per-tenant multitenancy. This is the
2250/// mirror of [`run_for_schema_in`] (which migrates the *tenant* apps into a
2251/// tenant schema): here only plugins IN `shared_apps` migrate into `public`,
2252/// so a tenant app's tables — and crucially its M2M junctions — are NEVER
2253/// created in `public`. They live only in each tenant schema, where a junction's
2254/// FK to a SHARED child resolves via the `<schema>, public` search-path.
2255///
2256/// Use this instead of the unfiltered [`run`]/[`run_in`] when running a
2257/// schema-per-tenant app: `run_shared` (shared → public) then `migrate_schemas`
2258/// (tenant apps → each schema). On a non-multitenant app the two are equivalent
2259/// only if every app is shared; otherwise prefer plain [`run`].
2260pub async fn run_shared(
2261 shared_apps: &std::collections::HashSet<String>,
2262) -> Result<u64, MigrateError> {
2263 run_shared_in(Path::new(MIGRATIONS_DIR), shared_apps).await
2264}
2265
2266/// [`run_shared`] against an explicit migrations directory (tests / tooling).
2267pub async fn run_shared_in(
2268 dir: &Path,
2269 shared_apps: &std::collections::HashSet<String>,
2270) -> Result<u64, MigrateError> {
2271 let mut total: u64 = 0;
2272 for alias in crate::db::registered_aliases() {
2273 match crate::db::pool_for_dispatched(&alias) {
2274 crate::db::DbPool::Sqlite(p) => {
2275 total += run_in_sqlite_for_alias(dir, &alias, p, Some(shared_apps)).await?
2276 }
2277 crate::db::DbPool::Postgres(p) => {
2278 total += run_in_postgres_for_alias(dir, &alias, p, Some(shared_apps)).await?
2279 }
2280 }
2281 }
2282 Ok(total)
2283}
2284
2285/// Predicate: does `op` target a table that lives on `alias`?
2286///
2287/// Routing rule: look up the table → alias mapping via
2288/// [`table_alias`]. Tables not owned by any registered model fall
2289/// through to `"default"` so the migration engine's own
2290/// `umbral_migrations` book-keeping stays in the main DB.
2291///
2292/// A second gate consults the installed [`DatabaseRouter`]: if the
2293/// router's [`allow_migrate`](crate::db::DatabaseRouter::allow_migrate)
2294/// returns `false` for this (alias, model) pair the operation is
2295/// excluded from the alias's run. Junction / unowned tables (no
2296/// registered `ModelMeta`) are always allowed — the router has no
2297/// model to inspect.
2298fn op_targets_alias(op: &Operation, alias: &str) -> bool {
2299 if table_alias(op.table_name()) != alias {
2300 return false;
2301 }
2302 // Let the router veto migrating this table on this alias.
2303 match model_meta_for_table(op.table_name()) {
2304 Some(meta) => crate::db::router::router().allow_migrate(alias, &meta),
2305 None => true, // junction / unowned table — migrate on its alias
2306 }
2307}
2308
2309/// SQLite per-alias variant. Same shape as the legacy `run_in_sqlite`
2310/// but: filters ops to those routed to `alias`; skips files whose op
2311/// list contains nothing for this DB (so we don't stuff orphan
2312/// tracking rows into pools that didn't run any SQL).
2313async fn run_in_sqlite_for_alias(
2314 dir: &Path,
2315 alias: &str,
2316 pool: &sqlx::SqlitePool,
2317 shared_only: Option<&std::collections::HashSet<String>>,
2318) -> Result<u64, MigrateError> {
2319 ensure_tracking_table_sqlite(pool).await?;
2320 let applied = applied_names_sqlite(pool).await?;
2321
2322 let mut applied_count: u64 = 0;
2323 for plugin in plugin_order() {
2324 if let Some(shared) = shared_only {
2325 if !shared.contains(&plugin) {
2326 continue;
2327 }
2328 }
2329 let plugin_dir = dir.join(&plugin);
2330 let paths = list_migration_files(&plugin_dir)?;
2331
2332 // Read the plugin's full file set, then plan squash decisions in one
2333 // pass (gaps2 #100). `squash_plan` decides Apply / Skip / RecordOnly per
2334 // file, honoring `replaces` so a squash and its originals never both
2335 // apply. Pure and backend-agnostic — the Postgres runner plans the same.
2336 let files: Vec<MigrationFile> = paths
2337 .iter()
2338 .map(|p| read_migration_file(p))
2339 .collect::<Result<_, _>>()?;
2340 let plan = squash_plan(&files, &applied)?;
2341
2342 for (file, decision) in files.iter().zip(plan) {
2343 if decision == ApplyDecision::Skip {
2344 continue;
2345 }
2346
2347 let ops_for_this_db: Vec<&Operation> = file
2348 .operations
2349 .iter()
2350 .filter(|op| op_targets_alias(op, alias))
2351 .collect();
2352 if ops_for_this_db.is_empty() {
2353 // File's content all targets some other DB. Don't
2354 // record it here — re-runs will re-evaluate cleanly
2355 // once the right DB picks it up. The tracking rows
2356 // per-DB stay accurate to "what actually ran here."
2357 continue;
2358 }
2359
2360 let snapshot_hash = file.snapshot_after.hash();
2361 let applied_at = chrono::Utc::now().to_rfc3339();
2362 // RecordOnly: the squash's replaced set is already applied, so its
2363 // schema exists — insert the tracking row without running any ops
2364 // (an empty op slice inserts the row and commits).
2365 let ops_to_run: &[&Operation] = if decision == ApplyDecision::RecordOnly {
2366 &[]
2367 } else {
2368 &ops_for_this_db
2369 };
2370 apply_sqlite_migration_tx(
2371 pool,
2372 ops_to_run,
2373 &file.plugin,
2374 &file.id,
2375 &applied_at,
2376 &snapshot_hash,
2377 )
2378 .await?;
2379 applied_count += 1;
2380 }
2381 }
2382 Ok(applied_count)
2383}
2384
2385/// Apply one migration file's SQLite `ops` in a single transaction, then record
2386/// it in the tracking table.
2387///
2388/// When any op is an `AlterColumn` (the table-recreation dance), the
2389/// transaction is bracketed with `PRAGMA foreign_keys=OFF` … `PRAGMA
2390/// foreign_key_check` … `PRAGMA foreign_keys=ON` on a **pinned** connection
2391/// (SQLite's official recipe), so step 3's `DROP TABLE` on a table with inbound
2392/// FKs doesn't fail with `FOREIGN KEY constraint failed` (error 787, gaps3
2393/// #13). The pragma MUST be toggled outside the tx (it's a no-op inside one),
2394/// and enforcement is restored even on failure so the pooled connection never
2395/// returns with FK checks disabled. `foreign_key_check` before commit keeps the
2396/// integrity guarantee: a migration that genuinely orphans a row is aborted.
2397async fn apply_sqlite_migration_tx(
2398 pool: &sqlx::SqlitePool,
2399 ops: &[&Operation],
2400 plugin: &str,
2401 name: &str,
2402 applied_at: &str,
2403 snapshot_hash: &str,
2404) -> Result<(), MigrateError> {
2405 use sqlx::Acquire as _;
2406
2407 let needs_fk_off = ops
2408 .iter()
2409 .any(|op| matches!(op, Operation::AlterColumn { .. }));
2410
2411 let mut conn = pool.acquire().await?;
2412 if needs_fk_off {
2413 sqlx::query("PRAGMA foreign_keys=OFF")
2414 .execute(&mut *conn)
2415 .await?;
2416 }
2417
2418 let result: Result<(), MigrateError> = async {
2419 let mut tx = conn.begin().await?;
2420 for op in ops {
2421 for sql in render_operation_for(op, "sqlite") {
2422 sqlx::query(&sql).execute(&mut *tx).await?;
2423 }
2424 }
2425 if needs_fk_off {
2426 // Enforcement was off during the dance; verify the recreation left
2427 // no dangling references before we commit.
2428 let violations = sqlx::query("PRAGMA foreign_key_check")
2429 .fetch_all(&mut *tx)
2430 .await?;
2431 if !violations.is_empty() {
2432 return Err(MigrateError::Sqlx(sqlx::Error::Protocol(format!(
2433 "migration `{plugin}/{name}` would leave {} dangling foreign-key \
2434 reference(s); aborted",
2435 violations.len()
2436 ))));
2437 }
2438 }
2439 sqlx::query(
2440 "INSERT INTO umbral_migrations (plugin, name, applied_at, snapshot_hash) \
2441 VALUES (?, ?, ?, ?)",
2442 )
2443 .bind(plugin)
2444 .bind(name)
2445 .bind(applied_at)
2446 .bind(snapshot_hash)
2447 .execute(&mut *tx)
2448 .await?;
2449 tx.commit().await?;
2450 Ok(())
2451 }
2452 .await;
2453
2454 if needs_fk_off {
2455 // Restore enforcement before the connection returns to the pool, even
2456 // on failure — but never mask the primary error with a pragma error.
2457 let _ = sqlx::query("PRAGMA foreign_keys=ON")
2458 .execute(&mut *conn)
2459 .await;
2460 }
2461 result
2462}
2463
2464/// A stable 64-bit key for the Postgres migration advisory lock, derived from a
2465/// fixed namespace + a `discriminator` (the pool alias or tenant schema). FNV-1a
2466/// with fixed constants — deterministic and process-independent, so every
2467/// migrator computes the SAME key for the same target and they mutually exclude.
2468/// Different aliases/schemas get different keys, so unrelated logical databases
2469/// migrate concurrently (audit_2 core-migrate #7).
2470fn pg_migration_lock_key(discriminator: &str) -> i64 {
2471 let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
2472 for b in b"umbral_migrations\0"
2473 .iter()
2474 .copied()
2475 .chain(discriminator.bytes())
2476 {
2477 hash ^= b as u64;
2478 hash = hash.wrapping_mul(0x0000_0100_0000_01b3);
2479 }
2480 hash as i64
2481}
2482
2483/// How long a migrator waits for the advisory lock before giving up. Generous
2484/// (another replica's migration set can take a while), bounded so a deploy can't
2485/// hang forever on a wedged migrator. Override with
2486/// `UMBRAL_MIGRATION_LOCK_TIMEOUT_SECS`.
2487fn pg_migration_lock_timeout() -> std::time::Duration {
2488 let secs = std::env::var("UMBRAL_MIGRATION_LOCK_TIMEOUT_SECS")
2489 .ok()
2490 .and_then(|v| v.trim().parse::<u64>().ok())
2491 .filter(|&n| n > 0)
2492 .unwrap_or(300);
2493 std::time::Duration::from_secs(secs)
2494}
2495
2496/// Acquire the session-level Postgres advisory lock keyed on `discriminator`,
2497/// holding it on `conn` for the caller to release. Non-blocking polls
2498/// (`pg_try_advisory_lock`) with a short sleep between attempts, bounded by
2499/// [`pg_migration_lock_timeout`] — so a stuck migrator surfaces as a clear
2500/// [`MigrateError::MigrationLockTimeout`] instead of an indefinite hang, and a
2501/// crashed migrator's lock auto-releases (session locks die with the backend).
2502async fn acquire_pg_migration_lock(
2503 conn: &mut sqlx::PgConnection,
2504 key: i64,
2505 discriminator: &str,
2506) -> Result<(), MigrateError> {
2507 let timeout = pg_migration_lock_timeout();
2508 let start = std::time::Instant::now();
2509 let mut warned = false;
2510 loop {
2511 let got: bool = sqlx::query_scalar("SELECT pg_try_advisory_lock($1)")
2512 .bind(key)
2513 .fetch_one(&mut *conn)
2514 .await?;
2515 if got {
2516 return Ok(());
2517 }
2518 let waited = start.elapsed();
2519 if waited >= timeout {
2520 return Err(MigrateError::MigrationLockTimeout {
2521 discriminator: discriminator.to_string(),
2522 waited_secs: waited.as_secs(),
2523 });
2524 }
2525 if !warned {
2526 tracing::info!(
2527 discriminator,
2528 "umbral migrate: another process holds the migration lock; waiting for it to \
2529 finish before applying migrations…"
2530 );
2531 warned = true;
2532 }
2533 tokio::time::sleep(std::time::Duration::from_millis(200)).await;
2534 }
2535}
2536
2537/// Release the advisory lock acquired by [`acquire_pg_migration_lock`]. Best
2538/// effort — dropping `conn` also releases a session lock, so a failure here is
2539/// logged, not propagated (it must never mask the migration's own result).
2540async fn release_pg_migration_lock(conn: &mut sqlx::PgConnection, key: i64) {
2541 if let Err(e) = sqlx::query("SELECT pg_advisory_unlock($1)")
2542 .bind(key)
2543 .execute(&mut *conn)
2544 .await
2545 {
2546 tracing::warn!("umbral migrate: failed to release the migration advisory lock: {e}");
2547 }
2548}
2549
2550/// Postgres per-alias variant. Mirror of `run_in_sqlite_for_alias`.
2551///
2552/// Wraps the apply loop in a session advisory lock (audit_2 core-migrate #7) so
2553/// two replicas deploying at once can't both read the applied set and race the
2554/// same DDL (the loser errors "relation already exists" mid-deploy). The lock is
2555/// held on a dedicated connection for the whole run; the per-migration
2556/// transactions use their own pooled connections and are unaffected.
2557async fn run_in_postgres_for_alias(
2558 dir: &Path,
2559 alias: &str,
2560 pool: &sqlx::PgPool,
2561 shared_only: Option<&std::collections::HashSet<String>>,
2562) -> Result<u64, MigrateError> {
2563 let key = pg_migration_lock_key(alias);
2564 let mut lock_conn = pool.acquire().await?;
2565 acquire_pg_migration_lock(&mut lock_conn, key, alias).await?;
2566 let result = run_in_postgres_for_alias_locked(dir, alias, pool, shared_only).await;
2567 release_pg_migration_lock(&mut lock_conn, key).await;
2568 result
2569}
2570
2571/// Run `f` under the Postgres migration advisory lock keyed on `discriminator`.
2572/// Shared by the checked-run and schema-per-tenant apply paths so they get the
2573/// same cross-process serialization as [`run_in_postgres_for_alias`].
2574async fn with_pg_migration_lock<F, Fut>(
2575 pool: &sqlx::PgPool,
2576 discriminator: &str,
2577 f: F,
2578) -> Result<u64, MigrateError>
2579where
2580 F: FnOnce() -> Fut,
2581 Fut: std::future::Future<Output = Result<u64, MigrateError>>,
2582{
2583 let key = pg_migration_lock_key(discriminator);
2584 let mut lock_conn = pool.acquire().await?;
2585 acquire_pg_migration_lock(&mut lock_conn, key, discriminator).await?;
2586 let result = f().await;
2587 release_pg_migration_lock(&mut lock_conn, key).await;
2588 result
2589}
2590
2591/// The unlocked body of [`run_in_postgres_for_alias`] — runs while the caller
2592/// holds the migration advisory lock.
2593async fn run_in_postgres_for_alias_locked(
2594 dir: &Path,
2595 alias: &str,
2596 pool: &sqlx::PgPool,
2597 shared_only: Option<&std::collections::HashSet<String>>,
2598) -> Result<u64, MigrateError> {
2599 ensure_tracking_table_postgres(pool).await?;
2600 let applied = applied_names_postgres(pool).await?;
2601
2602 let mut applied_count: u64 = 0;
2603 for plugin in plugin_order() {
2604 // Shared-filtered public migrate (multitenancy): when a shared-app set
2605 // is given, migrate ONLY those plugins into this pool, so a tenant
2606 // app's tables (and its M2M junctions) are NOT created in `public` —
2607 // they belong only in each tenant schema. `None` = migrate everything
2608 // (the default single-DB behaviour, byte-identical to before).
2609 if let Some(shared) = shared_only {
2610 if !shared.contains(&plugin) {
2611 continue;
2612 }
2613 }
2614 let plugin_dir = dir.join(&plugin);
2615 let paths = list_migration_files(&plugin_dir)?;
2616
2617 // Plan squash decisions for the plugin's whole file set (gaps2 #100),
2618 // identical to the SQLite runner — `squash_plan` is backend-agnostic.
2619 let files: Vec<MigrationFile> = paths
2620 .iter()
2621 .map(|p| read_migration_file(p))
2622 .collect::<Result<_, _>>()?;
2623 let plan = squash_plan(&files, &applied)?;
2624
2625 for (file, decision) in files.iter().zip(plan) {
2626 if decision == ApplyDecision::Skip {
2627 continue;
2628 }
2629
2630 let ops_for_this_db: Vec<&Operation> = file
2631 .operations
2632 .iter()
2633 .filter(|op| op_targets_alias(op, alias))
2634 .collect();
2635 if ops_for_this_db.is_empty() {
2636 continue;
2637 }
2638
2639 let mut tx = pool.begin().await?;
2640 // RecordOnly (squash whose replaced set is already applied): its
2641 // schema exists, so run no DDL — just insert the tracking row.
2642 if decision != ApplyDecision::RecordOnly {
2643 for op in &ops_for_this_db {
2644 for sql in render_operation(op) {
2645 sqlx::query(&sql).execute(&mut *tx).await?;
2646 }
2647 }
2648 }
2649 let snapshot_hash = file.snapshot_after.hash();
2650 let applied_at = chrono::Utc::now().to_rfc3339();
2651 sqlx::query(
2652 "INSERT INTO umbral_migrations (plugin, name, applied_at, snapshot_hash) \
2653 VALUES ($1, $2, $3, $4)",
2654 )
2655 .bind(&file.plugin)
2656 .bind(&file.id)
2657 .bind(&applied_at)
2658 .bind(&snapshot_hash)
2659 .execute(&mut *tx)
2660 .await?;
2661 tx.commit().await?;
2662 applied_count += 1;
2663 }
2664 }
2665 Ok(applied_count)
2666}
2667
2668/// Migrate the **tenant** apps into a named Postgres schema (schema-per-tenant
2669/// style). The migration engine owns all schema DDL; this is the
2670/// sanctioned `CREATE SCHEMA` / `SET search_path` exception (a plugin calls this
2671/// rather than writing raw schema SQL itself).
2672///
2673/// Steps, all inside one transaction per migration file (mirroring
2674/// [`run_in_postgres_for_alias`]):
2675/// 1. `CREATE SCHEMA IF NOT EXISTS "<schema>"` (the `Schema` was already
2676/// validated to a safe PG identifier, but is still emitted quoted).
2677/// 2. `SET LOCAL search_path TO "<schema>"` so every unqualified
2678/// `CREATE TABLE` **and** the `umbral_migrations` ledger land *inside*
2679/// `<schema>` — per-schema migration tracking falls out for free.
2680/// 3. Apply pending migrations, **filtered to the tenant apps** — every plugin
2681/// NOT in `shared_apps` (those tables live in `public` and are migrated by
2682/// the normal [`run`]). A file with no tenant-app ops for this schema is
2683/// skipped without a tracking row.
2684///
2685/// Idempotent: re-running applies only the migrations the schema's own
2686/// `umbral_migrations` ledger hasn't recorded. Postgres-only — schemas don't
2687/// exist on SQLite, so a SQLite pool is a clear error
2688/// ([`MigrateError::SchemaUnsupportedOnSqlite`]).
2689pub async fn run_for_schema(
2690 schema: &crate::db::Schema,
2691 shared_apps: &std::collections::HashSet<String>,
2692) -> Result<u64, MigrateError> {
2693 run_for_schema_in(Path::new(MIGRATIONS_DIR), schema, shared_apps).await
2694}
2695
2696/// Same as [`run_for_schema`] but takes an explicit migrations base directory.
2697/// The entry tests drive.
2698pub async fn run_for_schema_in(
2699 dir: &Path,
2700 schema: &crate::db::Schema,
2701 shared_apps: &std::collections::HashSet<String>,
2702) -> Result<u64, MigrateError> {
2703 match crate::db::pool_dispatched() {
2704 crate::db::DbPool::Postgres(p) => {
2705 // audit_2 core-migrate #7: serialize concurrent migrators of THIS
2706 // tenant schema (keyed by schema name, so different tenants still
2707 // migrate concurrently). The shared/public run uses a different key
2708 // (its alias), so a tenant migrate and the public migrate don't
2709 // block each other.
2710 with_pg_migration_lock(p, schema.as_str(), || {
2711 run_tenant_apps_in_postgres_schema(dir, schema, shared_apps, p)
2712 })
2713 .await
2714 }
2715 crate::db::DbPool::Sqlite(_) => Err(MigrateError::SchemaUnsupportedOnSqlite {
2716 schema: schema.as_str().to_string(),
2717 }),
2718 }
2719}
2720
2721/// Postgres schema-scoped variant of [`run_in_postgres_for_alias`]. Creates the
2722/// schema, pins `search_path` to it for the transaction, and applies only the
2723/// tenant apps' migrations (plugins not in `shared_apps`). The `umbral_migrations`
2724/// ledger is read/written *inside* the schema (search_path is set first), so
2725/// tracking is per-schema with no extra book-keeping.
2726async fn run_tenant_apps_in_postgres_schema(
2727 dir: &Path,
2728 schema: &crate::db::Schema,
2729 shared_apps: &std::collections::HashSet<String>,
2730 pool: &sqlx::PgPool,
2731) -> Result<u64, MigrateError> {
2732 let quoted = format!("\"{}\"", schema.as_str());
2733
2734 // Create the schema once, outside the per-file loop. IF NOT EXISTS makes
2735 // the whole call idempotent.
2736 sqlx::query(&format!("CREATE SCHEMA IF NOT EXISTS {quoted}"))
2737 .execute(pool)
2738 .await?;
2739
2740 // Ensure + read the ledger INSIDE the schema. Each block runs in its own
2741 // transaction with `SET LOCAL search_path` so the tracking table is created
2742 // in (and read from) `<schema>`, not `public` — AND the search_path is
2743 // transaction-scoped, so the pooled connection is NOT left pinned to this
2744 // schema when it returns to the pool. A plain session-level `SET` here
2745 // pollutes the pool: the next unqualified ORM query that reuses the
2746 // connection would resolve against `<schema>` instead of `public` (e.g. an
2747 // insert into the public `tenant` registry failing with "relation does not
2748 // exist") — a real cross-tenant bug, caught only against live Postgres.
2749 {
2750 let mut tx = pool.begin().await?;
2751 sqlx::query(&format!("SET LOCAL search_path TO {quoted}"))
2752 .execute(&mut *tx)
2753 .await?;
2754 ensure_tracking_table_pg_conn(&mut tx).await?;
2755 tx.commit().await?;
2756 }
2757 let applied = {
2758 let mut tx = pool.begin().await?;
2759 sqlx::query(&format!("SET LOCAL search_path TO {quoted}"))
2760 .execute(&mut *tx)
2761 .await?;
2762 let rows: Vec<(String, String)> =
2763 sqlx::query_as("SELECT plugin, name FROM umbral_migrations")
2764 .fetch_all(&mut *tx)
2765 .await?;
2766 tx.commit().await?;
2767 rows.into_iter().collect::<std::collections::HashSet<_>>()
2768 };
2769
2770 let mut applied_count: u64 = 0;
2771 for plugin in plugin_order() {
2772 // Tenant apps only — shared apps live in `public`.
2773 if shared_apps.contains(&plugin) {
2774 continue;
2775 }
2776 let plugin_dir = dir.join(&plugin);
2777 let paths = list_migration_files(&plugin_dir)?;
2778
2779 // Squash-aware planning (gaps2 #100), same as every other apply loop.
2780 let files: Vec<MigrationFile> = paths
2781 .iter()
2782 .map(|p| read_migration_file(p))
2783 .collect::<Result<_, _>>()?;
2784 let plan = squash_plan(&files, &applied)?;
2785
2786 for (file, decision) in files.iter().zip(plan) {
2787 if decision == ApplyDecision::Skip {
2788 continue;
2789 }
2790 // Belt-and-braces: skip a file whose declared plugin is shared.
2791 if shared_apps.contains(&file.plugin) {
2792 continue;
2793 }
2794
2795 let mut tx = pool.begin().await?;
2796 // Pin search_path for THIS transaction, tenant schema FIRST with
2797 // `public` as a fallback. `CREATE TABLE` / `INSERT` still land in
2798 // the tenant schema (it's first), but an unqualified reference that
2799 // ISN'T in the tenant schema resolves against `public` — which is
2800 // what makes a CROSS-BOUNDARY foreign key work: a tenant-owned
2801 // table (or an M2M junction) with an FK `REFERENCES <shared_child>`
2802 // resolves the shared child in `public` instead of erroring
2803 // `relation does not exist`. It also lets a (future) RunSql data
2804 // migration in a tenant schema read SHARED/`public` lookup tables.
2805 // The tenant-first ordering means a tenant table still shadows a
2806 // same-named public table, so no behaviour changes for the common
2807 // case where tenant and shared table names are distinct.
2808 sqlx::query(&format!("SET LOCAL search_path TO {quoted}, public"))
2809 .execute(&mut *tx)
2810 .await?;
2811 // RecordOnly: schema already built by the replaced originals — run
2812 // no DDL, just record the squash.
2813 if decision != ApplyDecision::RecordOnly {
2814 for op in &file.operations {
2815 for sql in render_operation_for(op, "postgres") {
2816 sqlx::query(&sql).execute(&mut *tx).await?;
2817 }
2818 }
2819 }
2820 let snapshot_hash = file.snapshot_after.hash();
2821 let applied_at = chrono::Utc::now().to_rfc3339();
2822 sqlx::query(
2823 "INSERT INTO umbral_migrations (plugin, name, applied_at, snapshot_hash) \
2824 VALUES ($1, $2, $3, $4)",
2825 )
2826 .bind(&file.plugin)
2827 .bind(&file.id)
2828 .bind(&applied_at)
2829 .bind(&snapshot_hash)
2830 .execute(&mut *tx)
2831 .await?;
2832 tx.commit().await?;
2833 applied_count += 1;
2834 }
2835 }
2836 Ok(applied_count)
2837}
2838
2839/// Migrate the **tenant** apps into the pool registered under `alias`
2840/// (database-per-tenant). The db-per-tenant sibling of [`run_for_schema`]:
2841/// where the schema variant pins `search_path` inside one shared Postgres
2842/// database, this targets a *whole separate database/pool* registered at
2843/// runtime via [`register_tenant_pool`](crate::db::register_tenant_pool) and
2844/// resolved here through [`pool_for_dispatched`](crate::db::pool_for_dispatched)
2845/// (which sees dynamic pools). No schema games — per-database migration
2846/// tracking is just that database's own `umbral_migrations` table.
2847///
2848/// Like the schema variant it applies only the **tenant apps**: every plugin
2849/// NOT in `shared_apps` (the shared registry/auth tables live in the default
2850/// DB and are migrated there by the normal [`run`]). A migration file whose
2851/// declared plugin is shared is skipped without a tracking row. Idempotent:
2852/// re-running applies only what the tenant DB's own ledger hasn't recorded.
2853///
2854/// Works on both backends — a tenant pool can be Postgres (the production case)
2855/// or SQLite (tests). Unlike the alias-routed [`run_in`], this does NOT filter
2856/// ops by [`table_alias`]: a tenant-owned model's static alias is still
2857/// `"default"`, so the per-alias filter would wrongly exclude it from the
2858/// tenant DB. The shared/tenant split is the *only* filter here.
2859pub async fn migrate_apps_into_pool(
2860 alias: &str,
2861 shared_apps: &std::collections::HashSet<String>,
2862) -> Result<u64, MigrateError> {
2863 migrate_apps_into_pool_in(Path::new(MIGRATIONS_DIR), alias, shared_apps).await
2864}
2865
2866/// Same as [`migrate_apps_into_pool`] but takes an explicit migrations base
2867/// directory. The entry tests drive.
2868pub async fn migrate_apps_into_pool_in(
2869 dir: &Path,
2870 alias: &str,
2871 shared_apps: &std::collections::HashSet<String>,
2872) -> Result<u64, MigrateError> {
2873 match crate::db::pool_for_dispatched(alias) {
2874 crate::db::DbPool::Postgres(p) => {
2875 migrate_tenant_apps_into_pg_pool(dir, shared_apps, p).await
2876 }
2877 crate::db::DbPool::Sqlite(p) => {
2878 migrate_tenant_apps_into_sqlite_pool(dir, shared_apps, p).await
2879 }
2880 }
2881}
2882
2883/// Postgres tenant-DB apply loop. Mirrors [`run_in_postgres_for_alias`] but the
2884/// only filter is the shared/tenant split — every plugin not in `shared_apps`
2885/// is applied in full into this database.
2886async fn migrate_tenant_apps_into_pg_pool(
2887 dir: &Path,
2888 shared_apps: &std::collections::HashSet<String>,
2889 pool: &sqlx::PgPool,
2890) -> Result<u64, MigrateError> {
2891 ensure_tracking_table_postgres(pool).await?;
2892 let applied = applied_names_postgres(pool).await?;
2893
2894 let mut applied_count: u64 = 0;
2895 for plugin in plugin_order() {
2896 if shared_apps.contains(&plugin) {
2897 continue;
2898 }
2899 let plugin_dir = dir.join(&plugin);
2900 let files: Vec<MigrationFile> = list_migration_files(&plugin_dir)?
2901 .iter()
2902 .map(|p| read_migration_file(p))
2903 .collect::<Result<_, _>>()?;
2904 let plan = squash_plan(&files, &applied)?;
2905 for (file, decision) in files.iter().zip(plan) {
2906 if decision == ApplyDecision::Skip {
2907 continue;
2908 }
2909 if shared_apps.contains(&file.plugin) {
2910 continue;
2911 }
2912 let mut tx = pool.begin().await?;
2913 if decision != ApplyDecision::RecordOnly {
2914 for op in &file.operations {
2915 for sql in render_operation_for(op, "postgres") {
2916 sqlx::query(&sql).execute(&mut *tx).await?;
2917 }
2918 }
2919 }
2920 let snapshot_hash = file.snapshot_after.hash();
2921 let applied_at = chrono::Utc::now().to_rfc3339();
2922 sqlx::query(
2923 "INSERT INTO umbral_migrations (plugin, name, applied_at, snapshot_hash) \
2924 VALUES ($1, $2, $3, $4)",
2925 )
2926 .bind(&file.plugin)
2927 .bind(&file.id)
2928 .bind(&applied_at)
2929 .bind(&snapshot_hash)
2930 .execute(&mut *tx)
2931 .await?;
2932 tx.commit().await?;
2933 applied_count += 1;
2934 }
2935 }
2936 Ok(applied_count)
2937}
2938
2939/// SQLite tenant-DB apply loop (tests). Same shape as the Postgres variant.
2940async fn migrate_tenant_apps_into_sqlite_pool(
2941 dir: &Path,
2942 shared_apps: &std::collections::HashSet<String>,
2943 pool: &sqlx::SqlitePool,
2944) -> Result<u64, MigrateError> {
2945 ensure_tracking_table_sqlite(pool).await?;
2946 let applied = applied_names_sqlite(pool).await?;
2947
2948 let mut applied_count: u64 = 0;
2949 for plugin in plugin_order() {
2950 if shared_apps.contains(&plugin) {
2951 continue;
2952 }
2953 let plugin_dir = dir.join(&plugin);
2954 let files: Vec<MigrationFile> = list_migration_files(&plugin_dir)?
2955 .iter()
2956 .map(|p| read_migration_file(p))
2957 .collect::<Result<_, _>>()?;
2958 let plan = squash_plan(&files, &applied)?;
2959 for (file, decision) in files.iter().zip(plan) {
2960 if decision == ApplyDecision::Skip {
2961 continue;
2962 }
2963 if shared_apps.contains(&file.plugin) {
2964 continue;
2965 }
2966 let snapshot_hash = file.snapshot_after.hash();
2967 let applied_at = chrono::Utc::now().to_rfc3339();
2968 // RecordOnly → empty op slice: inserts the tracking row, no DDL.
2969 let ops: Vec<&Operation> = if decision == ApplyDecision::RecordOnly {
2970 Vec::new()
2971 } else {
2972 file.operations.iter().collect()
2973 };
2974 apply_sqlite_migration_tx(
2975 pool,
2976 &ops,
2977 &file.plugin,
2978 &file.id,
2979 &applied_at,
2980 &snapshot_hash,
2981 )
2982 .await?;
2983 applied_count += 1;
2984 }
2985 }
2986 Ok(applied_count)
2987}
2988
2989/// `ensure_tracking_table_postgres` against an explicit connection (so the
2990/// caller can pin `search_path` first and have the table created in the tenant
2991/// schema rather than `public`).
2992async fn ensure_tracking_table_pg_conn(conn: &mut sqlx::PgConnection) -> Result<(), MigrateError> {
2993 sqlx::query(
2994 "CREATE TABLE IF NOT EXISTS umbral_migrations (
2995 plugin TEXT NOT NULL,
2996 name TEXT NOT NULL,
2997 applied_at TEXT NOT NULL,
2998 snapshot_hash TEXT NOT NULL,
2999 PRIMARY KEY (plugin, name)
3000 )",
3001 )
3002 .execute(conn)
3003 .await?;
3004 Ok(())
3005}
3006
3007/// SQLite drift-checking path for `run_checked_in`.
3008///
3009/// Reads the applied set, runs `detect_all_drift`, and either errors
3010/// (if `allow_drift = false` and critical drift is found) or logs a
3011/// warning and proceeds (if `allow_drift = true`). Then delegates to
3012/// `run_in_sqlite` for the actual apply loop.
3013async fn run_in_sqlite_checked(
3014 dir: &Path,
3015 pool: &sqlx::SqlitePool,
3016 allow_drift: bool,
3017 alias: &str,
3018) -> Result<u64, MigrateError> {
3019 ensure_tracking_table_sqlite(pool).await?;
3020 let applied = applied_names_sqlite(pool).await?;
3021 let report = detect_all_drift(&applied, dir)?;
3022
3023 if report.has_critical_drift() {
3024 if allow_drift {
3025 let missing = report.missing_on_disk();
3026 for entry in &missing {
3027 eprintln!(
3028 "warning: umbral migrate --allow-drift: migration {}/{} is recorded in \
3029 the tracking table but the file is missing from disk; proceeding.",
3030 entry.plugin, entry.name
3031 );
3032 }
3033 } else {
3034 let missing: Vec<(String, String)> = report
3035 .missing_on_disk()
3036 .iter()
3037 .map(|e| (e.plugin.clone(), e.name.clone()))
3038 .collect();
3039 return Err(MigrateError::DriftDetected { missing });
3040 }
3041 }
3042
3043 // Emit warnings for out-of-order files.
3044 for entry in report
3045 .entries
3046 .iter()
3047 .filter(|e| e.status == MigrationStatus::OutOfOrder)
3048 {
3049 eprintln!(
3050 "warning: umbral migrate: migration {}/{} is on disk but appears before the \
3051 last applied migration for this plugin; it looks like a file was restored \
3052 after a teammate already applied later ones.",
3053 entry.plugin, entry.name
3054 );
3055 }
3056
3057 run_in_sqlite_for_alias(dir, alias, pool, None).await
3058}
3059
3060/// Postgres drift-checking path for `run_checked_in`. Same logic as
3061/// `run_in_sqlite_checked` but uses the Postgres applied-set reader.
3062async fn run_in_postgres_checked(
3063 dir: &Path,
3064 pool: &sqlx::PgPool,
3065 allow_drift: bool,
3066 alias: &str,
3067) -> Result<u64, MigrateError> {
3068 ensure_tracking_table_postgres(pool).await?;
3069 let applied = applied_names_postgres(pool).await?;
3070 let report = detect_all_drift(&applied, dir)?;
3071
3072 if report.has_critical_drift() {
3073 if allow_drift {
3074 let missing = report.missing_on_disk();
3075 for entry in &missing {
3076 eprintln!(
3077 "warning: umbral migrate --allow-drift: migration {}/{} is recorded in \
3078 the tracking table but the file is missing from disk; proceeding.",
3079 entry.plugin, entry.name
3080 );
3081 }
3082 } else {
3083 let missing: Vec<(String, String)> = report
3084 .missing_on_disk()
3085 .iter()
3086 .map(|e| (e.plugin.clone(), e.name.clone()))
3087 .collect();
3088 return Err(MigrateError::DriftDetected { missing });
3089 }
3090 }
3091
3092 for entry in report
3093 .entries
3094 .iter()
3095 .filter(|e| e.status == MigrationStatus::OutOfOrder)
3096 {
3097 eprintln!(
3098 "warning: umbral migrate: migration {}/{} is on disk but appears before the \
3099 last applied migration for this plugin; it looks like a file was restored \
3100 after a teammate already applied later ones.",
3101 entry.plugin, entry.name
3102 );
3103 }
3104
3105 run_in_postgres_for_alias(dir, alias, pool, None).await
3106}
3107
3108/// Record a migration as applied in the `umbral_migrations` tracking
3109/// table without running its operations. The "mark as applied" path
3110/// `inspectdb --mark-applied` uses to register the introspected
3111/// `0001_initial` against an already-populated database. Idempotent:
3112/// if the `(plugin, name)` row already exists, the call is a no-op.
3113pub async fn record_applied(
3114 plugin: &str,
3115 name: &str,
3116 snapshot_hash: &str,
3117) -> Result<(), MigrateError> {
3118 let applied_at = chrono::Utc::now().to_rfc3339();
3119 match crate::db::pool_dispatched() {
3120 crate::db::DbPool::Sqlite(pool) => {
3121 ensure_tracking_table_sqlite(pool).await?;
3122 sqlx::query(
3123 "INSERT OR IGNORE INTO umbral_migrations \
3124 (plugin, name, applied_at, snapshot_hash) \
3125 VALUES (?, ?, ?, ?)",
3126 )
3127 .bind(plugin)
3128 .bind(name)
3129 .bind(&applied_at)
3130 .bind(snapshot_hash)
3131 .execute(pool)
3132 .await?;
3133 }
3134 crate::db::DbPool::Postgres(pool) => {
3135 ensure_tracking_table_postgres(pool).await?;
3136 sqlx::query(
3137 "INSERT INTO umbral_migrations \
3138 (plugin, name, applied_at, snapshot_hash) \
3139 VALUES ($1, $2, $3, $4) \
3140 ON CONFLICT (plugin, name) DO NOTHING",
3141 )
3142 .bind(plugin)
3143 .bind(name)
3144 .bind(&applied_at)
3145 .bind(snapshot_hash)
3146 .execute(pool)
3147 .await?;
3148 }
3149 }
3150 Ok(())
3151}
3152
3153// =========================================================================
3154// Drift detection — gap 24.
3155// =========================================================================
3156
3157/// Compute the drift report for a single plugin directory. Compares the
3158/// set of `(plugin, name)` pairs recorded in the tracking table against
3159/// the migration files present on disk and classifies each into one of
3160/// the four [`MigrationStatus`] states.
3161///
3162/// `applied` is the full set of `(plugin, name)` tuples already read
3163/// from the tracking table (shared across plugins to avoid extra DB
3164/// round-trips). `plugin_dir` is the on-disk directory for this plugin;
3165/// an absent directory is treated the same as an empty one.
3166///
3167/// # Classification
3168///
3169/// - File present + in DB → `Applied`
3170/// - File absent + in DB → `AppliedButMissing`
3171/// - File present + not in DB + seq ≤ max_applied_seq → `OutOfOrder`
3172/// - File present + not in DB + seq > max_applied_seq → `Pending`
3173///
3174/// The sequence number is the numeric prefix of the migration name
3175/// (e.g. `0001` in `0001_create_post`). Absence of any applied
3176/// migration for this plugin means `max_applied_seq = 0`.
3177pub fn detect_drift(
3178 plugin: &str,
3179 applied: &std::collections::HashSet<(String, String)>,
3180 plugin_dir: &Path,
3181) -> Result<Vec<MigrationEntry>, MigrateError> {
3182 // Collect on-disk migration names (the id, not the full path).
3183 let paths = list_migration_files(plugin_dir)?;
3184 let mut on_disk: Vec<String> = Vec::new();
3185 // Migrations shadowed by an APPLIED squash: an applied `000X_squashed`
3186 // replaces its originals, which are kept on disk (Django-style, non-
3187 // destructive) but get no tracking row. Without this they'd be
3188 // misclassified `Pending` forever and block umbral-health readiness on
3189 // a freshly-squashed database. Mirrors `squash_plan`'s `shadowed` set.
3190 let mut shadowed: std::collections::HashSet<String> = std::collections::HashSet::new();
3191 for path in &paths {
3192 let file = read_migration_file(path)?;
3193 if applied.contains(&(plugin.to_string(), file.id.clone())) {
3194 for m in &file.replaces {
3195 shadowed.insert(m.migration.clone());
3196 }
3197 }
3198 on_disk.push(file.id.clone());
3199 }
3200
3201 // Pull every tracking-table entry for this plugin.
3202 let plugin_applied: Vec<&str> = applied
3203 .iter()
3204 .filter(|(p, _)| p == plugin)
3205 .map(|(_, n)| n.as_str())
3206 .collect();
3207
3208 // Highest sequence number among applied migrations for this plugin.
3209 let max_applied_seq: u32 = plugin_applied
3210 .iter()
3211 .filter_map(|name| name.split('_').next()?.parse::<u32>().ok())
3212 .max()
3213 .unwrap_or(0);
3214
3215 let on_disk_set: std::collections::HashSet<&str> = on_disk.iter().map(|s| s.as_str()).collect();
3216
3217 let mut entries: Vec<MigrationEntry> = Vec::new();
3218
3219 // Walk on-disk files in order.
3220 for name in &on_disk {
3221 let key = (plugin.to_string(), name.clone());
3222 let status = if applied.contains(&key) || shadowed.contains(name) {
3223 // Applied directly, or shadowed by an applied squash (its
3224 // originals are covered even without their own tracking row).
3225 MigrationStatus::Applied
3226 } else {
3227 // Determine this migration's sequence number.
3228 let seq: u32 = name
3229 .split('_')
3230 .next()
3231 .and_then(|s| s.parse().ok())
3232 .unwrap_or(0);
3233 if seq <= max_applied_seq && max_applied_seq > 0 {
3234 MigrationStatus::OutOfOrder
3235 } else {
3236 MigrationStatus::Pending
3237 }
3238 };
3239 entries.push(MigrationEntry {
3240 plugin: plugin.to_string(),
3241 name: name.clone(),
3242 status,
3243 });
3244 }
3245
3246 // Walk applied entries not present on disk.
3247 for name in &plugin_applied {
3248 if !on_disk_set.contains(*name) {
3249 entries.push(MigrationEntry {
3250 plugin: plugin.to_string(),
3251 name: (*name).to_string(),
3252 status: MigrationStatus::AppliedButMissing,
3253 });
3254 }
3255 }
3256
3257 // Sort: applied-but-missing entries bubble after their expected
3258 // position is not determinable; sort all entries by name for a
3259 // deterministic order. In practice, applied-but-missing names
3260 // are still prefixed with the numeric sequence so lexical sort
3261 // yields the right display order.
3262 entries.sort_by(|a, b| a.name.cmp(&b.name));
3263
3264 Ok(entries)
3265}
3266
3267/// Detect drift across every registered plugin and return a combined
3268/// [`DriftReport`]. Called by `run_in_checked` before executing SQL
3269/// and by `show_in` when displaying the four-state list.
3270///
3271/// `applied` is already fetched from the DB; `dir` is the migrations
3272/// root directory.
3273pub fn detect_all_drift(
3274 applied: &std::collections::HashSet<(String, String)>,
3275 dir: &Path,
3276) -> Result<DriftReport, MigrateError> {
3277 let mut all_entries: Vec<MigrationEntry> = Vec::new();
3278
3279 // Also surface any tracking-table entries whose plugin directory
3280 // doesn't appear in the registered-plugins list — a plugin was
3281 // removed entirely but its DB rows remain.
3282 let mut seen_plugins: std::collections::HashSet<String> = std::collections::HashSet::new();
3283
3284 for plugin in plugin_order() {
3285 seen_plugins.insert(plugin.clone());
3286 let plugin_dir = dir.join(&plugin);
3287 let entries = detect_drift(&plugin, applied, &plugin_dir)?;
3288 all_entries.extend(entries);
3289 }
3290
3291 // Any applied entries whose plugin is not in the registered set at
3292 // all — treat them as AppliedButMissing (the whole plugin is gone).
3293 for (plugin, name) in applied {
3294 if !seen_plugins.contains(plugin.as_str()) {
3295 all_entries.push(MigrationEntry {
3296 plugin: plugin.clone(),
3297 name: name.clone(),
3298 status: MigrationStatus::AppliedButMissing,
3299 });
3300 }
3301 }
3302
3303 Ok(DriftReport {
3304 entries: all_entries,
3305 })
3306}
3307
3308/// Record a migration as applied in the tracking table WITHOUT running
3309/// its SQL operations. The `--fake` recovery path: the schema already
3310/// exists (e.g. the migration was run outside umbral, or the DB was
3311/// bootstrapped from a dump) and the operator wants to bring the
3312/// tracking table into sync without re-executing the DDL.
3313///
3314/// Idempotent: if `(plugin, name)` is already in the table the call
3315/// is a no-op (same behaviour as `record_applied`).
3316///
3317/// The snapshot hash is derived from the migration file on disk.
3318/// Returns `MigrateError::Io` if the file can't be found (the caller
3319/// should verify the name before calling this).
3320pub async fn fake_apply(plugin: &str, name: &str) -> Result<(), MigrateError> {
3321 fake_apply_in(plugin, name, Path::new(MIGRATIONS_DIR)).await
3322}
3323
3324/// Same as [`fake_apply`] but takes an explicit migrations base dir.
3325/// Used by tests and by the CLI when `--migrations-dir` is passed.
3326pub async fn fake_apply_in(plugin: &str, name: &str, dir: &Path) -> Result<(), MigrateError> {
3327 let path = dir.join(plugin).join(format!("{name}.json"));
3328 let file = read_migration_file(&path)?;
3329 let snapshot_hash = file.snapshot_after.hash();
3330 record_applied(plugin, name, &snapshot_hash).await
3331}
3332
3333/// For every registered plugin's first migration (`0001_*`), check
3334/// whether the tables that migration would create already exist in the
3335/// database. If they do, fake-apply the migration (mark it applied
3336/// without running its SQL).
3337///
3338/// This is the `--fake-initial` path: the operator has a database
3339/// bootstrapped outside umbral (a dump restore, a manual `CREATE TABLE`,
3340/// or a previous schema manager) and wants to bring the tracking table
3341/// into sync so subsequent `migrate` calls apply only the genuine
3342/// deltas.
3343///
3344/// Returns the number of plugins whose `0001_*` migration was
3345/// fake-applied. Zero means either no `0001_*` file exists or the
3346/// target tables were absent (in which case normal `migrate` should be
3347/// run to create them).
3348pub async fn fake_initial() -> Result<u64, MigrateError> {
3349 fake_initial_in(Path::new(MIGRATIONS_DIR)).await
3350}
3351
3352/// Same as [`fake_initial`] but takes an explicit migrations base dir.
3353pub async fn fake_initial_in(dir: &Path) -> Result<u64, MigrateError> {
3354 match crate::db::pool_dispatched() {
3355 crate::db::DbPool::Sqlite(pool) => fake_initial_sqlite(dir, pool).await,
3356 crate::db::DbPool::Postgres(pool) => fake_initial_postgres(dir, pool).await,
3357 }
3358}
3359
3360/// SQLite path for [`fake_initial_in`].
3361async fn fake_initial_sqlite(dir: &Path, pool: &sqlx::SqlitePool) -> Result<u64, MigrateError> {
3362 ensure_tracking_table_sqlite(pool).await?;
3363 let applied = applied_names_sqlite(pool).await?;
3364 let mut count: u64 = 0;
3365
3366 for plugin in plugin_order() {
3367 let plugin_dir = dir.join(&plugin);
3368 let paths = list_migration_files(&plugin_dir)?;
3369
3370 // Find the first migration file (lowest sequence number).
3371 let first = paths.first();
3372 let first = match first {
3373 Some(p) => p,
3374 None => continue,
3375 };
3376 let file = read_migration_file(first)?;
3377
3378 // Skip if already applied.
3379 if applied.contains(&(file.plugin.clone(), file.id.clone())) {
3380 continue;
3381 }
3382
3383 // Check whether the tables the first migration would create
3384 // already exist in the database.
3385 let tables_to_create: Vec<&str> = file
3386 .operations
3387 .iter()
3388 .filter_map(|op| match op {
3389 Operation::CreateTable { table, .. } => Some(table.as_str()),
3390 _ => None,
3391 })
3392 .collect();
3393
3394 if tables_to_create.is_empty() {
3395 continue;
3396 }
3397
3398 // All tables present → fake-apply.
3399 let mut all_present = true;
3400 for table in &tables_to_create {
3401 let exists: Option<(String,)> =
3402 sqlx::query_as("SELECT name FROM sqlite_master WHERE type = 'table' AND name = ?")
3403 .bind(*table)
3404 .fetch_optional(pool)
3405 .await?;
3406 if exists.is_none() {
3407 all_present = false;
3408 break;
3409 }
3410 }
3411
3412 if all_present {
3413 let snapshot_hash = file.snapshot_after.hash();
3414 let applied_at = chrono::Utc::now().to_rfc3339();
3415 sqlx::query(
3416 "INSERT OR IGNORE INTO umbral_migrations \
3417 (plugin, name, applied_at, snapshot_hash) VALUES (?, ?, ?, ?)",
3418 )
3419 .bind(&file.plugin)
3420 .bind(&file.id)
3421 .bind(&applied_at)
3422 .bind(&snapshot_hash)
3423 .execute(pool)
3424 .await?;
3425 count += 1;
3426 }
3427 }
3428
3429 Ok(count)
3430}
3431
3432/// Postgres path for [`fake_initial_in`].
3433async fn fake_initial_postgres(dir: &Path, pool: &sqlx::PgPool) -> Result<u64, MigrateError> {
3434 ensure_tracking_table_postgres(pool).await?;
3435 let applied = applied_names_postgres(pool).await?;
3436 let mut count: u64 = 0;
3437
3438 for plugin in plugin_order() {
3439 let plugin_dir = dir.join(&plugin);
3440 let paths = list_migration_files(&plugin_dir)?;
3441
3442 let first = paths.first();
3443 let first = match first {
3444 Some(p) => p,
3445 None => continue,
3446 };
3447 let file = read_migration_file(first)?;
3448
3449 if applied.contains(&(file.plugin.clone(), file.id.clone())) {
3450 continue;
3451 }
3452
3453 let tables_to_create: Vec<&str> = file
3454 .operations
3455 .iter()
3456 .filter_map(|op| match op {
3457 Operation::CreateTable { table, .. } => Some(table.as_str()),
3458 _ => None,
3459 })
3460 .collect();
3461
3462 if tables_to_create.is_empty() {
3463 continue;
3464 }
3465
3466 let mut all_present = true;
3467 for table in &tables_to_create {
3468 let exists: Option<(String,)> = sqlx::query_as(
3469 "SELECT table_name FROM information_schema.tables \
3470 WHERE table_schema = 'public' AND table_name = $1",
3471 )
3472 .bind(*table)
3473 .fetch_optional(pool)
3474 .await?;
3475 if exists.is_none() {
3476 all_present = false;
3477 break;
3478 }
3479 }
3480
3481 if all_present {
3482 let snapshot_hash = file.snapshot_after.hash();
3483 let applied_at = chrono::Utc::now().to_rfc3339();
3484 sqlx::query(
3485 "INSERT INTO umbral_migrations \
3486 (plugin, name, applied_at, snapshot_hash) VALUES ($1, $2, $3, $4) \
3487 ON CONFLICT (plugin, name) DO NOTHING",
3488 )
3489 .bind(&file.plugin)
3490 .bind(&file.id)
3491 .bind(&applied_at)
3492 .bind(&snapshot_hash)
3493 .execute(pool)
3494 .await?;
3495 count += 1;
3496 }
3497 }
3498
3499 Ok(count)
3500}
3501
3502/// Print the per-migration state, applied or pending. Output goes to
3503/// stdout; the return value is the count of pending migrations so a
3504/// CLI can `exit(n)` on need.
3505pub async fn show() -> Result<u64, MigrateError> {
3506 show_in(Path::new(MIGRATIONS_DIR)).await
3507}
3508
3509/// Read-only migration status for a health / readiness probe (Kikosi #5,
3510/// gaps3 #38).
3511///
3512/// Returns the same [`DriftReport`] `showmigrations` computes — one entry per
3513/// migration, categorised into the four [`MigrationStatus`] states — but prints
3514/// nothing, so it is safe to call on every `/readyz` hit. Reads the
3515/// `umbral_migrations` tracking table (ensuring it exists) and the on-disk
3516/// `migrations/` tree. Use [`DriftReport::pending`] to gate readiness.
3517pub async fn drift_report() -> Result<DriftReport, MigrateError> {
3518 drift_report_in(Path::new(MIGRATIONS_DIR)).await
3519}
3520
3521/// [`drift_report`] against an explicit migrations directory. The entry tests
3522/// drive this so they can point at a temp tree.
3523pub async fn drift_report_in(dir: &Path) -> Result<DriftReport, MigrateError> {
3524 let applied = match crate::db::pool_dispatched() {
3525 crate::db::DbPool::Sqlite(pool) => {
3526 ensure_tracking_table_sqlite(pool).await?;
3527 applied_names_sqlite(pool).await?
3528 }
3529 crate::db::DbPool::Postgres(pool) => {
3530 ensure_tracking_table_postgres(pool).await?;
3531 applied_names_postgres(pool).await?
3532 }
3533 };
3534 detect_all_drift(&applied, dir)
3535}
3536
3537/// Same as [`show`] but takes an explicit base directory. Walks every
3538/// registered plugin in sorted-by-name order, printing one section per
3539/// plugin that owns at least one migration file; empty plugins are
3540/// skipped silently rather than emitting a bare header.
3541///
3542/// Four-state output (gap 24):
3543///
3544/// - `[X]` applied and file present on disk (normal)
3545/// - `[ ]` pending (on disk, not yet applied, sequence after last applied)
3546/// - `[!]` applied but missing on disk (drift — tracking table ahead of VCS)
3547/// - `[?]` on disk but out of order (sequence before last applied, not in DB)
3548pub async fn show_in(dir: &Path) -> Result<u64, MigrateError> {
3549 let report = drift_report_in(dir).await?;
3550
3551 // Group by plugin for display.
3552 let mut by_plugin: std::collections::BTreeMap<&str, Vec<&MigrationEntry>> =
3553 std::collections::BTreeMap::new();
3554 for entry in &report.entries {
3555 by_plugin
3556 .entry(entry.plugin.as_str())
3557 .or_default()
3558 .push(entry);
3559 }
3560
3561 let mut pending: u64 = 0;
3562 for (plugin, entries) in &by_plugin {
3563 if entries.is_empty() {
3564 continue;
3565 }
3566 println!("# plugin: {plugin}");
3567 for entry in entries {
3568 let marker = match entry.status {
3569 MigrationStatus::Applied => "[X]",
3570 MigrationStatus::Pending => {
3571 pending += 1;
3572 "[ ]"
3573 }
3574 MigrationStatus::AppliedButMissing => "[!]",
3575 MigrationStatus::OutOfOrder => "[?]",
3576 };
3577 println!("{marker} {}/{}", entry.plugin, entry.name);
3578 }
3579 }
3580 Ok(pending)
3581}
3582
3583/// Safety classification for a single pending migration operation.
3584///
3585/// Feature #65 (blue-green / zero-downtime). The `checkmigrations`
3586/// command walks every pending operation and tags it so an operator
3587/// deploying without a maintenance window can tell which changes are safe
3588/// under a rolling deploy (old and new code serving traffic at once) and
3589/// which need the expand-contract dance. This is advisory triage — the
3590/// engine still *applies* every op exactly as written; nothing here gates
3591/// `migrate`.
3592#[derive(Debug, Clone, PartialEq, Eq)]
3593pub enum OpSafety {
3594 /// Additive and backward-compatible — safe while old code still runs.
3595 Safe,
3596 /// Applies cleanly but can break still-running old code, lock a large
3597 /// table, or fail against unexpected production data. Review first.
3598 Warning(String),
3599 /// Destroys data or is irreversible; old code referencing the dropped
3600 /// surface errors immediately.
3601 Unsafe(String),
3602}
3603
3604impl OpSafety {
3605 /// The advisory reason for a `Warning` / `Unsafe`; empty for `Safe`.
3606 pub fn reason(&self) -> &str {
3607 match self {
3608 OpSafety::Safe => "",
3609 OpSafety::Warning(r) | OpSafety::Unsafe(r) => r,
3610 }
3611 }
3612
3613 /// True for the destructive / irreversible tier only.
3614 pub fn is_unsafe(&self) -> bool {
3615 matches!(self, OpSafety::Unsafe(_))
3616 }
3617
3618 /// True for the review-before-deploy tier only.
3619 pub fn is_warning(&self) -> bool {
3620 matches!(self, OpSafety::Warning(_))
3621 }
3622}
3623
3624/// One pending operation tagged with its [`OpSafety`] and the migration
3625/// that introduced it. The unit of output for `checkmigrations`.
3626#[derive(Debug, Clone)]
3627pub struct ClassifiedOp {
3628 pub plugin: String,
3629 pub migration: String,
3630 pub op: Operation,
3631 pub safety: OpSafety,
3632}
3633
3634/// Classify one operation for zero-downtime safety. Pure — no DB access,
3635/// no file reads — so it is trivially unit-testable and reused by both
3636/// the CLI report and any plugin that wants to gate its own deploys.
3637pub fn classify_operation(op: &Operation) -> OpSafety {
3638 match op {
3639 // Brand-new tables touch no existing rows and no old code reads
3640 // them yet.
3641 Operation::CreateTable { .. } | Operation::CreateM2MTable { .. } => OpSafety::Safe,
3642
3643 // A column comment is metadata: no rows read or written, no lock worth
3644 // naming, and no code path — old or new — reads it. A docstring edit
3645 // must never gate a zero-downtime deploy.
3646 Operation::SetColumnComment { .. } => OpSafety::Safe,
3647
3648 // Adding a column is additive — unless it's NOT NULL with no
3649 // default, in which case old code inserting a row without the
3650 // column fails. (The engine refuses such an add against a
3651 // populated SQLite table at apply time; this surfaces the same
3652 // hazard *before* the operator runs it, and for Postgres too.)
3653 Operation::AddColumn { table, column } => {
3654 if !column.nullable && column.default.is_empty() {
3655 OpSafety::Warning(format!(
3656 "adds NOT NULL column `{}.{}` with no default — old code inserting without it will fail. Add it nullable (or with a default), backfill, then tighten",
3657 table, column.name
3658 ))
3659 } else {
3660 OpSafety::Safe
3661 }
3662 }
3663
3664 // A view holds no rows of its own (a materialized one holds only derived,
3665 // recomputable rows), so creating one is free and dropping one destroys
3666 // nothing. What a DropView CAN break is old code still selecting from it —
3667 // and since every view edit lowers to drop+create, classifying this Unsafe
3668 // would trip the zero-downtime gate on a routine SQL tweak. Warning is the
3669 // honest severity: worth reading, not worth blocking.
3670 Operation::CreateView { .. } => OpSafety::Safe,
3671 Operation::DropView { name, .. } => OpSafety::Warning(format!(
3672 "drops view `{name}` — no rows are lost (a view stores none), but old code \
3673 still selecting from it breaks until the recreate lands in the same migration"
3674 )),
3675
3676 // Destructive / irreversible: data loss the moment it runs.
3677 Operation::DropTable { table } => OpSafety::Unsafe(format!(
3678 "drops table `{table}` and every row in it — irreversible, and old code still reading it breaks. Stop using it, deploy, then drop in a later migration"
3679 )),
3680 Operation::DropM2MTable { junction_table } => OpSafety::Unsafe(format!(
3681 "drops join table `{junction_table}` and every row in it — irreversible"
3682 )),
3683 Operation::DropColumn { table, column } => OpSafety::Unsafe(format!(
3684 "drops column `{table}.{column}` and its data — old code reading it breaks. Expand-contract: stop writing it, deploy, then drop"
3685 )),
3686
3687 // Renames apply atomically in the DB but NOT atomically with a
3688 // code deploy: between the migration and the rollout, one of the
3689 // two code versions references the missing name.
3690 Operation::RenameTable { from, to } => OpSafety::Warning(format!(
3691 "renames table `{from}` → `{to}` — not atomic with a code deploy; old code references `{from}`. Expand-contract: add `{to}`, dual-write, switch, then drop `{from}`"
3692 )),
3693 Operation::RenameColumn {
3694 table, from, to, ..
3695 } => OpSafety::Warning(format!(
3696 "renames column `{table}.{from}` → `{to}` — old code references `{from}`. Expand-contract: add `{to}`, backfill, switch reads, then drop `{from}`"
3697 )),
3698
3699 // An alter can rewrite a column (table lock on large data) and a
3700 // nullable→NOT NULL tightening fails on existing NULLs.
3701 Operation::AlterColumn { table, column, .. } => OpSafety::Warning(format!(
3702 "alters column `{table}.{column}` — a type change rewrites the column (locks the table on large data) and a NOT NULL tightening fails on existing NULLs; verify against production data first"
3703 )),
3704
3705 // A hand-authored data migration runs arbitrary SQL — the
3706 // engine can't reason about its row impact, so flag it for
3707 // human review (it may rewrite or delete data, and re-running
3708 // the rollout while it's mid-flight can double-apply).
3709 Operation::RunSql { .. } => OpSafety::Warning(
3710 "runs a hand-authored data migration (raw SQL) — review its row impact, ensure it's idempotent or guarded, and verify it against production data first".to_string(),
3711 ),
3712
3713 // Adding a composite UNIQUE constraint fails at apply time if
3714 // existing rows already violate it — same hazard as a single-column
3715 // UNIQUE add. A plain (non-unique) index is purely additive.
3716 Operation::AddIndex {
3717 table,
3718 columns,
3719 unique: true,
3720 } => OpSafety::Warning(format!(
3721 "adds a composite UNIQUE constraint on `{table}` ({}) — fails on existing duplicate rows; de-duplicate first or the migration aborts",
3722 columns.join(", ")
3723 )),
3724 Operation::AddIndex { unique: false, .. } => OpSafety::Safe,
3725
3726 // Dropping an index / UNIQUE constraint touches no rows. It removes
3727 // a guarantee (a later duplicate becomes insertable) but that is the
3728 // intent when a `unique_together` is removed, and no data is lost.
3729 Operation::DropIndex { .. } => OpSafety::Safe,
3730 }
3731}
3732
3733/// Classify every operation across all pending migrations against the
3734/// ambient pool. Reads the same applied-set + on-disk diff that
3735/// `migrate` / `showmigrations` use, then loads each pending migration
3736/// file and classifies its operations in order. Powers `checkmigrations`.
3737pub async fn check_pending_safety() -> Result<Vec<ClassifiedOp>, MigrateError> {
3738 check_pending_safety_in(Path::new(MIGRATIONS_DIR)).await
3739}
3740
3741/// [`check_pending_safety`] against an explicit migrations directory.
3742/// The seam tests use to point at a fixture tree.
3743pub async fn check_pending_safety_in(dir: &Path) -> Result<Vec<ClassifiedOp>, MigrateError> {
3744 let applied = match crate::db::pool_dispatched() {
3745 crate::db::DbPool::Sqlite(pool) => {
3746 ensure_tracking_table_sqlite(pool).await?;
3747 applied_names_sqlite(pool).await?
3748 }
3749 crate::db::DbPool::Postgres(pool) => {
3750 ensure_tracking_table_postgres(pool).await?;
3751 applied_names_postgres(pool).await?
3752 }
3753 };
3754
3755 let report = detect_all_drift(&applied, dir)?;
3756
3757 let mut out: Vec<ClassifiedOp> = Vec::new();
3758 for entry in &report.entries {
3759 if entry.status != MigrationStatus::Pending {
3760 continue;
3761 }
3762 let path = dir.join(&entry.plugin).join(format!("{}.json", entry.name));
3763 let file = read_migration_file(&path)?;
3764 for op in &file.operations {
3765 out.push(ClassifiedOp {
3766 plugin: entry.plugin.clone(),
3767 migration: entry.name.clone(),
3768 op: op.clone(),
3769 safety: classify_operation(op),
3770 });
3771 }
3772 }
3773 Ok(out)
3774}
3775
3776// =========================================================================
3777// Internal helpers. Crate-private; the public surface above is the only
3778// thing the rest of umbral calls into.
3779// =========================================================================
3780
3781/// Return every `*.json` migration file in `plugin_dir`, sorted by
3782/// filename (lexical sort matches numeric order because the prefix is
3783/// zero-padded). Returns an empty vec if the directory is missing.
3784fn list_migration_files(plugin_dir: &Path) -> Result<Vec<PathBuf>, MigrateError> {
3785 if !plugin_dir.exists() {
3786 return Ok(Vec::new());
3787 }
3788 let mut paths: Vec<PathBuf> = Vec::new();
3789 for entry in std::fs::read_dir(plugin_dir)? {
3790 let entry = entry?;
3791 let path = entry.path();
3792 if path.extension().and_then(|s| s.to_str()) == Some("json") {
3793 paths.push(path);
3794 }
3795 }
3796 paths.sort();
3797 Ok(paths)
3798}
3799
3800/// Read and parse one migration file.
3801fn read_migration_file(path: &Path) -> Result<MigrationFile, MigrateError> {
3802 let text = std::fs::read_to_string(path)?;
3803 let file: MigrationFile = serde_json::from_str(&text)?;
3804 Ok(file)
3805}
3806
3807/// Diff the previous snapshot against the current one and produce the
3808/// ordered operation list.
3809///
3810/// Emits `CreateTable` / `DropTable` for whole-model changes (M5 v1),
3811/// and `AddColumn` / `DropColumn` for column-level changes on a model
3812/// that appears in both snapshots (M8 v1). A column whose name stays
3813/// the same but whose type or nullable flag changed surfaces as
3814/// [`MigrateError::UnsafeAlter`]: SQLite can't ALTER COLUMN TYPE in
3815/// place, and a nullable flip on a populated table is destructive.
3816///
3817/// Gap 30 adds two-pass rename detection. `Model::NAME` (the Rust struct
3818/// name) is the stable identity key across snapshots; the SQL table name
3819/// in `Model::TABLE` may change (e.g. via the `#[umbral(plugin = "...")]`
3820/// opt-in). The two passes are:
3821///
3822/// - **First pass — struct-name match.** If a model present in `current`
3823/// but absent from `previous` (by `Model::NAME`) has the same NAME as
3824/// a model present in `previous` but absent from `current`, the table
3825/// name changed: emit `RenameTable { from, to }` instead of DropTable +
3826/// CreateTable. A stdout message names the rename so the developer can
3827/// audit `makemigrations` output.
3828/// - **Second pass — column-shape match.** Among unpaired drops and
3829/// creates, if a drop candidate and a create candidate have bit-identical
3830/// column shapes (same column names, types, nullable, fk_target), emit
3831/// `RenameTable` and log a warning so the developer can verify the
3832/// intent. Struct names differ; the shape heuristic fills in for cases
3833/// like a wholesale model rename (Foo → Bar, identical fields).
3834/// - **No-match.** Drop and create as today.
3835///
3836/// `pub` (not `pub(crate)`) so integration tests can drive the diff
3837/// directly with hand-built snapshots. Spec 06 calls the diff the
3838/// engine's contract; exposing it lets the tests pin every scenario
3839/// without laundering snapshots through the process-wide registry.
3840/// audit_2 H23 — how `diff` should treat an ambiguous column-shape rename: an
3841/// unpaired dropped model and an unpaired created model with *identical* shapes.
3842/// Driven by `UMBRAL_MIGRATIONS_ASSUME_RENAMES`.
3843enum RenameIntent {
3844 /// Auto-pair every shape match into a `RenameTable` (the pre-H23 default).
3845 Assume,
3846 /// Treat the pair as unrelated: emit drop + create, no row transfer.
3847 Independent,
3848 /// Not configured → `diff` fails closed with `AmbiguousRename`.
3849 Undecided,
3850}
3851
3852fn rename_intent() -> RenameIntent {
3853 match std::env::var("UMBRAL_MIGRATIONS_ASSUME_RENAMES") {
3854 Ok(v) => match v.trim().to_ascii_lowercase().as_str() {
3855 "assume" | "1" | "true" | "yes" | "rename" => RenameIntent::Assume,
3856 "independent" | "0" | "false" | "no" | "drop" => RenameIntent::Independent,
3857 _ => RenameIntent::Undecided,
3858 },
3859 Err(_) => RenameIntent::Undecided,
3860 }
3861}
3862
3863/// Emit `AddIndex` / `DropIndex` ops for changes to a model's TABLE-level
3864/// `unique_together` and `indexes` between two snapshots.
3865///
3866/// Column-shape changes are handled by [`diff_columns`]; this covers the
3867/// constraint-only deltas it can't see. Before this existed, adding a
3868/// `unique_together` group or a multi-column `indexes` entry with no column
3869/// change produced NO migration at all — `makemigrations` said "no changes"
3870/// and the constraint was silently never created.
3871///
3872/// A group is identified by its ORDERED column list, so re-ordering a group
3873/// (`[a, b]` → `[b, a]`) reads as drop-old + add-new — correct, since column
3874/// order changes which queries the index serves. `AddIndex`/`DropIndex`
3875/// render `IF NOT EXISTS` / `IF EXISTS`, so when a same-table `AlterColumn`
3876/// in the same migration already rebuilt the table with the new constraint
3877/// set (the SQLite dance), these ops are harmless no-ops.
3878fn diff_indexes(previous: &ModelMeta, current: &ModelMeta) -> Vec<Operation> {
3879 use std::collections::BTreeSet;
3880
3881 let mut ops: Vec<Operation> = Vec::new();
3882
3883 let mut emit = |prev_groups: &[Vec<String>], curr_groups: &[Vec<String>], unique: bool| {
3884 let prev_set: BTreeSet<&Vec<String>> = prev_groups.iter().collect();
3885 let curr_set: BTreeSet<&Vec<String>> = curr_groups.iter().collect();
3886 // Added groups → AddIndex (on the current table name).
3887 for group in curr_set.difference(&prev_set) {
3888 ops.push(Operation::AddIndex {
3889 table: current.table.clone(),
3890 columns: (*group).clone(),
3891 unique,
3892 });
3893 }
3894 // Removed groups → DropIndex (on the previous table name — a rename
3895 // in the same diff emits its own RenameTable first, and the index
3896 // travels with the table, so the drop names the post-rename table;
3897 // use current.table for that reason).
3898 for group in prev_set.difference(&curr_set) {
3899 ops.push(Operation::DropIndex {
3900 table: current.table.clone(),
3901 columns: (*group).clone(),
3902 unique,
3903 });
3904 }
3905 };
3906
3907 emit(&previous.unique_together, ¤t.unique_together, true);
3908 emit(&previous.indexes, ¤t.indexes, false);
3909
3910 // Single-column `#[umbral(index)]` flag flips. A column that gains the
3911 // flag → AddIndex; one that loses it (but still exists) → DropIndex. PK
3912 // and UNIQUE columns are excluded — they carry their own index, matching
3913 // `should_emit_btree_index`. A column that was DROPPED entirely takes its
3914 // index with it (DROP COLUMN cascades on both backends), so we don't emit a
3915 // redundant DropIndex for it. The `idx_<table>_<col>` name matches the one
3916 // `create_index_stmt` uses at CreateTable time, so add/drop stay symmetric.
3917 let indexed = |m: &ModelMeta| -> std::collections::BTreeSet<String> {
3918 m.fields
3919 .iter()
3920 .filter(|c| c.index && !c.primary_key && !c.unique)
3921 .map(|c| c.name.clone())
3922 .collect()
3923 };
3924 let curr_col_names: std::collections::BTreeSet<&str> =
3925 current.fields.iter().map(|c| c.name.as_str()).collect();
3926 let prev_indexed = indexed(previous);
3927 let curr_indexed = indexed(current);
3928 for name in curr_indexed.difference(&prev_indexed) {
3929 ops.push(Operation::AddIndex {
3930 table: current.table.clone(),
3931 columns: vec![name.clone()],
3932 unique: false,
3933 });
3934 }
3935 for name in prev_indexed.difference(&curr_indexed) {
3936 if curr_col_names.contains(name.as_str()) {
3937 ops.push(Operation::DropIndex {
3938 table: current.table.clone(),
3939 columns: vec![name.clone()],
3940 unique: false,
3941 });
3942 }
3943 }
3944 ops
3945}
3946
3947pub fn diff(previous: &Snapshot, current: &Snapshot) -> Result<Vec<Operation>, MigrateError> {
3948 use std::collections::{BTreeMap, HashSet};
3949
3950 // features #73 — views are diffed on a separate track, for a reason that runs
3951 // deeper than tidiness: every pass below exists to AVOID destroying data (the
3952 // rename heuristic, the safe-cast table, the data-preserving ALTER). A view
3953 // stores nothing, so none of that machinery applies to it, and letting a view
3954 // through it would be actively wrong — Pass 1 could "detect a rename" between a
3955 // dropped table and a created view and emit an ALTER against a relation that has
3956 // no rows to alter.
3957 //
3958 // So: hide the views from the table passes entirely, then handle them in Pass 5
3959 // with the only two operations a view can ever need.
3960 let view_prev = previous;
3961 let view_curr = current;
3962 let prev_tables = Snapshot {
3963 models: previous
3964 .models
3965 .iter()
3966 .filter(|m| m.view.is_none())
3967 .cloned()
3968 .collect(),
3969 };
3970 let curr_tables = Snapshot {
3971 models: current
3972 .models
3973 .iter()
3974 .filter(|m| m.view.is_none())
3975 .cloned()
3976 .collect(),
3977 };
3978 let previous = &prev_tables;
3979 let current = &curr_tables;
3980
3981 let prev_by_name: BTreeMap<&str, &ModelMeta> = previous
3982 .models
3983 .iter()
3984 .map(|m| (m.name.as_str(), m))
3985 .collect();
3986 let curr_by_name: BTreeMap<&str, &ModelMeta> = current
3987 .models
3988 .iter()
3989 .map(|m| (m.name.as_str(), m))
3990 .collect();
3991
3992 let mut ops: Vec<Operation> = Vec::new();
3993
3994 // gaps.md #93: every table rename `diff` decides on (old_table → new_table),
3995 // recorded so Pass 4 can rename the parent's M2M junctions instead of
3996 // dropping + recreating them (which would destroy every relationship row).
3997 let mut renamed_tables: BTreeMap<String, String> = BTreeMap::new();
3998
3999 // ---- Pass 0: Walk models present in both snapshots (same NAME). ----
4000 // Same-name models with a different table produce a first-pass rename.
4001 // Same-name models with identical table+columns produce nothing.
4002 // Same-name models with column changes produce column-level ops.
4003
4004 let mut drop_candidates: Vec<&ModelMeta> = Vec::new(); // in prev, not curr
4005 let mut create_candidates: Vec<&ModelMeta> = Vec::new(); // in curr, not prev
4006
4007 // Creates and column-level diffs, in deterministic name order.
4008 for (name, curr) in &curr_by_name {
4009 match prev_by_name.get(name) {
4010 None => {
4011 // In current but not previous — might be a create or a first-pass rename.
4012 create_candidates.push(curr);
4013 }
4014 Some(prev) if prev.table != curr.table => {
4015 // Same struct name, different table name → first-pass rename.
4016 println!(
4017 "umbral makemigrations: rename detected (struct-name match): \
4018 table `{}` → `{}`",
4019 prev.table, curr.table
4020 );
4021 ops.push(Operation::RenameTable {
4022 from: prev.table.clone(),
4023 to: curr.table.clone(),
4024 });
4025 renamed_tables.insert(prev.table.clone(), curr.table.clone());
4026 // After the rename the columns might also have changed; diff them.
4027 let col_ops = diff_columns(name, prev, curr)?;
4028 ops.extend(col_ops);
4029 // ...and the table-level constraints (unique_together/indexes).
4030 ops.extend(diff_indexes(prev, curr));
4031 }
4032 Some(prev) if prev == curr => {}
4033 Some(prev) => {
4034 ops.extend(diff_columns(name, prev, curr)?);
4035 // Constraint-only deltas (a new/removed unique_together or
4036 // composite index with no column change) — otherwise silent.
4037 ops.extend(diff_indexes(prev, curr));
4038 }
4039 }
4040 }
4041
4042 // Drops — models in prev but not curr (by NAME).
4043 for (name, prev) in &prev_by_name {
4044 if !curr_by_name.contains_key(name) {
4045 drop_candidates.push(prev);
4046 }
4047 }
4048
4049 // ---- Pass 1: Column-shape heuristic for unpaired drops + creates. ----
4050 // A sorted, canonical serialisation of (name, ty, nullable, fk_target)
4051 // is the "shape" fingerprint. Bit-identical shapes → likely a model
4052 // rename where the struct name also changed.
4053
4054 // audit_2 H23 — a bit-identical column shape between a dropped and a
4055 // created model is genuinely ambiguous (rename vs. two unrelated models).
4056 // The pre-H23 code auto-emitted a `RenameTable` with only an `eprintln!`,
4057 // silently handing one model's rows to another and skipping the intended
4058 // drop. Now the ambiguity is resolved by explicit operator intent: `assume`
4059 // restores auto-pairing, `independent` treats the pair as unrelated
4060 // (drop + create), and the unset default fails closed so no destructive
4061 // guess is ever applied silently.
4062 let intent = rename_intent();
4063 let mut paired_drop_tables: HashSet<&str> = HashSet::new();
4064 let mut paired_create_tables: HashSet<&str> = HashSet::new();
4065
4066 'creates: for create in &create_candidates {
4067 let create_shape = column_shape(&create.fields);
4068 for drop in &drop_candidates {
4069 if paired_drop_tables.contains(drop.table.as_str()) {
4070 continue;
4071 }
4072 if column_shape(&drop.fields) != create_shape {
4073 continue;
4074 }
4075 match intent {
4076 RenameIntent::Assume => {
4077 eprintln!(
4078 "umbral makemigrations: rename ASSUMED (column-shape match): \
4079 `{}` → `{}` — moving the old table's rows to the new name. Set \
4080 UMBRAL_MIGRATIONS_ASSUME_RENAMES=independent if these are unrelated \
4081 models that merely share a shape.",
4082 drop.table, create.table
4083 );
4084 ops.push(Operation::RenameTable {
4085 from: drop.table.clone(),
4086 to: create.table.clone(),
4087 });
4088 renamed_tables.insert(drop.table.clone(), create.table.clone());
4089 paired_drop_tables.insert(drop.table.as_str());
4090 paired_create_tables.insert(create.table.as_str());
4091 continue 'creates;
4092 }
4093 RenameIntent::Independent => {
4094 eprintln!(
4095 "umbral makemigrations: column-shape match `{}` ↔ `{}` treated as \
4096 UNRELATED (drop + create) per UMBRAL_MIGRATIONS_ASSUME_RENAMES=\
4097 independent. Set it to `assume` (or hand-write a RenameTable) if this \
4098 is actually a rename — otherwise `{}`'s rows are dropped.",
4099 drop.table, create.table, drop.table
4100 );
4101 // Leave both unpaired → Pass 2 creates, Pass 3 drops.
4102 continue 'creates;
4103 }
4104 RenameIntent::Undecided => {
4105 return Err(MigrateError::AmbiguousRename {
4106 from_table: drop.table.clone(),
4107 to_table: create.table.clone(),
4108 });
4109 }
4110 }
4111 }
4112 }
4113
4114 // ---- Pass 2: Emit plain CreateTable for unpaired creates. ----
4115 //
4116 // Sort the create list topologically by FK dependency so that a
4117 // table referenced by another table in this batch is created first.
4118 // Without this, Postgres rejects the second CreateTable with
4119 // `relation "<target>" does not exist`. (SQLite tolerates the wrong
4120 // order when `foreign_keys=OFF`, the historical default; once
4121 // we turned foreign_keys ON in connect_sqlite, SQLite agrees with
4122 // Postgres on the order requirement.)
4123 //
4124 // Kahn's algorithm on (table → set of FK-target tables that are
4125 // ALSO in the create batch). Self-references and FK targets outside
4126 // the batch are skipped (they're either harmless or already exist
4127 // by the time this migration runs).
4128 let creates: Vec<&&ModelMeta> = create_candidates
4129 .iter()
4130 .filter(|c| !paired_create_tables.contains(c.table.as_str()))
4131 .collect();
4132 let batch_tables: HashSet<&str> = creates.iter().map(|c| c.table.as_str()).collect();
4133 let mut deps: BTreeMap<&str, HashSet<&str>> = BTreeMap::new();
4134 for create in &creates {
4135 let mut in_batch: HashSet<&str> = HashSet::new();
4136 for col in &create.fields {
4137 if let Some(target) = col.fk_target.as_deref()
4138 && target != create.table.as_str()
4139 && batch_tables.contains(target)
4140 {
4141 in_batch.insert(target);
4142 }
4143 }
4144 deps.insert(create.table.as_str(), in_batch);
4145 }
4146 // Kahn: repeatedly pop tables with no remaining deps in the batch.
4147 // BTreeMap iteration is alphabetical → ties break alphabetically,
4148 // keeping the output stable.
4149 let mut ordered: Vec<&&ModelMeta> = Vec::with_capacity(creates.len());
4150 while !deps.is_empty() {
4151 let ready: Vec<&str> = deps
4152 .iter()
4153 .filter(|(_, d)| d.is_empty())
4154 .map(|(t, _)| *t)
4155 .collect();
4156 if ready.is_empty() {
4157 // Cyclic FK or other unresolvable dep — fall through to
4158 // the original order rather than dropping models. A cycle
4159 // here means the user's schema can't be created with
4160 // plain CreateTable anyway (Postgres needs deferrable
4161 // constraints), so we surface the user-visible error at
4162 // apply time instead of silently looping.
4163 for create in &creates {
4164 if deps.contains_key(create.table.as_str()) {
4165 ordered.push(create);
4166 }
4167 }
4168 break;
4169 }
4170 for t in &ready {
4171 if let Some(create) = creates.iter().find(|c| c.table.as_str() == *t) {
4172 ordered.push(create);
4173 }
4174 deps.remove(t);
4175 }
4176 for (_, set) in deps.iter_mut() {
4177 for t in &ready {
4178 set.remove(t);
4179 }
4180 }
4181 }
4182 for create in ordered {
4183 ops.push(Operation::CreateTable {
4184 table: create.table.clone(),
4185 columns: create.fields.clone(),
4186 unique_together: create.unique_together.clone(),
4187 indexes: create.indexes.clone(),
4188 });
4189 }
4190
4191 // ---- Pass 3: Emit plain DropTable for unpaired drops. ----
4192 for drop in &drop_candidates {
4193 if !paired_drop_tables.contains(drop.table.as_str()) {
4194 ops.push(Operation::DropTable {
4195 table: drop.table.clone(),
4196 });
4197 }
4198 }
4199
4200 // ---- Pass 4: Diff M2M relations. Closes the remaining BUG-16 gap. ----
4201 //
4202 // Treat each (parent_table, field_name) pair as a junction-table identity.
4203 // Compare the flattened set across snapshots and emit CreateM2MTable /
4204 // DropM2MTable per delta.
4205 //
4206 // gaps.md #93: when the PARENT model was renamed (Pass 0/1), its junction's
4207 // key moves from `(old_table, field)` to `(new_table, field)` — which would
4208 // otherwise read as one junction dropped and a different one created,
4209 // destroying every relationship row. Detect that case via `renamed_tables`
4210 // and emit a plain `RenameTable` on the junction instead. The junction's
4211 // columns are generic (`parent_id`/`child_id`) and its FK to the parent is
4212 // auto-updated by the parent's own rename, so a table rename is sufficient.
4213 let prev_m2m = collect_m2m_pairs(previous);
4214 let curr_m2m = collect_m2m_pairs(current);
4215
4216 // Prev junction keys consumed by a junction rename below — skip their drop.
4217 let mut renamed_prev_junctions: HashSet<(String, String)> = HashSet::new();
4218
4219 for (key, spec) in &curr_m2m {
4220 if prev_m2m.contains_key(key) {
4221 continue;
4222 }
4223 let (new_parent, field) = key;
4224 // Was this junction's parent renamed FROM some old table? If the prior
4225 // snapshot has the same field on that old parent, targeting the same
4226 // table, it's a junction rename, not a fresh create.
4227 let renamed_from = renamed_tables
4228 .iter()
4229 .find(|(_, new)| *new == new_parent)
4230 .and_then(|(old, _)| {
4231 let old_key = (old.clone(), field.clone());
4232 prev_m2m.get(&old_key).and_then(|old_spec| {
4233 (old_spec.target_table == spec.target_table).then_some(old_key)
4234 })
4235 });
4236 if let Some(old_key) = renamed_from {
4237 let old_junction = prev_m2m[&old_key].junction_table.clone();
4238 ops.push(Operation::RenameTable {
4239 from: old_junction,
4240 to: spec.junction_table.clone(),
4241 });
4242 renamed_prev_junctions.insert(old_key);
4243 continue;
4244 }
4245 // New M2M field on an existing or new model. Resolve the
4246 // target's PK column from the current snapshot.
4247 match build_create_m2m_op(spec, current) {
4248 Ok(op) => ops.push(op),
4249 Err(e) => return Err(e),
4250 }
4251 }
4252 for (key, spec) in &prev_m2m {
4253 if curr_m2m.contains_key(key) || renamed_prev_junctions.contains(key) {
4254 continue;
4255 }
4256 // M2M field removed (or its parent was dropped). The junction
4257 // table goes away.
4258 ops.push(Operation::DropM2MTable {
4259 junction_table: spec.junction_table.clone(),
4260 });
4261 }
4262
4263 // ---- Pass 5: views (features #73). ----
4264 //
4265 // Ordering is the whole feature. Drops go BEFORE the table ops and creates go
4266 // AFTER, because Postgres refuses to drop or retype a column that a live view
4267 // selects from — so a migration that alters `matchday.score` while a view reads
4268 // it fails unless the view is out of the way first, and is only whole again once
4269 // the view is back.
4270 let (view_drops, view_creates) = diff_views(view_prev, view_curr, &ops)?;
4271 let mut all = view_drops;
4272 all.extend(ops);
4273 all.extend(view_creates);
4274 Ok(all)
4275}
4276
4277/// The relations a view's SQL selects from, recovered by scanning its body for the
4278/// names of relations we know about.
4279///
4280/// The SELECT body is an opaque string — nobody is writing a SQL parser here. But
4281/// the engine already knows every table and view name in the snapshot, so it can ask
4282/// the cheaper question: *does this view's SQL mention `matchday`?* A whole-word
4283/// match (so `matchday` does not match `matchday_team`) recovers the dependency edge
4284/// that ordering needs.
4285///
4286/// The failure modes are asymmetric, which is why this is worth doing:
4287/// - A **false positive** (the name appears in a string literal or a comment) costs
4288/// one needless `DROP VIEW` + `CREATE VIEW`. A view stores nothing. Nobody notices.
4289/// - A **false negative** would order a view before the table it reads, and the
4290/// migration fails at apply time with a clear "relation does not exist".
4291///
4292/// Both are safe. Neither can corrupt data. That asymmetry is what makes a heuristic
4293/// acceptable here and unacceptable in the table passes next door.
4294fn view_dependencies(sql: &str, known: &[String], self_table: &str) -> Vec<String> {
4295 let haystack = sql.to_ascii_lowercase();
4296 let bytes = haystack.as_bytes();
4297 let is_word = |b: u8| b.is_ascii_alphanumeric() || b == b'_';
4298
4299 let mut deps = Vec::new();
4300 for name in known {
4301 if name == self_table {
4302 continue;
4303 }
4304 let needle = name.to_ascii_lowercase();
4305 if needle.is_empty() {
4306 continue;
4307 }
4308 let mut from = 0usize;
4309 while let Some(hit) = haystack[from..].find(&needle) {
4310 let start = from + hit;
4311 let end = start + needle.len();
4312 let before_ok = start == 0 || !is_word(bytes[start - 1]);
4313 let after_ok = end == bytes.len() || !is_word(bytes[end]);
4314 if before_ok && after_ok {
4315 deps.push(name.clone());
4316 break;
4317 }
4318 from = end;
4319 }
4320 }
4321 deps
4322}
4323
4324/// Diff the view models of two snapshots into (drops, creates).
4325///
4326/// `table_ops` is what the table passes decided; a view that reads from a table
4327/// those ops touch is recreated even if its own SQL never changed.
4328fn diff_views(
4329 previous: &Snapshot,
4330 current: &Snapshot,
4331 table_ops: &[Operation],
4332) -> Result<(Vec<Operation>, Vec<Operation>), MigrateError> {
4333 use std::collections::{BTreeMap, HashSet};
4334
4335 let view_models = |snap: &Snapshot| -> Vec<ModelMeta> {
4336 snap.models
4337 .iter()
4338 .filter(|m| m.view.is_some())
4339 .cloned()
4340 .collect()
4341 };
4342 let prev_views = view_models(previous);
4343 let curr_views = view_models(current);
4344 if prev_views.is_empty() && curr_views.is_empty() {
4345 return Ok((Vec::new(), Vec::new()));
4346 }
4347
4348 let touched: HashSet<&str> = table_ops.iter().map(|o| o.table_name()).collect();
4349 let known: Vec<String> = current.models.iter().map(|m| m.table.clone()).collect();
4350
4351 let deps: BTreeMap<&str, Vec<String>> = curr_views
4352 .iter()
4353 .map(|m| {
4354 (
4355 m.name.as_str(),
4356 view_dependencies(m.view.as_deref().unwrap_or(""), &known, &m.table),
4357 )
4358 })
4359 .collect();
4360
4361 let prev_by_name: BTreeMap<&str, &ModelMeta> =
4362 prev_views.iter().map(|m| (m.name.as_str(), m)).collect();
4363
4364 // Which views are stale? Either their own definition moved, or something they
4365 // read did.
4366 let mut recreate: HashSet<&str> = HashSet::new();
4367 for m in &curr_views {
4368 let definition_changed = match prev_by_name.get(m.name.as_str()) {
4369 None => true, // brand new, or was a table until this change
4370 Some(prev) => {
4371 prev.view != m.view || prev.materialized != m.materialized || prev.table != m.table
4372 }
4373 };
4374 let reads_a_touched_table = deps[m.name.as_str()]
4375 .iter()
4376 .any(|d| touched.contains(d.as_str()));
4377 if definition_changed || reads_a_touched_table {
4378 recreate.insert(m.name.as_str());
4379 }
4380 }
4381
4382 // A view built on a view that is being recreated is itself stale. Iterate to a
4383 // fixpoint rather than assuming one level of nesting.
4384 loop {
4385 let mut grew = false;
4386 for m in &curr_views {
4387 if recreate.contains(m.name.as_str()) {
4388 continue;
4389 }
4390 let reads_a_stale_view = deps[m.name.as_str()].iter().any(|d| {
4391 curr_views
4392 .iter()
4393 .any(|v| &v.table == d && recreate.contains(v.name.as_str()))
4394 });
4395 if reads_a_stale_view {
4396 recreate.insert(m.name.as_str());
4397 grew = true;
4398 }
4399 }
4400 if !grew {
4401 break;
4402 }
4403 }
4404
4405 // ---- Creates, in dependency order (a view after the views it reads). ----
4406 let mut to_create: Vec<&ModelMeta> = curr_views
4407 .iter()
4408 .filter(|m| recreate.contains(m.name.as_str()))
4409 .collect();
4410 to_create.sort_by(|a, b| a.name.cmp(&b.name)); // deterministic before topo
4411
4412 let mut creates: Vec<Operation> = Vec::new();
4413 let mut placed: HashSet<&str> = HashSet::new();
4414 while placed.len() < to_create.len() {
4415 let before = placed.len();
4416 for m in &to_create {
4417 if placed.contains(m.table.as_str()) {
4418 continue;
4419 }
4420 let blocked = deps[m.name.as_str()].iter().any(|d| {
4421 to_create
4422 .iter()
4423 .any(|other| &other.table == d && !placed.contains(other.table.as_str()))
4424 });
4425 if blocked {
4426 continue;
4427 }
4428 placed.insert(m.table.as_str());
4429 creates.push(Operation::CreateView {
4430 name: m.table.clone(),
4431 sql: m.view.clone().unwrap_or_default(),
4432 materialized: m.materialized,
4433 });
4434 }
4435 if placed.len() == before {
4436 // A cycle. Two views cannot select from each other, so this means the
4437 // SQL scan found a self-referential edge we cannot honour. Name the
4438 // views rather than looping forever.
4439 let stuck: Vec<&str> = to_create
4440 .iter()
4441 .filter(|m| !placed.contains(m.table.as_str()))
4442 .map(|m| m.table.as_str())
4443 .collect();
4444 return Err(MigrateError::UnsupportedChange(format!(
4445 "views {stuck:?} form a dependency cycle — each one's SQL selects from \
4446 another in the set, so there is no order in which they can be created. \
4447 Break the cycle by inlining one of the definitions."
4448 )));
4449 }
4450 }
4451
4452 // ---- Drops: every view we are about to recreate, plus every view that is gone
4453 // (removed, or converted back into a table). Emitted in reverse create order so
4454 // a dependant view is dropped before the view it reads.
4455 let mut drops: Vec<Operation> = Vec::new();
4456 let curr_view_names: HashSet<&str> = curr_views.iter().map(|m| m.name.as_str()).collect();
4457 let mut to_drop: Vec<&ModelMeta> = prev_views
4458 .iter()
4459 .filter(|m| {
4460 !curr_view_names.contains(m.name.as_str()) || recreate.contains(m.name.as_str())
4461 })
4462 .collect();
4463 // Reverse-topological by the CREATE order we just computed; anything not in it
4464 // (a removed view) sorts first, since nothing we keep can depend on it.
4465 let create_pos = |table: &str| -> usize {
4466 creates
4467 .iter()
4468 .position(|op| matches!(op, Operation::CreateView { name, .. } if name == table))
4469 .map(|i| usize::MAX - i)
4470 .unwrap_or(0)
4471 };
4472 to_drop.sort_by_key(|m| (create_pos(&m.table), m.table.clone()));
4473 for m in to_drop {
4474 drops.push(Operation::DropView {
4475 name: m.table.clone(),
4476 materialized: m.materialized,
4477 });
4478 }
4479
4480 Ok((drops, creates))
4481}
4482
4483/// A flat-resolved M2M descriptor used by [`diff`] to compare snapshots.
4484/// Owns its strings so it can be keyed in a map without lifetime
4485/// gymnastics.
4486#[derive(Debug, Clone)]
4487struct M2MPair {
4488 parent_table: String,
4489 parent_pk: String,
4490 field_name: String,
4491 target_table: String,
4492 junction_table: String,
4493}
4494
4495/// Walk a snapshot and produce one [`M2MPair`] per declared M2M field.
4496/// Keyed on `(parent_table, field_name)` since that uniquely identifies
4497/// a junction table — two models can't share the same parent_table, and
4498/// one model can't declare two M2M fields with the same name.
4499fn collect_m2m_pairs(snap: &Snapshot) -> std::collections::BTreeMap<(String, String), M2MPair> {
4500 let mut out = std::collections::BTreeMap::new();
4501 for model in &snap.models {
4502 let parent_pk = model
4503 .fields
4504 .iter()
4505 .find(|c| c.primary_key)
4506 .map(|c| c.name.clone())
4507 .unwrap_or_else(|| "id".to_string());
4508 for rel in &model.m2m_relations {
4509 let key = (model.table.clone(), rel.field_name.clone());
4510 out.insert(
4511 key,
4512 M2MPair {
4513 parent_table: model.table.clone(),
4514 parent_pk: parent_pk.clone(),
4515 field_name: rel.field_name.clone(),
4516 target_table: rel.target_table.clone(),
4517 junction_table: format!("{}_{}", model.table, rel.field_name),
4518 },
4519 );
4520 }
4521 }
4522 out
4523}
4524
4525/// Lift an [`M2MPair`] into a fully-specified [`Operation::CreateM2MTable`].
4526/// The target table's PK column name is resolved from `current` (the
4527/// snapshot the diff is computing toward) — without it the DDL would
4528/// reference a column the child table doesn't have.
4529fn build_create_m2m_op(spec: &M2MPair, current: &Snapshot) -> Result<Operation, MigrateError> {
4530 // Resolve the target's PK from the current snapshot, FALLING BACK to the
4531 // global model registry. Migrations are generated per-plugin, so a
4532 // CROSS-PLUGIN M2M (parent owned by app A, target model owned by app B —
4533 // e.g. a tenant model with an M2M to a SHARED lookup table, or any app's
4534 // M2M to `umbral-auth`'s `User`) has its target in a *different* plugin's
4535 // snapshot, absent from `current`. The global registry sees every
4536 // registered model, so the junction DDL resolves the child PK no matter
4537 // which plugin owns the target. (Cross-plugin FK ordering already lets the
4538 // junction migration run after the target table's own migration.)
4539 let pk_col_and_ty = |m: &ModelMeta| -> (String, crate::orm::SqlType) {
4540 let pk = m.fields.iter().find(|c| c.primary_key);
4541 (
4542 pk.map(|c| c.name.clone())
4543 .unwrap_or_else(|| "id".to_string()),
4544 pk.map(|c| c.ty).unwrap_or(crate::orm::SqlType::BigInt),
4545 )
4546 };
4547 let (child_pk_col, child_ty) = current
4548 .models
4549 .iter()
4550 .find(|m| m.table == spec.target_table)
4551 .map(&pk_col_and_ty)
4552 .or_else(|| {
4553 // Non-panicking global lookup. `registered_models()` panics if the
4554 // registry isn't initialised (unit tests that call `diff` directly,
4555 // with no `App::build`); a `None` registry simply yields no global
4556 // fallback, so a TRULY-unregistered target is still rejected below.
4557 REGISTRY.get().and_then(|reg| {
4558 reg.iter()
4559 .find(|(_, m)| m.table == spec.target_table)
4560 .map(|(_, m)| pk_col_and_ty(m))
4561 })
4562 })
4563 .ok_or_else(|| {
4564 MigrateError::UnsupportedChange(format!(
4565 "M2M `{}.{}` targets table `{}` which is not registered \
4566 anywhere — register the target model via \
4567 `AppBuilder::model::<{}>()` or its owning plugin.",
4568 spec.parent_table, spec.field_name, spec.target_table, spec.target_table,
4569 ))
4570 })?;
4571 let parent_model = current
4572 .models
4573 .iter()
4574 .find(|m| m.table == spec.parent_table)
4575 .expect("parent model exists in snapshot — collect_m2m_pairs iterated it");
4576 let parent_ty = parent_model
4577 .fields
4578 .iter()
4579 .find(|c| c.primary_key)
4580 .map(|c| c.ty)
4581 .unwrap_or(crate::orm::SqlType::BigInt);
4582 Ok(Operation::CreateM2MTable {
4583 junction_table: spec.junction_table.clone(),
4584 parent_table: spec.parent_table.clone(),
4585 parent_col: spec.parent_pk.clone(),
4586 child_table: spec.target_table.clone(),
4587 child_col: child_pk_col,
4588 parent_ty,
4589 child_ty,
4590 })
4591}
4592
4593/// Compute a canonical, sorted column-shape fingerprint for rename
4594/// heuristic detection in `diff`. Two models whose column fingerprints
4595/// are identical are candidates for a rename (second-pass detection).
4596///
4597/// The fingerprint is a sorted `Vec` of `(name, ty, nullable, fk_target)`
4598/// tuples. Sorting by name ensures the fingerprint is independent of
4599/// declaration order.
4600fn column_shape(fields: &[Column]) -> Vec<(String, SqlType, bool, Option<String>)> {
4601 let mut shape: Vec<(String, SqlType, bool, Option<String>)> = fields
4602 .iter()
4603 .map(|c| (c.name.clone(), c.ty, c.nullable, c.fk_target.clone()))
4604 .collect();
4605 shape.sort_by(|a, b| a.0.cmp(&b.0));
4606 shape
4607}
4608
4609/// Type changes the migration engine can apply without user
4610/// intervention. The contract: every entry in this whitelist must be
4611/// data-preserving on both backends.
4612///
4613/// SQLite handles every entry trivially via the table-recreation
4614/// dance: its dynamic typing means whatever lives in a column today
4615/// reads back fine under a new column type affinity. Postgres needs
4616/// `ALTER COLUMN ... TYPE new_type USING column::new_type`, which the
4617/// renderer emits when this returns `true`.
4618///
4619/// What's *not* here is deliberate:
4620/// - `Text -> BigInt` / numeric parses can fail at runtime on non-
4621/// numeric rows. Force the user to write the migration so they own
4622/// the validation.
4623/// - Bigger int -> smaller int truncates silently.
4624/// - `Text -> Date` / `Text -> Uuid` are format-dependent.
4625/// - Anything -> JSON. Even if existing rows are JSON-shaped, that's
4626/// the user's invariant to assert.
4627fn is_safe_cast(from: SqlType, to: SqlType) -> bool {
4628 use SqlType::*;
4629 if from == to {
4630 return true;
4631 }
4632 match (from, to) {
4633 // Stringify: every scalar serialises to text losslessly. Read-
4634 // path code that wants the typed value parses it back; the
4635 // cast itself never fails.
4636 (
4637 SmallInt | Integer | BigInt | Real | Double | Boolean | Date | Time | Timestamptz
4638 | Timestamp | Uuid | Inet | Cidr | MacAddr | ForeignKey,
4639 Text,
4640 ) => true,
4641 // A naive/aware timestamp swap keeps the same wall-clock text, so the
4642 // cast is data-preserving in both directions.
4643 (Timestamp, Timestamptz) | (Timestamptz, Timestamp) => true,
4644 // Integer widening — no data loss.
4645 (SmallInt, Integer | BigInt) => true,
4646 (Integer, BigInt) => true,
4647 // Float widening.
4648 (Real, Double) => true,
4649 // ForeignKey is stored as BigInt under the hood, so the two
4650 // directions are storage-identical. The Rust-side type is
4651 // different but the bytes on disk are not.
4652 (ForeignKey, BigInt) => true,
4653 (BigInt, ForeignKey) => true,
4654 _ => false,
4655 }
4656}
4657
4658/// Postgres type name for an `ALTER COLUMN ... TYPE <name> USING …`
4659/// clause. Matches what sea-query's `PostgresQueryBuilder` emits for
4660/// the same `SqlType` inside a `CREATE TABLE`, so the resulting
4661/// schema after the alter is identical to a freshly created table.
4662/// The full Postgres type for a column, INCLUDING the varchar bound. A
4663/// `Text` column with `max_length > 0` is `character varying(n)` — the same
4664/// mapping the CreateTable path uses — so alters land on the type the table
4665/// would have been created with, not a bare `text` that silently drops the
4666/// bound.
4667fn postgres_type_sql(col: &Column) -> String {
4668 match col.ty {
4669 SqlType::Text if col.max_length > 0 => format!("character varying({})", col.max_length),
4670 // PostGIS spatial types carry their subtype + SRID in the type modifier,
4671 // e.g. `geometry(Point,4326)`. Reuse the same renderer the CreateTable
4672 // path uses so an ALTER lands on the identical column type.
4673 SqlType::Geometry(spec) => crate::backend::pg_spatial_type("geometry", spec),
4674 SqlType::Geography(spec) => crate::backend::pg_spatial_type("geography", spec),
4675 // Caller-chosen numeric dimensions, so an ALTER lands on the same
4676 // `numeric(p,s)` the CreateTable path emits.
4677 SqlType::DecimalN(spec) => format!("numeric({}, {})", spec.precision, spec.scale),
4678 _ => postgres_type_name(col.ty).to_string(),
4679 }
4680}
4681
4682fn postgres_type_name(ty: SqlType) -> &'static str {
4683 use SqlType::*;
4684 match ty {
4685 SmallInt => "smallint",
4686 Integer => "integer",
4687 BigInt | ForeignKey => "bigint",
4688 Real => "real",
4689 Double => "double precision",
4690 Boolean => "boolean",
4691 Text => "text",
4692 Date => "date",
4693 Time => "time",
4694 // sea-query's Postgres builder emits `timestamp with time zone`
4695 // for the equivalent column type; both spellings are accepted
4696 // by Postgres, but mirroring the builder keeps the surface
4697 // consistent if a test ever round-trips DDL.
4698 Timestamptz => "timestamp with time zone",
4699 Timestamp => "timestamp without time zone",
4700 Uuid => "uuid",
4701 Json => "jsonb",
4702 Inet => "inet",
4703 Cidr => "cidr",
4704 MacAddr => "macaddr",
4705 // gaps2 #70: text-backed Postgres types. `bit varying` mirrors
4706 // what sea-query's builder emits for the CREATE TABLE path.
4707 Xml => "xml",
4708 Ltree => "ltree",
4709 Bit => "bit varying",
4710 FullText => "tsvector",
4711 Bytes => "bytea",
4712 // BUG-10: NUMERIC(19, 4) — same dimensions as the CREATE TABLE
4713 // build path. Used by the `ALTER COLUMN ... TYPE ...` render
4714 // when the safe-cast diff allows transitioning to/from
4715 // Decimal.
4716 Decimal => "numeric(19, 4)",
4717 // Arbitrary-precision `numeric` (no dimensions). Mirrors the
4718 // unbounded `ColumnType::Decimal(None)` the create path emits.
4719 BigDecimal => "numeric",
4720 // Bare fallback; the dimensioned `numeric(p,s)` form is rendered by
4721 // `postgres_type_sql`, which special-cases DecimalN before reaching here.
4722 DecimalN(_) => "numeric",
4723 // Bare spatial base type. The subtype+SRID-aware form is rendered by
4724 // `postgres_type_sql`, which special-cases these before reaching here;
4725 // this arm is the unconstrained fallback and keeps the match total.
4726 Geometry(_) => "geometry",
4727 Geography(_) => "geography",
4728 // Arrays render as `<inner>[]` in Postgres. The migration
4729 // engine doesn't model nested element types deeply enough to
4730 // emit a precise inner type here at v1; fall back to `text[]`
4731 // and rely on the column-def renderer for the real shape when
4732 // recreating the column.
4733 Array(_) => "text[]",
4734 }
4735}
4736
4737/// Per-model column diff. Same-name columns whose type or nullable
4738/// flag changed return `UnsafeAlter` (no `AlterColumn` until M8 v1.1
4739/// covers the table-recreation dance for SQLite plus native ALTER for
4740/// Postgres). New-named columns emit `AddColumn`; missing-name columns
4741/// emit `DropColumn`. The ordering is: drops first, then adds, so a
4742/// rename-as-drop+add doesn't violate a uniqueness constraint mid-
4743/// migration on a single-row table.
4744fn diff_columns(
4745 model: &str,
4746 previous: &ModelMeta,
4747 current: &ModelMeta,
4748) -> Result<Vec<Operation>, MigrateError> {
4749 use std::collections::BTreeMap;
4750
4751 let prev_cols: BTreeMap<&str, &Column> = previous
4752 .fields
4753 .iter()
4754 .map(|c| (c.name.as_str(), c))
4755 .collect();
4756 let curr_cols: BTreeMap<&str, &Column> = current
4757 .fields
4758 .iter()
4759 .map(|c| (c.name.as_str(), c))
4760 .collect();
4761
4762 // Walk the intersection by name. Two questions per shared column:
4763 // - did the type change? If so, is the change in the safe-cast
4764 // whitelist (e.g. BigInt -> Text, SmallInt -> Integer)? Safe
4765 // casts emit AlterColumn; unsafe ones still UnsafeAlter so the
4766 // user is forced to write the data-preserving migration by
4767 // hand.
4768 // - did the nullable flag flip? AlterColumn either way.
4769 // Primary-key changes still UnsafeAlter (a PK rebuild is its own
4770 // dance and isn't shipped yet).
4771 let mut alter_columns: Vec<&str> = Vec::new();
4772 // Columns whose `#[umbral(help = "...")]` changed (gaps3 #43).
4773 let mut comment_columns: Vec<&str> = Vec::new();
4774 for (name, prev_col) in &prev_cols {
4775 if let Some(curr_col) = curr_cols.get(name) {
4776 if prev_col.primary_key != curr_col.primary_key {
4777 return Err(MigrateError::UnsafeAlter {
4778 model: model.to_string(),
4779 column: (*name).to_string(),
4780 reason: "primary-key flips need a manual data-preserving migration".to_string(),
4781 });
4782 }
4783 let type_changed = prev_col.ty != curr_col.ty;
4784 if type_changed && !is_safe_cast(prev_col.ty, curr_col.ty) {
4785 return Err(MigrateError::UnsafeAlter {
4786 model: model.to_string(),
4787 column: (*name).to_string(),
4788 reason: format!(
4789 "type change {prev_ty:?} -> {curr_ty:?} is not in the safe-cast whitelist — write a data-preserving migration by hand",
4790 prev_ty = prev_col.ty,
4791 curr_ty = curr_col.ty,
4792 ),
4793 });
4794 }
4795 if prev_col.nullable && !curr_col.nullable && curr_col.default.is_empty() {
4796 return Err(MigrateError::UnsafeAlter {
4797 model: model.to_string(),
4798 column: (*name).to_string(),
4799 reason: "nullable → NOT NULL requires a default/backfill before tightening; otherwise existing NULL rows abort the migration".to_string(),
4800 });
4801 }
4802 if !prev_col.unique && curr_col.unique {
4803 return Err(MigrateError::UnsafeAlter {
4804 model: model.to_string(),
4805 column: (*name).to_string(),
4806 reason: "adding UNIQUE to an existing column requires a duplicate pre-check/backfill migration; otherwise existing duplicate values abort the migration".to_string(),
4807 });
4808 }
4809 // gaps3 #43: `help` DOES have a DB effect now — it renders as a
4810 // Postgres `COMMENT ON COLUMN` — but it is not a column rewrite.
4811 // It gets its own cheap op rather than dragging the column through
4812 // an `AlterColumn` (a full table-recreation dance on SQLite).
4813 if prev_col.help != curr_col.help {
4814 comment_columns.push(*name);
4815 }
4816
4817 // `max_length` IS schema-meaningful for Text columns: CreateTable
4818 // renders `Text + max_length > 0` as `character varying(n)` on
4819 // Postgres, so treating it as UI-only let models drift from DDL —
4820 // raising a bound never emitted an ALTER, and every FRESH database
4821 // enforced the stale one (found live loading TaskFlow's dev dump:
4822 // four columns rejected rows the app had been writing for weeks;
4823 // SQLite hid it because its varchar bounds are advisory).
4824 // WIDENING (n -> bigger, or n -> unbounded) is a lossless in-place
4825 // alter; NARROWING (smaller, or unbounded -> bounded) can abort on
4826 // existing rows mid-migration, so it stays a hand-written job.
4827 let length_changed = !type_changed
4828 && prev_col.ty == SqlType::Text
4829 && prev_col.max_length != curr_col.max_length;
4830 if length_changed
4831 && curr_col.max_length != 0
4832 && (prev_col.max_length == 0 || curr_col.max_length < prev_col.max_length)
4833 {
4834 return Err(MigrateError::UnsafeAlter {
4835 model: model.to_string(),
4836 column: (*name).to_string(),
4837 reason: format!(
4838 "narrowing max_length {prev} -> {curr} can abort on existing oversized rows — write a data-preserving (truncate/validate) migration by hand",
4839 prev = if prev_col.max_length == 0 {
4840 "unbounded".to_string()
4841 } else {
4842 prev_col.max_length.to_string()
4843 },
4844 curr = curr_col.max_length,
4845 ),
4846 });
4847 }
4848
4849 // Any schema-meaningful field change triggers AlterColumn.
4850 // UI-only flags (`noform`, `noedit`,
4851 // `is_string_repr`, `is_multichoice`) are intentionally
4852 // excluded — they affect admin / OpenAPI rendering but
4853 // not the database schema, so emitting an ALTER would do
4854 // no DB work. The single-column `index` flag is ALSO excluded
4855 // here: an index add/remove is not a column rewrite. Folding it
4856 // into `AlterColumn` created no index on Postgres (its native
4857 // ALTER handles TYPE/nullable/UNIQUE/DEFAULT/FK/CHECK but not
4858 // indexes) and forced a full table-recreation dance on SQLite.
4859 // `diff_indexes` now emits a proper `AddIndex`/`DropIndex` for it,
4860 // which is correct and cheap on both backends.
4861 if type_changed
4862 || length_changed
4863 || prev_col.nullable != curr_col.nullable
4864 || prev_col.fk_target != curr_col.fk_target
4865 || prev_col.unique != curr_col.unique
4866 || prev_col.default != curr_col.default
4867 || prev_col.choices != curr_col.choices
4868 || prev_col.choice_labels != curr_col.choice_labels
4869 || prev_col.on_delete != curr_col.on_delete
4870 || prev_col.on_update != curr_col.on_update
4871 {
4872 alter_columns.push(*name);
4873 }
4874 }
4875 }
4876
4877 let mut ops: Vec<Operation> = Vec::new();
4878
4879 // AlterColumn ops first, in name order. One AlterColumn per
4880 // changed column; each carries the full new schema so the render
4881 // can rebuild without further context. Multiple nullable flips on
4882 // one table generate multiple AlterColumns; the apply loop runs
4883 // them sequentially (each is a table-recreation, so back-to-back
4884 // alters drop and recreate twice; the cost is acceptable while
4885 // M5.1 ships the simple case).
4886 //
4887 // audit_2 H21: the SQLite recreation dance rebuilds the table by
4888 // `INSERT INTO tmp (new_columns) SELECT new_columns FROM <old>`, so
4889 // every name in `new_columns` MUST exist in the old table at alter
4890 // time. Using `current.fields` here broke any diff that combined an
4891 // alter with an add or drop on the same table: a newly-ADDED column
4892 // is in `current` but not the old table (the SELECT hits "no such
4893 // column"), and a DROPPED column is absent from `current` so the
4894 // rebuild removed it early — the subsequent `DropColumn` op then
4895 // failed on the already-gone column. Instead, shape `new_columns`
4896 // like the PREVIOUS table (exactly the old table's columns), but
4897 // apply the CURRENT definition to every column that survives so the
4898 // type/nullable/default change still lands. A to-be-dropped column
4899 // keeps its old definition and rides through the rebuild; its
4900 // `DropColumn` op (emitted below, so it runs after) removes it. A
4901 // to-be-added column is intentionally absent; its `AddColumn` op
4902 // (also below) adds it after. Ordering alter → drop → add is what
4903 // the existing op emission already does.
4904 let new_columns: Vec<Column> = previous
4905 .fields
4906 .iter()
4907 .map(|prev_col| {
4908 curr_cols
4909 .get(prev_col.name.as_str())
4910 .map(|c| (*c).clone())
4911 .unwrap_or_else(|| prev_col.clone())
4912 })
4913 .collect();
4914 let prev_columns_snapshot: Vec<Column> = previous.fields.clone();
4915 for name in alter_columns {
4916 ops.push(Operation::AlterColumn {
4917 table: current.table.clone(),
4918 column: name.to_string(),
4919 new_columns: new_columns.clone(),
4920 prev_columns: Some(prev_columns_snapshot.clone()),
4921 // audit_2 core-migrate #10: carry the table-level constraints so the
4922 // SQLite recreation dance re-creates them instead of dropping them.
4923 unique_together: current.unique_together.clone(),
4924 indexes: current.indexes.clone(),
4925 });
4926 }
4927
4928 // gaps3 #43: comment ops, in name order. After any `AlterColumn` on the same
4929 // column, because SQLite's alter recreates the table and Postgres's rewrites
4930 // the column — a comment set first would be describing a column that is
4931 // about to be replaced.
4932 for name in comment_columns {
4933 ops.push(Operation::SetColumnComment {
4934 table: current.table.clone(),
4935 column: name.to_string(),
4936 comment: curr_cols
4937 .get(name)
4938 .map(|c| c.help.clone())
4939 .unwrap_or_default(),
4940 });
4941 }
4942
4943 // Collect the dropped + added column names. We need both lists in
4944 // memory so the rename heuristic can pair them.
4945 let mut dropped: Vec<&Column> = Vec::new();
4946 let mut added: Vec<&Column> = Vec::new();
4947 for (name, prev_col) in &prev_cols {
4948 if !curr_cols.contains_key(name) {
4949 dropped.push(prev_col);
4950 }
4951 }
4952 for col in ¤t.fields {
4953 if !prev_cols.contains_key(col.name.as_str()) {
4954 added.push(col);
4955 }
4956 }
4957
4958 // Gap 88 — column rename detection. When the same diff yields
4959 // exactly one drop and one add whose column shapes (sans name)
4960 // match bit-for-bit, the most likely interpretation is a rename
4961 // rather than a coincidental drop+add of two unrelated columns.
4962 // Emit RenameColumn instead and warn the user so they can
4963 // verify. Anything more ambiguous (multiple drops or adds, or
4964 // mismatched shapes) falls back to the drop+add path so the
4965 // rename is never inferred against the user's actual intent.
4966 //
4967 // The heuristic deliberately stays conservative: some tools ask
4968 // interactively in this case; we don't have
4969 // a prompt at v1, so the conservative auto-pair is the safest
4970 // shape. Users can always override by writing the
4971 // `RenameColumn` op into the migration file by hand.
4972 let mut paired_drop: Option<&str> = None;
4973 let mut paired_add: Option<&str> = None;
4974 if dropped.len() == 1 && added.len() == 1 {
4975 let d = dropped[0];
4976 let a = added[0];
4977 if column_shape_matches(d, a) {
4978 eprintln!(
4979 "umbral makemigrations: column rename detected on `{}`: \
4980 `{}` → `{}` — verify this is a rename and not a coincidental \
4981 shape match; edit the migration file if it's wrong",
4982 current.table, d.name, a.name,
4983 );
4984 ops.push(Operation::RenameColumn {
4985 table: current.table.clone(),
4986 from: d.name.clone(),
4987 to: a.name.clone(),
4988 column: Some(a.clone()),
4989 });
4990 paired_drop = Some(d.name.as_str());
4991 paired_add = Some(a.name.as_str());
4992 }
4993 }
4994
4995 // Drops first so a same-position add can reuse the column slot.
4996 for col in &dropped {
4997 if Some(col.name.as_str()) == paired_drop {
4998 continue;
4999 }
5000 ops.push(Operation::DropColumn {
5001 table: current.table.clone(),
5002 column: col.name.clone(),
5003 });
5004 }
5005
5006 // Then adds, in current declaration order so the schema retains
5007 // the user-written column order even after re-runs.
5008 for col in &added {
5009 if Some(col.name.as_str()) == paired_add {
5010 continue;
5011 }
5012 // Gap 97 — refuse to add a NOT NULL column without a default
5013 // (and without `auto_now_add` / `auto_now`, which fill the
5014 // column server-side at insert). SQLite + Postgres both
5015 // reject the ADD on a non-empty table; we surface the same
5016 // failure at diff time with actionable guidance so the user
5017 // doesn't ship a migration that bricks every deploy.
5018 if !col.nullable
5019 && col.default.is_empty()
5020 && !col.auto_now_add
5021 && !col.auto_now
5022 && !col.primary_key
5023 {
5024 return Err(MigrateError::UnsafeAlter {
5025 model: model.to_string(),
5026 column: col.name.clone(),
5027 reason: format!(
5028 "adding NOT NULL column `{}` without a default to existing \
5029 table `{}` would fail on every populated row. Pick one: \
5030 (a) make the field `Option<T>`, (b) add `#[umbral(default = \
5031 \"...\")]` so the migration backfills, or (c) add \
5032 `#[umbral(auto_now_add)]` for timestamp columns",
5033 col.name, current.table,
5034 ),
5035 });
5036 }
5037 ops.push(Operation::AddColumn {
5038 table: current.table.clone(),
5039 column: (*col).clone(),
5040 });
5041 }
5042
5043 Ok(ops)
5044}
5045
5046/// Gap 88 helper: compare two column snapshots for shape identity (every
5047/// schema-meaningful attribute except `name`). Used by the rename-
5048/// detection heuristic — bit-identical attrs are the signal that a
5049/// dropped column matches an added column and the diff is actually a
5050/// rename. Excludes UI-only flags (`noform`, `noedit`, `max_length`,
5051/// `is_string_repr`, `help`, `example`, `slug_from`) for the same
5052/// reason the AlterColumn diff excludes them: they have no DB effect.
5053fn column_shape_matches(a: &Column, b: &Column) -> bool {
5054 a.ty == b.ty
5055 && a.primary_key == b.primary_key
5056 && a.nullable == b.nullable
5057 && a.fk_target == b.fk_target
5058 && a.choices == b.choices
5059 && a.choice_labels == b.choice_labels
5060 && a.default == b.default
5061 && a.is_multichoice == b.is_multichoice
5062 && a.unique == b.unique
5063 && a.on_delete == b.on_delete
5064 && a.on_update == b.on_update
5065 && a.index == b.index
5066 && a.auto_now_add == b.auto_now_add
5067 && a.auto_now == b.auto_now
5068 && a.min == b.min
5069 && a.max == b.max
5070 && a.text_format == b.text_format
5071}
5072
5073/// Pick the suffix used in a migration filename. Single-op migrations
5074/// get a descriptive suffix; multi-op migrations fall back to `auto`.
5075fn suffix_for(ops: &[Operation]) -> String {
5076 match ops {
5077 [Operation::CreateTable { table, .. }] => format!("create_{table}"),
5078 [Operation::DropTable { table }] => format!("drop_{table}"),
5079 // features #73. The drop+create pair is what an EDIT to a view's SQL looks
5080 // like — by far the commonest view migration — so name it for what the user
5081 // did, not for the two statements it lowers to.
5082 [Operation::CreateView { name, .. }] => format!("create_view_{name}"),
5083 [Operation::DropView { name, .. }] => format!("drop_view_{name}"),
5084 [
5085 Operation::DropView { name: dropped, .. },
5086 Operation::CreateView { name: created, .. },
5087 ] if dropped == created => format!("update_view_{created}"),
5088 [Operation::AddColumn { table, column }] => format!("add_{}_{}", table, column.name),
5089 [Operation::DropColumn { table, column }] => format!("drop_{table}_{column}"),
5090 [Operation::AlterColumn { table, column, .. }] => format!("alter_{table}_{column}"),
5091 [Operation::RenameTable { from, to }] => format!("rename_{from}_to_{to}"),
5092 [
5093 Operation::RenameColumn {
5094 table, from, to, ..
5095 },
5096 ] => format!("rename_{table}_{from}_to_{to}"),
5097 [Operation::SetColumnComment { table, column, .. }] => {
5098 format!("comment_{table}_{column}")
5099 }
5100 [Operation::RunSql { .. }] => "run_sql".to_string(),
5101 [Operation::AddIndex { table, columns, .. }] => {
5102 format!("add_index_{table}_{}", columns.join("_"))
5103 }
5104 [Operation::DropIndex { table, columns, .. }] => {
5105 format!("drop_index_{table}_{}", columns.join("_"))
5106 }
5107 _ => "auto".to_string(),
5108 }
5109}
5110
5111/// Create the tracking table if it isn't there already. The DDL is
5112/// dialect-neutral (TEXT + composite PK is valid SQL on both shipped
5113/// backends), but the executor type isn't — sqlx::query is generic
5114/// over the database, so each backend gets its own thin wrapper.
5115///
5116/// Kept inline because this table is a chicken-and-egg case: every
5117/// other migration needs the tracking row written, so the table
5118/// itself can't be a migration.
5119async fn ensure_tracking_table_sqlite(pool: &sqlx::SqlitePool) -> Result<(), MigrateError> {
5120 sqlx::query(
5121 "CREATE TABLE IF NOT EXISTS umbral_migrations (
5122 plugin TEXT NOT NULL,
5123 name TEXT NOT NULL,
5124 applied_at TEXT NOT NULL,
5125 snapshot_hash TEXT NOT NULL,
5126 PRIMARY KEY (plugin, name)
5127 )",
5128 )
5129 .execute(pool)
5130 .await?;
5131 Ok(())
5132}
5133
5134/// Postgres counterpart to [`ensure_tracking_table_sqlite`].
5135async fn ensure_tracking_table_postgres(pool: &sqlx::PgPool) -> Result<(), MigrateError> {
5136 sqlx::query(
5137 "CREATE TABLE IF NOT EXISTS umbral_migrations (
5138 plugin TEXT NOT NULL,
5139 name TEXT NOT NULL,
5140 applied_at TEXT NOT NULL,
5141 snapshot_hash TEXT NOT NULL,
5142 PRIMARY KEY (plugin, name)
5143 )",
5144 )
5145 .execute(pool)
5146 .await?;
5147 Ok(())
5148}
5149
5150/// Pull the set of `(plugin, name)` tuples already recorded in the
5151/// tracking table (SQLite).
5152async fn applied_names_sqlite(
5153 pool: &sqlx::SqlitePool,
5154) -> Result<std::collections::HashSet<(String, String)>, MigrateError> {
5155 let rows: Vec<(String, String)> = sqlx::query_as("SELECT plugin, name FROM umbral_migrations")
5156 .fetch_all(pool)
5157 .await?;
5158 Ok(rows.into_iter().collect())
5159}
5160
5161/// Postgres counterpart to [`applied_names_sqlite`].
5162async fn applied_names_postgres(
5163 pool: &sqlx::PgPool,
5164) -> Result<std::collections::HashSet<(String, String)>, MigrateError> {
5165 let rows: Vec<(String, String)> = sqlx::query_as("SELECT plugin, name FROM umbral_migrations")
5166 .fetch_all(pool)
5167 .await?;
5168 Ok(rows.into_iter().collect())
5169}
5170
5171/// Render one operation to a list of SQL statements via sea-query.
5172///
5173/// Dispatches on the ambient backend's [`crate::backend::active`]
5174/// name; SQLite and Postgres are the two shipped dialects. Most ops
5175/// produce one statement; `AlterColumn` produces either the SQLite
5176/// table-recreation dance (`CREATE _umbral_new` + `INSERT ... SELECT`
5177/// + `DROP` + `RENAME`) or a single native `ALTER TABLE ... ALTER
5178/// COLUMN ... SET/DROP NOT NULL` on Postgres.
5179///
5180/// The apply loop in `run_in` executes each statement in order inside
5181/// the same transaction.
5182///
5183/// `AddColumn` ignores the `primary_key` flag: neither SQLite nor
5184/// Postgres lets a primary key be added to an existing table without
5185/// a table-recreation step, and the autodetector won't route a
5186/// pk-flagged column through `AddColumn` anyway. A hand-edited
5187/// migration that sets the flag is taken to mean "the user is taking
5188/// responsibility".
5189fn render_operation(op: &Operation) -> Vec<String> {
5190 render_operation_for(op, crate::backend::active().name())
5191}
5192
5193fn should_emit_btree_index(col: &Column) -> bool {
5194 !col.primary_key
5195 && !col.unique
5196 && (col.index || matches!(col.ty, SqlType::ForeignKey) || col.name == "deleted_at")
5197}
5198
5199/// A PostGIS spatial column — indexed with GiST rather than B-tree.
5200fn is_spatial(col: &Column) -> bool {
5201 matches!(col.ty, SqlType::Geometry(_) | SqlType::Geography(_))
5202}
5203
5204/// Render one operation against an explicit backend name. The
5205/// dispatching seam — the public [`render_operation`] is just
5206/// `render_operation_for(op, backend::active().name())`. Splitting
5207/// the two lets tests render Postgres DDL without installing the
5208/// process-wide ambient backend (the `OnceLock` can only be set once,
5209/// so `App::build` and tests would otherwise collide).
5210///
5211/// Panics on unknown backend names; only `"sqlite"` and `"postgres"`
5212/// are shipped in Phase 2.
5213pub fn render_operation_for(op: &Operation, backend_name: &str) -> Vec<String> {
5214 match backend_name {
5215 "sqlite" => render_operation_sqlite(op),
5216 "postgres" => render_operation_postgres(op),
5217 other => panic!(
5218 "umbral::migrate: no DDL renderer for backend `{other}`; \
5219 Phase 2 ships sqlite and postgres only"
5220 ),
5221 }
5222}
5223
5224/// SQLite-dialect rendering for one operation.
5225fn render_operation_sqlite(op: &Operation) -> Vec<String> {
5226 use sea_query::{Alias, SqliteQueryBuilder, Table};
5227
5228 match op {
5229 // features #73. A view is DDL the ORM cannot express as a sea-query
5230 // statement (there is no ViewCreateStatement), and the SELECT body is an
5231 // opaque string the user wrote. This is exception (1) in CLAUDE.md's raw-SQL
5232 // rule: schema DDL, owned by the migration engine.
5233 //
5234 // A MATERIALIZED view never reaches here — the `model.materialized_view`
5235 // system check fails the boot on SQLite. Rendering it as a plain view would
5236 // be the exact "SQLite branch that quietly diverges" the design principles
5237 // forbid: the same query, silently recomputed on every read.
5238 Operation::CreateView {
5239 name,
5240 sql,
5241 materialized,
5242 } => {
5243 debug_assert!(
5244 !materialized,
5245 "materialized views are Postgres-only; the model.materialized_view \
5246 system check should have failed the boot before reaching the renderer"
5247 );
5248 vec![format!("CREATE VIEW \"{name}\" AS {sql}")]
5249 }
5250 Operation::DropView { name, .. } => vec![format!("DROP VIEW IF EXISTS \"{name}\"")],
5251 Operation::CreateTable {
5252 table,
5253 columns,
5254 unique_together,
5255 indexes,
5256 } => {
5257 // sea-query's TableCreateStatement renders columns inline.
5258 let mut stmt = Table::create();
5259 stmt.table(Alias::new(table));
5260 for col in columns {
5261 let mut def = build_column_def_sqlite(col);
5262 stmt.col(&mut def);
5263 }
5264 let mut stmts = vec![stmt.build(SqliteQueryBuilder)];
5265 // `unique_together` groups render as follow-up
5266 // `CREATE UNIQUE INDEX` statements rather than inline table
5267 // constraints. A named unique index enforces the SAME
5268 // constraint, but — unlike an inline `UNIQUE(...)` (which becomes
5269 // an un-droppable auto-index on SQLite / an implicit constraint
5270 // on Postgres) — it can be added and DROPPED by name after the
5271 // table exists. That is what makes `unique_together`
5272 // autodetection reversible: the exact statement an `AddIndex`
5273 // would emit, so a later `DropIndex` names the same index.
5274 for group in unique_together {
5275 stmts.push(add_index_stmt(table, group, true));
5276 }
5277 // Single-column explicit indexes plus ORM-required helper
5278 // indexes follow the CREATE TABLE. FK columns need indexes
5279 // for reverse/select-related queries, and soft-delete
5280 // models read through `deleted_at IS NULL` by default.
5281 for col in columns {
5282 if should_emit_btree_index(col) {
5283 stmts.push(create_index_stmt(table, &col.name));
5284 }
5285 }
5286 // BUG-7: multi-column indexes follow as plain CREATE INDEX.
5287 for group in indexes {
5288 stmts.push(create_multi_index_stmt(table, group));
5289 }
5290 stmts
5291 }
5292 // gaps3 #60 — `IF EXISTS`, matching `DROP INDEX` and `DROP VIEW`, which have always
5293 // said it. A migration's job is to make a statement about the DESIRED end state:
5294 // "this table should not exist". Erroring because the table is already gone is a
5295 // complaint that the world already agrees with you.
5296 //
5297 // It is not masking a bug. Forward application has to be idempotent, or a ledger
5298 // that drifts even slightly — a table dropped out of band, a rename replayed, an
5299 // app whose migration history was rewritten — becomes permanently unmigratable,
5300 // and the only escape the operator can see is deleting the database.
5301 Operation::DropTable { table } => vec![
5302 Table::drop()
5303 .table(Alias::new(table))
5304 .if_exists()
5305 .build(SqliteQueryBuilder),
5306 ],
5307 Operation::AddColumn { table, column } => {
5308 // SQLite-specific limitation: `ALTER TABLE ADD COLUMN`
5309 // requires a CONSTANT default. `CURRENT_TIMESTAMP` is
5310 // non-constant ("Cannot add a column with non-constant
5311 // default"). So when we're adding a NOT NULL auto_now /
5312 // auto_now_add column on top of an existing table, we
5313 // emit a two-statement sequence:
5314 // 1. ADD COLUMN as NULLABLE (no default needed).
5315 // 2. UPDATE every existing row to `datetime('now')`.
5316 // The column ends up NULL-permitting at the DB level on
5317 // SQLite — but the Rust type stays `DateTime<Utc>` (not
5318 // Option), and every INSERT through the ORM supplies a
5319 // value via the macro-emitted auto_now arm. The DB-side
5320 // NOT NULL guarantee is lost only for direct-SQL writers,
5321 // which umbral already discourages (see CLAUDE.md "Plugins
5322 // use the ORM"). Postgres has no such restriction —
5323 // `DEFAULT now()` works there in ALTER, no backfill
5324 // statement needed (see the Postgres render below).
5325 let needs_backfill = (column.auto_now || column.auto_now_add)
5326 && !column.nullable
5327 && matches!(
5328 column.ty,
5329 SqlType::Timestamptz | SqlType::Timestamp | SqlType::Date | SqlType::Time
5330 );
5331
5332 let mut stmts = if needs_backfill {
5333 let mut nullable_col = column.clone();
5334 nullable_col.nullable = true;
5335 let mut stmt = Table::alter();
5336 stmt.table(Alias::new(table));
5337 let mut def = build_column_def_sqlite(&nullable_col);
5338 stmt.add_column(&mut def);
5339 let add_sql = stmt.build(SqliteQueryBuilder);
5340
5341 // Manual UPDATE — sea-query's update builder is
5342 // overkill for a single SET col = datetime('now').
5343 let table_quoted = table.replace('"', "\"\"");
5344 let col_quoted = column.name.replace('"', "\"\"");
5345 let backfill_sql = format!(
5346 "UPDATE \"{table_quoted}\" SET \"{col_quoted}\" = datetime('now') \
5347 WHERE \"{col_quoted}\" IS NULL"
5348 );
5349 vec![add_sql, backfill_sql]
5350 } else {
5351 let mut stmt = Table::alter();
5352 stmt.table(Alias::new(table));
5353 // SQLite rejects `ADD COLUMN … UNIQUE`. Build the column def
5354 // WITHOUT the inline UNIQUE and apply uniqueness as a
5355 // separate `CREATE UNIQUE INDEX` below — semantically
5356 // identical, and it applies on a populated table.
5357 let mut col_no_unique = column.clone();
5358 col_no_unique.unique = false;
5359 let mut def = build_column_def_sqlite(&col_no_unique);
5360 stmt.add_column(&mut def);
5361 vec![stmt.build(SqliteQueryBuilder)]
5362 };
5363 if column.unique && !column.primary_key {
5364 stmts.push(create_unique_index_stmt(table, &column.name));
5365 }
5366 if should_emit_btree_index(column) {
5367 stmts.push(create_index_stmt(table, &column.name));
5368 }
5369 stmts
5370 }
5371 Operation::DropColumn { table, column } => vec![
5372 Table::alter()
5373 .table(Alias::new(table))
5374 .drop_column(Alias::new(column))
5375 .build(SqliteQueryBuilder),
5376 ],
5377 // gaps3 #43: SQLite has no `COMMENT` statement — no column comments, no
5378 // table comments, nothing. Rendering zero statements is the whole
5379 // implementation. This is not a silent divergence of the kind the raw-SQL
5380 // rule warns about: the columns, types, constraints and rows are
5381 // identical on both backends. Only the `psql \d+` annotation is absent,
5382 // because SQLite has nowhere to put it.
5383 Operation::SetColumnComment { .. } => Vec::new(),
5384 Operation::AlterColumn {
5385 table,
5386 column,
5387 new_columns,
5388 prev_columns,
5389 unique_together,
5390 indexes,
5391 } => {
5392 // gaps3 #24: choices are Rust-enforced on SQLite (build_column_def_sqlite
5393 // emits no CHECK), so an alter that ONLY changes this column's
5394 // choices/labels is invisible to the schema. Skip the otherwise
5395 // byte-identical table-recreation dance. Postgres still swaps its
5396 // CHECK constraint via the Postgres renderer, so the op is still
5397 // emitted and recorded — only the SQLite render short-circuits.
5398 if prev_columns
5399 .as_ref()
5400 .is_some_and(|prev| alter_is_choices_only(column, prev, new_columns))
5401 {
5402 Vec::new()
5403 } else {
5404 render_alter_column_dance_sqlite(table, new_columns, unique_together, indexes)
5405 }
5406 }
5407 Operation::CreateM2MTable {
5408 junction_table,
5409 parent_table,
5410 parent_col,
5411 child_table,
5412 child_col,
5413 parent_ty,
5414 child_ty,
5415 } => {
5416 // Junction table for many-to-many: two FK columns + composite PK.
5417 // Column types follow the referenced PKs — `BigInt` → `INTEGER`
5418 // (SQLite affinity), `Text` → `TEXT`, `Uuid` → `TEXT` on SQLite
5419 // / `UUID` on Postgres. Raw DDL is the simplest expression of
5420 // the composite-PK + per-side cascade FK shape; sea-query's
5421 // builder can't express it cleanly in one call.
5422 vec![format!(
5423 r#"CREATE TABLE "{jt}" (
5424 "parent_id" {pty} NOT NULL REFERENCES "{pt}"("{pc}") ON DELETE CASCADE,
5425 "child_id" {cty} NOT NULL REFERENCES "{ct}"("{cc}") ON DELETE CASCADE,
5426 PRIMARY KEY ("parent_id", "child_id")
5427)"#,
5428 jt = junction_table.replace('"', "\"\""),
5429 pt = parent_table.replace('"', "\"\""),
5430 pc = parent_col.replace('"', "\"\""),
5431 ct = child_table.replace('"', "\"\""),
5432 cc = child_col.replace('"', "\"\""),
5433 pty = m2m_pk_sql_type_sqlite(*parent_ty),
5434 cty = m2m_pk_sql_type_sqlite(*child_ty),
5435 )]
5436 }
5437 // gaps3 #60 — idempotent, same reasoning as `DropTable`.
5438 Operation::DropM2MTable { junction_table } => vec![
5439 Table::drop()
5440 .table(Alias::new(junction_table))
5441 .if_exists()
5442 .build(SqliteQueryBuilder),
5443 ],
5444 Operation::RenameTable { from, to } => {
5445 use sea_query::{Alias, SqliteQueryBuilder, Table};
5446 vec![
5447 Table::rename()
5448 .table(Alias::new(from.as_str()), Alias::new(to.as_str()))
5449 .build(SqliteQueryBuilder),
5450 ]
5451 }
5452 Operation::RenameColumn {
5453 table, from, to, ..
5454 } => {
5455 // SQLite 3.25+ supports `ALTER TABLE ... RENAME COLUMN`
5456 // natively. Quote both sides to allow names that need
5457 // escaping; sea-query's column-rename builder isn't
5458 // exposed cleanly so we render the DDL string directly.
5459 let t = table.replace('"', "\"\"");
5460 let f = from.replace('"', "\"\"");
5461 let tn = to.replace('"', "\"\"");
5462 vec![format!(
5463 "ALTER TABLE \"{t}\" RENAME COLUMN \"{f}\" TO \"{tn}\""
5464 )]
5465 }
5466 // A data migration renders to its raw forward SQL verbatim —
5467 // the author owns portability across backends.
5468 Operation::RunSql { sql, .. } => vec![sql.clone()],
5469 // Composite index / UNIQUE constraint add + drop. The
5470 // `CREATE [UNIQUE] INDEX IF NOT EXISTS` / `DROP INDEX IF EXISTS`
5471 // forms are identical on SQLite and Postgres, so both render arms
5472 // share the same helpers.
5473 Operation::AddIndex {
5474 table,
5475 columns,
5476 unique,
5477 } => vec![add_index_stmt(table, columns, *unique)],
5478 Operation::DropIndex {
5479 table,
5480 columns,
5481 unique,
5482 } => vec![drop_index_stmt(&index_name(table, columns, *unique))],
5483 }
5484}
5485
5486/// Postgres-dialect rendering for one operation.
5487///
5488/// Postgres has native `ALTER COLUMN` so `AlterColumn` doesn't need
5489/// the SQLite table-recreation dance; it lowers to a single statement.
5490/// Integer primary keys use sea-query's `auto_increment()` flag, which
5491/// the Postgres query builder lowers to `BIGSERIAL` / `SERIAL` rather
5492/// than SQLite's `INTEGER PRIMARY KEY AUTOINCREMENT` quirk.
5493fn render_operation_postgres(op: &Operation) -> Vec<String> {
5494 use sea_query::{Alias, PostgresQueryBuilder, Table};
5495
5496 match op {
5497 // features #73 — see the SQLite renderer for why this is raw DDL.
5498 Operation::CreateView {
5499 name,
5500 sql,
5501 materialized,
5502 } => {
5503 let kind = if *materialized {
5504 "MATERIALIZED VIEW"
5505 } else {
5506 "VIEW"
5507 };
5508 vec![format!("CREATE {kind} \"{name}\" AS {sql}")]
5509 }
5510 Operation::DropView { name, materialized } => {
5511 let kind = if *materialized {
5512 "MATERIALIZED VIEW"
5513 } else {
5514 "VIEW"
5515 };
5516 // No CASCADE. A view another view depends on should fail loudly rather
5517 // than silently take its dependants down with it — the engine emits the
5518 // dependants' drops itself, in order, when it knows about them.
5519 vec![format!("DROP {kind} IF EXISTS \"{name}\"")]
5520 }
5521 Operation::CreateTable {
5522 table,
5523 columns,
5524 unique_together,
5525 indexes,
5526 } => {
5527 let mut stmt = Table::create();
5528 stmt.table(Alias::new(table));
5529 for col in columns {
5530 let mut def = build_column_def_postgres(col);
5531 stmt.col(&mut def);
5532 }
5533 let mut stmts = vec![stmt.build(PostgresQueryBuilder)];
5534 // gaps3 #35: a `#[umbral(case_insensitive)]` text column renders as
5535 // `citext`, which needs its extension. Emit the (idempotent) create
5536 // BEFORE the table so the citext type resolves. Requires a role with
5537 // CREATE privilege on the database — the operator pre-creates it once
5538 // if the runtime role is restricted.
5539 if columns
5540 .iter()
5541 .any(|c| c.case_insensitive && matches!(c.ty, crate::orm::SqlType::Text))
5542 {
5543 stmts.insert(0, "CREATE EXTENSION IF NOT EXISTS citext".to_string());
5544 }
5545 // gaps5 #22: a spatial column needs the PostGIS extension. Emit the
5546 // (idempotent) create BEFORE the table so the `geometry`/`geography`
5547 // types resolve — the same pattern as citext. A restricted runtime
5548 // role has the operator pre-install it; `IF NOT EXISTS` no-ops then.
5549 if columns.iter().any(|c| {
5550 matches!(
5551 c.ty,
5552 crate::orm::SqlType::Geometry(_) | crate::orm::SqlType::Geography(_)
5553 )
5554 }) {
5555 stmts.insert(0, "CREATE EXTENSION IF NOT EXISTS postgis".to_string());
5556 }
5557 // `unique_together` as follow-up `CREATE UNIQUE INDEX` (not an
5558 // inline constraint) so it is droppable by name — see the SQLite
5559 // render arm for the full rationale.
5560 for group in unique_together {
5561 stmts.push(add_index_stmt(table, group, true));
5562 }
5563 for col in columns {
5564 if matches!(col.ty, crate::orm::SqlType::FullText) {
5565 // tsvector columns get an auto-GIN index (#33) — they're
5566 // useless for search without one, so the engine never
5567 // makes the caller hand-write it.
5568 stmts.push(create_gin_index_stmt(table, &col.name));
5569 } else if is_spatial(col) && col.index {
5570 // A spatial column marked `#[umbral(index)]` gets a GiST
5571 // index — the only index type spatial predicates use. A
5572 // plain B-tree is meaningless on geometry (gaps5 #22).
5573 stmts.push(create_gist_index_stmt(table, &col.name));
5574 } else if should_emit_btree_index(col) {
5575 stmts.push(create_index_stmt(table, &col.name));
5576 }
5577 }
5578 for group in indexes {
5579 stmts.push(create_multi_index_stmt(table, group));
5580 }
5581 // gaps3 #43: `#[umbral(help = "...")]` as a column comment. Postgres
5582 // has no inline column-comment syntax, so these follow the CREATE —
5583 // commenting a column that doesn't exist yet is an error.
5584 for col in columns {
5585 if !col.help.is_empty() {
5586 stmts.push(comment_on_column_stmt(table, &col.name, &col.help));
5587 }
5588 }
5589 stmts
5590 }
5591 // gaps3 #60 — see the SQLite renderer.
5592 Operation::DropTable { table } => vec![
5593 Table::drop()
5594 .table(Alias::new(table))
5595 .if_exists()
5596 .build(PostgresQueryBuilder),
5597 ],
5598 Operation::AddColumn { table, column } => {
5599 let mut stmt = Table::alter();
5600 stmt.table(Alias::new(table));
5601 let mut def = build_column_def_postgres(column);
5602 stmt.add_column(&mut def);
5603 let mut stmts = vec![stmt.build(PostgresQueryBuilder)];
5604 if matches!(column.ty, crate::orm::SqlType::FullText) {
5605 // Auto-GIN for a tsvector column added later (#33).
5606 stmts.push(create_gin_index_stmt(table, &column.name));
5607 } else if is_spatial(column) && column.index {
5608 // GiST for a spatial column added later (gaps5 #22).
5609 stmts.push(create_gist_index_stmt(table, &column.name));
5610 } else if should_emit_btree_index(column) {
5611 stmts.push(create_index_stmt(table, &column.name));
5612 }
5613 if !column.help.is_empty() {
5614 stmts.push(comment_on_column_stmt(table, &column.name, &column.help));
5615 }
5616 stmts
5617 }
5618 Operation::SetColumnComment {
5619 table,
5620 column,
5621 comment,
5622 } => vec![comment_on_column_stmt(table, column, comment)],
5623 Operation::DropColumn { table, column } => vec![
5624 Table::alter()
5625 .table(Alias::new(table))
5626 .drop_column(Alias::new(column))
5627 .build(PostgresQueryBuilder),
5628 ],
5629 Operation::AlterColumn {
5630 table,
5631 column,
5632 new_columns,
5633 prev_columns,
5634 // Postgres alters in place — indexes/UNIQUE survive the ALTER, so
5635 // it doesn't re-create them (only the SQLite recreation dance does).
5636 unique_together: _,
5637 indexes: _,
5638 } => render_alter_column_postgres(table, column, new_columns, prev_columns.as_deref()),
5639 Operation::CreateM2MTable {
5640 junction_table,
5641 parent_table,
5642 parent_col,
5643 child_table,
5644 child_col,
5645 parent_ty,
5646 child_ty,
5647 } => {
5648 vec![format!(
5649 r#"CREATE TABLE "{jt}" (
5650 "parent_id" {pty} NOT NULL REFERENCES "{pt}"("{pc}") ON DELETE CASCADE DEFERRABLE INITIALLY IMMEDIATE,
5651 "child_id" {cty} NOT NULL REFERENCES "{ct}"("{cc}") ON DELETE CASCADE DEFERRABLE INITIALLY IMMEDIATE,
5652 PRIMARY KEY ("parent_id", "child_id")
5653)"#,
5654 jt = junction_table.replace('"', "\"\""),
5655 pt = parent_table.replace('"', "\"\""),
5656 pc = parent_col.replace('"', "\"\""),
5657 ct = child_table.replace('"', "\"\""),
5658 cc = child_col.replace('"', "\"\""),
5659 pty = m2m_pk_sql_type_postgres(*parent_ty),
5660 cty = m2m_pk_sql_type_postgres(*child_ty),
5661 )]
5662 }
5663 // gaps3 #60 — idempotent, same reasoning as `DropTable`.
5664 Operation::DropM2MTable { junction_table } => vec![
5665 Table::drop()
5666 .table(Alias::new(junction_table))
5667 .if_exists()
5668 .build(PostgresQueryBuilder),
5669 ],
5670 Operation::RenameTable { from, to } => {
5671 // Postgres: ALTER TABLE "<from>" RENAME TO "<to>"
5672 // sea-query's Table::rename() emits the right form.
5673 use sea_query::{Alias, PostgresQueryBuilder, Table};
5674 vec![
5675 Table::rename()
5676 .table(Alias::new(from.as_str()), Alias::new(to.as_str()))
5677 .build(PostgresQueryBuilder),
5678 ]
5679 }
5680 Operation::RenameColumn {
5681 table, from, to, ..
5682 } => {
5683 let t = table.replace('"', "\"\"");
5684 let f = from.replace('"', "\"\"");
5685 let tn = to.replace('"', "\"\"");
5686 vec![format!(
5687 "ALTER TABLE \"{t}\" RENAME COLUMN \"{f}\" TO \"{tn}\""
5688 )]
5689 }
5690 // A data migration renders to its raw forward SQL verbatim —
5691 // the author owns portability across backends.
5692 Operation::RunSql { sql, .. } => vec![sql.clone()],
5693 // Same `CREATE [UNIQUE] INDEX IF NOT EXISTS` / `DROP INDEX IF EXISTS`
5694 // as SQLite — Postgres accepts the identical form.
5695 Operation::AddIndex {
5696 table,
5697 columns,
5698 unique,
5699 } => vec![add_index_stmt(table, columns, *unique)],
5700 Operation::DropIndex {
5701 table,
5702 columns,
5703 unique,
5704 } => vec![drop_index_stmt(&index_name(table, columns, *unique))],
5705 }
5706}
5707
5708/// The SQLite table-recreation dance for `AlterColumn`. SQLite has no
5709/// in-place `ALTER COLUMN`, so the only safe way to flip a column's
5710/// nullable flag is to rebuild the table:
5711///
5712/// 1. `CREATE TABLE _umbral_new_<table>` with the new schema.
5713/// 2. `INSERT ... SELECT` to copy every row from the old table.
5714/// 3. `DROP TABLE <table>`.
5715/// 4. `ALTER TABLE _umbral_new_<table> RENAME TO <table>`.
5716///
5717/// Wrapped in a transaction by the caller, which — when the migration contains
5718/// an `AlterColumn` — brackets that transaction with `PRAGMA foreign_keys=OFF`
5719/// … `PRAGMA foreign_key_check` … `PRAGMA foreign_keys=ON` (SQLite's official
5720/// recipe), so step 3's `DROP TABLE` on a table with **inbound** FKs doesn't
5721/// trip `FOREIGN KEY constraint failed` (gaps3 #13). Indexes, triggers, and FK
5722/// targets aren't preserved at M5.1 because umbral-core's schema model
5723/// doesn't yet carry them; once it does, this routine picks them up
5724/// by rebuilding them at step 1.
5725///
5726/// Nullable `TRUE -> FALSE` fails at step 2 if any row holds NULL,
5727/// which is the correct data-integrity behaviour. Nullable
5728/// `FALSE -> TRUE` always succeeds.
5729/// gaps3 #24: true when an `AlterColumn` on `column` changes ONLY the closed-set
5730/// `choices` / `choice_labels` — every schema-relevant field (type, nullable,
5731/// unique, default, FK, on_delete/update, pk, length, multichoice) is identical.
5732/// On SQLite such a delta is a no-op: choices are enforced in Rust there, not by
5733/// a DB CHECK, so rebuilding the table would produce a byte-identical schema.
5734fn alter_is_choices_only(column: &str, prev: &[Column], new: &[Column]) -> bool {
5735 let (Some(p), Some(n)) = (
5736 prev.iter().find(|c| c.name == column),
5737 new.iter().find(|c| c.name == column),
5738 ) else {
5739 return false;
5740 };
5741 let choices_changed = p.choices != n.choices || p.choice_labels != n.choice_labels;
5742 let schema_same = p.ty == n.ty
5743 && p.nullable == n.nullable
5744 && p.unique == n.unique
5745 && p.default == n.default
5746 && p.fk_target == n.fk_target
5747 && p.on_delete == n.on_delete
5748 && p.on_update == n.on_update
5749 && p.primary_key == n.primary_key
5750 && p.max_length == n.max_length
5751 && p.is_multichoice == n.is_multichoice;
5752 choices_changed && schema_same
5753}
5754
5755fn render_alter_column_dance_sqlite(
5756 table: &str,
5757 new_columns: &[Column],
5758 unique_together: &[Vec<String>],
5759 indexes: &[Vec<String>],
5760) -> Vec<String> {
5761 use sea_query::{Alias, SqliteQueryBuilder, Table};
5762
5763 let tmp = format!("_umbral_new_{table}");
5764
5765 // Step 1 — CREATE TABLE _umbral_new_<table>.
5766 let mut create = Table::create();
5767 create.table(Alias::new(&tmp));
5768 for col in new_columns {
5769 let mut def = build_column_def_sqlite(col);
5770 create.col(&mut def);
5771 }
5772
5773 // Step 2 — INSERT ... SELECT. The INSERT target list is the plain column
5774 // names; each is double-quoted so SQLite identifier rules don't bite on
5775 // reserved words. The SELECT side backfills any NOT-NULL-with-default column
5776 // via `COALESCE(col, <default>)` (audit_2 core-migrate #5) — a
5777 // nullable→NOT NULL tightening whose existing rows hold NULL would otherwise
5778 // copy NULL into the new NOT NULL column and abort the rebuild. COALESCE is
5779 // a harmless no-op for a column that never held NULLs.
5780 let insert_cols = new_columns
5781 .iter()
5782 .map(|c| format!("\"{}\"", c.name.replace('"', "\"\"")))
5783 .collect::<Vec<_>>()
5784 .join(", ");
5785 let select_exprs = new_columns
5786 .iter()
5787 .map(|c| {
5788 let name = format!("\"{}\"", c.name.replace('"', "\"\""));
5789 if !c.nullable && !c.default.is_empty() {
5790 format!("COALESCE({name}, {})", default_sql_literal(c, false))
5791 } else {
5792 name
5793 }
5794 })
5795 .collect::<Vec<_>>()
5796 .join(", ");
5797 let insert_sql =
5798 format!("INSERT INTO \"{tmp}\" ({insert_cols}) SELECT {select_exprs} FROM \"{table}\"");
5799
5800 // Step 3 — DROP TABLE <table>.
5801 let drop_sql = Table::drop()
5802 .table(Alias::new(table))
5803 .build(SqliteQueryBuilder);
5804
5805 // Step 4 — ALTER TABLE _umbral_new_<table> RENAME TO <table>.
5806 let rename_sql = Table::rename()
5807 .table(Alias::new(&tmp), Alias::new(table))
5808 .build(SqliteQueryBuilder);
5809
5810 let mut stmts = vec![
5811 create.build(SqliteQueryBuilder),
5812 insert_sql,
5813 drop_sql,
5814 rename_sql,
5815 ];
5816 // Step 5 — audit_2 core-migrate #10: re-create the secondary indexes and
5817 // composite UNIQUE constraints the dropped table carried, or the rebuild
5818 // silently drops them (duplicates become insertable — integrity loss).
5819 // Single-column / FK / soft-delete indexes are derived from the columns
5820 // (same rule as CreateTable); `unique_together` re-emits as a named
5821 // `CREATE UNIQUE INDEX` and composite `indexes` as plain `CREATE INDEX`.
5822 // All are `IF NOT EXISTS`, so the step is idempotent.
5823 for col in new_columns {
5824 if should_emit_btree_index(col) {
5825 stmts.push(create_index_stmt(table, &col.name));
5826 }
5827 }
5828 for group in unique_together {
5829 stmts.push(add_index_stmt(table, group, true));
5830 }
5831 for group in indexes {
5832 stmts.push(create_multi_index_stmt(table, group));
5833 }
5834 stmts
5835}
5836
5837/// Native Postgres `AlterColumn`. Postgres supports
5838/// `ALTER TABLE x ALTER COLUMN y SET NOT NULL` and
5839/// `ALTER TABLE x ALTER COLUMN y DROP NOT NULL` in place, so the
5840/// SQLite table-recreation dance isn't needed. Lowers to a single
5841/// statement.
5842///
5843/// `SET NOT NULL` fails at the server if any row holds NULL on `y`,
5844/// matching SQLite's INSERT-time failure on the dance — the
5845/// data-integrity contract is identical between backends.
5846///
5847/// `column` is the field name that triggered the flip; `new_columns`
5848/// is the post-change schema (carried for parity with the SQLite
5849/// dance, though Postgres only needs the one column).
5850fn render_alter_column_postgres(
5851 table: &str,
5852 column: &str,
5853 new_columns: &[Column],
5854 prev_columns: Option<&[Column]>,
5855) -> Vec<String> {
5856 let new = new_columns.iter().find(|c| c.name == column).expect(
5857 "umbral::migrate: AlterColumn op references a column missing from new_columns; \
5858 this is a bug in `diff_columns`",
5859 );
5860 let prev = prev_columns.and_then(|cols| cols.iter().find(|c| c.name == column));
5861
5862 let q_table = quote_pg_ident(table);
5863 let q_column = quote_pg_ident(column);
5864
5865 let mut stmts: Vec<String> = Vec::new();
5866
5867 // TYPE change: only when we have a previous snapshot AND it differs
5868 // AND the change is in the safe-cast whitelist (diff_columns has
5869 // already gated unsafe ones). Emitted before nullable so a NOT
5870 // NULL flip against the just-cast column reads the new type.
5871 //
5872 // A same-type Text column whose `max_length` changed is ALSO a type
5873 // change on Postgres — varchar(n) carries the bound in the type — and
5874 // rendering nothing here is how widened models silently kept their old
5875 // bound on every fresh database (diff_columns has already refused the
5876 // narrowing direction). `postgres_type_sql` is length-aware, so a plain
5877 // safe-cast into a bounded Text lands on varchar(n) too, not bare text.
5878 if let Some(prev_col) = prev {
5879 let cast_changed = prev_col.ty != new.ty && is_safe_cast(prev_col.ty, new.ty);
5880 let length_changed = prev_col.ty == new.ty
5881 && new.ty == SqlType::Text
5882 && prev_col.max_length != new.max_length;
5883 if cast_changed || length_changed {
5884 let new_ty_sql = postgres_type_sql(new);
5885 stmts.push(format!(
5886 "ALTER TABLE {q_table} ALTER COLUMN {q_column} TYPE {new_ty_sql} USING {q_column}::{new_ty_sql}"
5887 ));
5888 }
5889 }
5890
5891 // NULL-flag change: skipped when prev is None (legacy migrations
5892 // with no snapshot — preserve the old "emit unconditionally" path
5893 // because it's idempotent on Postgres). With a snapshot, only emit
5894 // when the flag actually flipped.
5895 let nullable_changed = match prev {
5896 Some(prev_col) => prev_col.nullable != new.nullable,
5897 None => true,
5898 };
5899 if nullable_changed {
5900 // audit_2 core-migrate #5: backfill existing NULLs before tightening.
5901 // A nullable→NOT NULL flip whose column carries a default would abort on
5902 // any pre-existing NULL row (bare `SET NOT NULL` doesn't backfill, and
5903 // `SET DEFAULT` only affects future inserts). Emit the backfill UPDATE
5904 // first so the subsequent `SET NOT NULL` succeeds.
5905 if !new.nullable && !new.default.is_empty() {
5906 let lit = default_sql_literal(new, true);
5907 stmts.push(format!(
5908 "UPDATE {q_table} SET {q_column} = {lit} WHERE {q_column} IS NULL"
5909 ));
5910 }
5911 let clause = if new.nullable {
5912 "DROP NOT NULL"
5913 } else {
5914 "SET NOT NULL"
5915 };
5916 stmts.push(format!(
5917 "ALTER TABLE {q_table} ALTER COLUMN {q_column} {clause}"
5918 ));
5919 }
5920
5921 // From here down — all the gap #65 follow-up changes. Each branch
5922 // checks if `prev` exists (legacy migrations with no snapshot
5923 // skip these, matching the historical behaviour) and emits the
5924 // matching ALTER on real flips.
5925 if let Some(prev_col) = prev {
5926 // UNIQUE flag flip. Postgres autogen for column-level UNIQUE
5927 // at CREATE TABLE is `<table>_<col>_key`; we use the same
5928 // name when ADDing so a subsequent DROP finds it.
5929 if prev_col.unique != new.unique {
5930 let cname = format!("{table}_{column}_key");
5931 if new.unique {
5932 stmts.push(format!(
5933 "ALTER TABLE {q_table} ADD CONSTRAINT \"{cname}\" UNIQUE ({q_column})"
5934 ));
5935 } else {
5936 stmts.push(format!(
5937 "ALTER TABLE {q_table} DROP CONSTRAINT IF EXISTS \"{cname}\""
5938 ));
5939 }
5940 }
5941
5942 // DEFAULT change. Empty string in either snapshot means "no
5943 // default"; the canonical SET / DROP pair fully expresses
5944 // the transition.
5945 if prev_col.default != new.default {
5946 if new.default.is_empty() {
5947 stmts.push(format!(
5948 "ALTER TABLE {q_table} ALTER COLUMN {q_column} DROP DEFAULT"
5949 ));
5950 } else {
5951 let escaped = new.default.replace('\'', "''");
5952 stmts.push(format!(
5953 "ALTER TABLE {q_table} ALTER COLUMN {q_column} SET DEFAULT '{escaped}'"
5954 ));
5955 }
5956 }
5957
5958 // FK target / on_delete / on_update — these are all carried
5959 // on the same constraint, so any one of them flipping
5960 // requires a DROP + readd of the whole FK. Autogen name
5961 // convention `<table>_<col>_fkey` matches Postgres at CREATE
5962 // TABLE time. Only emitted when the new column is still a
5963 // FK; if the column stopped being a FK (ty changed away
5964 // from ForeignKey), the type-change branch above handles
5965 // it indirectly via the column type rewrite.
5966 let fk_changed = prev_col.fk_target != new.fk_target
5967 || prev_col.on_delete != new.on_delete
5968 || prev_col.on_update != new.on_update;
5969 if fk_changed && matches!(new.ty, SqlType::ForeignKey) {
5970 let cname = format!("{table}_{column}_fkey");
5971 stmts.push(format!(
5972 "ALTER TABLE {q_table} DROP CONSTRAINT IF EXISTS \"{cname}\""
5973 ));
5974 // gaps2 #22: only re-add the physical constraint when the FK
5975 // still wants one. A `db_constraint = false` FK keeps the
5976 // DROP (so flipping the flag tears down any prior constraint)
5977 // but emits no ADD CONSTRAINT.
5978 if let Some(target) = &new.fk_target
5979 && new.db_constraint
5980 {
5981 let q_target = quote_pg_ident(target);
5982 // Resolve the referenced PK column from the target model's
5983 // registered meta instead of hardcoding `"id"`. String/Uuid
5984 // PKs (e.g. `Permission.codename`) are first-class post-lift;
5985 // the CreateTable path already resolves via `fk_target_pk`
5986 // (build_column_def_postgres), so the re-add must match or it
5987 // aborts the migration ("column id does not exist") / attaches
5988 // the constraint to the wrong column.
5989 let (pk_col, _pk_ty) = fk_target_pk(&target.replace('"', "\"\""));
5990 let q_pk = quote_pg_ident(&pk_col);
5991 let on_delete_clause = new
5992 .on_delete
5993 .sql_keyword()
5994 .map(|k| format!(" ON DELETE {k}"))
5995 .unwrap_or_default();
5996 let on_update_clause = new
5997 .on_update
5998 .sql_keyword()
5999 .map(|k| format!(" ON UPDATE {k}"))
6000 .unwrap_or_default();
6001 stmts.push(format!(
6002 "ALTER TABLE {q_table} ADD CONSTRAINT \"{cname}\" \
6003 FOREIGN KEY ({q_column}) REFERENCES {q_target}({q_pk})\
6004 {on_delete_clause}{on_update_clause}{PG_FK_DEFERRABLE}"
6005 ));
6006 }
6007 }
6008
6009 // CHECK constraint (single-valued choices) change. MultiChoice
6010 // uses CSV storage which can't be expressed as a column-level
6011 // IN constraint; the runtime sqlx Decode path is the guard.
6012 if prev_col.choices != new.choices && !new.is_multichoice {
6013 let cname = format!("{table}_{column}_check");
6014 stmts.push(format!(
6015 "ALTER TABLE {q_table} DROP CONSTRAINT IF EXISTS \"{cname}\""
6016 ));
6017 if !new.choices.is_empty() {
6018 let values_sql = new
6019 .choices
6020 .iter()
6021 .map(|v| format!("'{}'", v.replace('\'', "''")))
6022 .collect::<Vec<_>>()
6023 .join(", ");
6024 stmts.push(format!(
6025 "ALTER TABLE {q_table} ADD CONSTRAINT \"{cname}\" \
6026 CHECK ({q_column} IN ({values_sql}))"
6027 ));
6028 }
6029 }
6030 }
6031
6032 // Defensive: if we somehow produced no statements (shouldn't
6033 // happen — diff_columns gates on at least one schema-meaningful
6034 // flag changing), fall back to a single redundant SET NULL flip
6035 // to match the legacy contract. Tests cover both branches; this
6036 // is belt-and-braces.
6037 if stmts.is_empty() {
6038 let clause = if new.nullable {
6039 "DROP NOT NULL"
6040 } else {
6041 "SET NOT NULL"
6042 };
6043 stmts.push(format!(
6044 "ALTER TABLE {q_table} ALTER COLUMN {q_column} {clause}"
6045 ));
6046 }
6047
6048 stmts
6049}
6050
6051/// Quote a SQL identifier the Postgres way: wrap in double quotes,
6052/// escape inner double quotes by doubling them. Matches sea-query's
6053/// `PostgresQueryBuilder` output for identifiers so the rendered
6054/// statements look uniform.
6055fn quote_pg_ident(ident: &str) -> String {
6056 format!("\"{}\"", ident.replace('"', "\"\""))
6057}
6058
6059/// Build a SQLite `ColumnDef`. SQLite has one important quirk: its
6060/// ROWID-alias mechanic (which gives a primary-key column auto-
6061/// increment behaviour out of the box) only fires when the column's
6062/// type is the exact text `INTEGER` — case-insensitive but no other
6063/// variant. `BIGINT PRIMARY KEY`, even on a column the M3 derive
6064/// declared as `i64`, does NOT auto-increment, so an `INSERT INTO t
6065/// (other_col) VALUES (...)` without an explicit PK value fails the
6066/// NOT NULL constraint. Every umbral user with an `id: i64` model
6067/// would hit this without the override.
6068///
6069/// The fix: when a column is a primary key with an integer SqlType
6070/// (Integer or BigInt), force the rendered type to `Integer` and
6071/// attach `auto_increment()` so the generated DDL reads `"id" integer
6072/// NOT NULL PRIMARY KEY AUTOINCREMENT`. SQLite stores both `i32` and
6073/// `i64` as INTEGER affinity anyway, so the override is a no-op
6074/// semantically — the rows that round-trip through `sqlx::FromRow`
6075/// deserialize back into `i64` cleanly.
6076///
6077/// For `SqlType::Uuid` PKs: SQLite stores UUIDs as TEXT. No
6078/// `DEFAULT gen_random_uuid()` is emitted; the application must supply
6079/// the UUID at create time (or pass `Uuid::nil()` to trigger the
6080/// omit-on-insert sentinel that leaves the column to a future default).
6081///
6082/// For `SqlType::ForeignKey` columns: rendered as `BIGINT` with a
6083/// `REFERENCES "<target>"("id")` suffix appended via `.extra()`. The
6084/// target table name comes from `col.fk_target`.
6085/// Look up the FK target model's primary-key column name and SQL
6086/// type. Walks the registered ModelMeta set to find the model whose
6087/// table matches `fk_target_table`, then picks the first column
6088/// marked `primary_key = true`. Falls back to `("id", BigInteger)`
6089/// when the target isn't registered (cross-plugin lookup miss, or
6090/// the FK points outside the framework's model registry).
6091///
6092/// Used by both the SQLite and Postgres FK column-def builders so the
6093/// generated `<col> <type> REFERENCES <tbl>(<pk_col>)` matches the
6094/// target's actual PK shape — gap #60 made non-`id`, non-i64 PKs
6095/// (e.g. `Permission.codename: String`) a real case.
6096fn fk_target_pk(fk_target_table: &str) -> (String, sea_query::ColumnType) {
6097 use sea_query::ColumnType;
6098 let unesc = fk_target_table.replace("\"\"", "\"");
6099 // Non-panicking registry read — `registered_models()` itself
6100 // panics when called outside an `App::build()` context, but the
6101 // migration engine's unit tests construct snapshots by hand and
6102 // call into DDL emit without booting the framework. Fall through
6103 // to the historical "id"/BigInteger default in that case.
6104 let Some(metas) = REGISTRY.get() else {
6105 return ("id".to_string(), ColumnType::BigInteger);
6106 };
6107 for meta in metas.iter().map(|(_, m)| m) {
6108 if meta.table != unesc {
6109 continue;
6110 }
6111 if let Some(pk) = meta.fields.iter().find(|c| c.primary_key) {
6112 // Map the PK's SqlType to a sea-query ColumnType. We can't
6113 // route through `SqliteBackend::map_column` because that
6114 // wants a `Column` and applies max_length / choices
6115 // metadata which is irrelevant to a FK column. Hand-roll
6116 // the few cases the framework supports for PKs.
6117 let ct = match pk.ty {
6118 SqlType::BigInt | SqlType::Integer => ColumnType::BigInteger,
6119 SqlType::SmallInt => ColumnType::SmallInteger,
6120 SqlType::Text => ColumnType::Text,
6121 SqlType::Uuid => ColumnType::Uuid,
6122 // Other PK types fall back to BigInteger as the
6123 // historical default. The compile-time PrimaryKey
6124 // trait keeps this list closed in practice.
6125 _ => ColumnType::BigInteger,
6126 };
6127 return (pk.name.clone(), ct);
6128 }
6129 }
6130 ("id".to_string(), ColumnType::BigInteger)
6131}
6132
6133fn build_column_def_sqlite(col: &Column) -> sea_query::ColumnDef {
6134 use sea_query::{Alias, ColumnDef, ColumnType};
6135
6136 // ForeignKey gets a special path: column type + inline REFERENCES
6137 // clause both derived from the target model's PK column.
6138 if matches!(col.ty, SqlType::ForeignKey) {
6139 let fk_target = col
6140 .fk_target
6141 .as_deref()
6142 .unwrap_or("_unknown_")
6143 .replace('"', "\"\"");
6144 let (pk_col_name, pk_col_type) = fk_target_pk(&fk_target);
6145 let mut def = ColumnDef::new_with_type(Alias::new(&col.name), pk_col_type);
6146 if !col.nullable {
6147 def.not_null();
6148 }
6149 // A FK column that is ALSO the primary key (an identifying relation —
6150 // Prisma's `@id` on a relation field) must carry the PRIMARY KEY
6151 // marking. The FK branch returns early, so it wouldn't otherwise reach
6152 // the general `col.primary_key` handling below. No AUTOINCREMENT: the
6153 // value is the referenced row's key, not a fresh sequence.
6154 if col.primary_key {
6155 def.primary_key();
6156 }
6157 // BUG-15: `#[umbral(unique)]` on a FK column is the
6158 // OneToOne idiom — emit UNIQUE inline so the
6159 // referencing-row uniqueness is enforced at the DB.
6160 // The FK branch used to skip this because it returned
6161 // before the non-FK unique branch ran. A PK column is already
6162 // unique, so don't double up.
6163 if col.unique && !col.primary_key {
6164 def.unique_key();
6165 }
6166 // gaps2 #22: `#[umbral(db_constraint = false)]` keeps the logical
6167 // FK (column type derived from the target PK, above) but emits
6168 // NO physical `REFERENCES` clause. This is the only valid shape
6169 // for a cross-database FK. The default (`true`) emits the
6170 // constraint as before.
6171 if col.db_constraint {
6172 def.extra(format!(
6173 "REFERENCES \"{fk_target}\"(\"{pk_col_name}\"){}",
6174 fk_action_suffix(col),
6175 ));
6176 }
6177 return def;
6178 }
6179
6180 let is_int_pk = col.primary_key && matches!(col.ty, SqlType::Integer | SqlType::BigInt);
6181
6182 let column_type = if is_int_pk {
6183 ColumnType::Integer
6184 } else {
6185 crate::backend::SqliteBackend.map_column(col)
6186 };
6187
6188 let mut def = ColumnDef::new_with_type(Alias::new(&col.name), column_type);
6189 if !col.nullable {
6190 def.not_null();
6191 }
6192 if col.primary_key {
6193 def.primary_key();
6194 if is_int_pk {
6195 def.auto_increment();
6196 }
6197 }
6198 // gaps3 #35: `#[umbral(case_insensitive)]` on SQLite lifts to a column-level
6199 // `COLLATE NOCASE`, so `=`, `UNIQUE`, and `ORDER BY` fold case while storage
6200 // keeps the original casing. Emitted before UNIQUE so the auto-created
6201 // unique index inherits the column's NOCASE collation. NOCASE folds ASCII
6202 // A–Z only (a boot check warns about Unicode); Postgres uses `citext`.
6203 if col.case_insensitive && matches!(col.ty, SqlType::Text) {
6204 def.extra("COLLATE NOCASE".to_string());
6205 }
6206 // `#[umbral(unique)]` lifts to a column-level UNIQUE clause.
6207 // Skipped on PK columns (already unique) so the DDL stays tidy.
6208 if col.unique && !col.primary_key {
6209 def.unique_key();
6210 }
6211 // IMP-3: `#[umbral(min = N)]` / `#[umbral(max = N)]` lift to a
6212 // column-level CHECK clause. Both SQLite and Postgres accept the
6213 // same syntax. The pre-validation in `insert_json`/`update_json`
6214 // catches violations earlier with a friendlier error; the CHECK
6215 // is the DB-side safety net against direct-SQL writers.
6216 if let Some(check) = check_min_max_sql(col) {
6217 def.extra(check);
6218 }
6219 // User-declared `#[umbral(default = "...")]` lifts to a DDL DEFAULT
6220 // clause. Required when emitting `ALTER TABLE ADD COLUMN` for a
6221 // NOT NULL column against a non-empty table (SQLite rejects the
6222 // ADD otherwise); on CREATE TABLE it sets the column-level default
6223 // the database uses when an INSERT omits the value.
6224 //
6225 // SQLite stores booleans as INTEGER; the literal `'true'` /
6226 // `'false'` would land as a TEXT default that fails type checks
6227 // on reads. Translate Boolean defaults to `1` / `0` so the
6228 // stored representation matches what sqlx expects on hydration
6229 // (closes IMP-2 in bugs/tests/testBugs.md).
6230 if !col.default.is_empty() {
6231 if matches!(col.ty, SqlType::Boolean) {
6232 // Pass an integer to sea-query so the rendered SQL is
6233 // `DEFAULT 1` / `DEFAULT 0` instead of the quoted-string
6234 // `DEFAULT '1'` (which sqlx rejects as TEXT on read of
6235 // a BOOLEAN column).
6236 def.default(sqlite_bool_default(&col.default));
6237 } else {
6238 def.default(col.default.clone());
6239 }
6240 }
6241 // NOTE: auto_now / auto_now_add deliberately does NOT emit a
6242 // `DEFAULT CURRENT_TIMESTAMP` here. SQLite rejects non-constant
6243 // defaults in `ALTER TABLE ADD COLUMN` ("Cannot add a column
6244 // with non-constant default") and that's the path that matters
6245 // for evolving an existing table. The SQLite `AddColumn` render
6246 // path handles the auto_now backfill via a two-statement
6247 // sequence (nullable ADD + UPDATE backfill). On CREATE TABLE
6248 // we don't need a default at all because every INSERT goes
6249 // through the macro-emitted Rust path which always supplies the
6250 // value. See `Operation::AddColumn` render below.
6251 def
6252}
6253
6254/// Render a column's `#[umbral(default = ...)]` value as a raw SQL literal for
6255/// a hand-built statement (the NOT-NULL backfill, audit_2 core-migrate #5).
6256/// sea-query quotes literals itself in the column-def path, but the backfill
6257/// `UPDATE`/`COALESCE` is a raw `format!`, so it needs the literal here.
6258/// Numeric and boolean types render unquoted (`0`, `true`); everything else is
6259/// a single-quoted string with inner quotes doubled. `is_postgres` only affects
6260/// booleans (`true`/`false` on PG, `1`/`0` on SQLite, matching each backend's
6261/// boolean storage).
6262fn default_sql_literal(col: &Column, is_postgres: bool) -> String {
6263 use crate::orm::SqlType::*;
6264 match col.ty {
6265 Boolean => {
6266 let truthy = matches!(
6267 col.default.trim().to_ascii_lowercase().as_str(),
6268 "true" | "1" | "t" | "yes"
6269 );
6270 if is_postgres {
6271 if truthy { "true" } else { "false" }.to_string()
6272 } else if truthy {
6273 "1".to_string()
6274 } else {
6275 "0".to_string()
6276 }
6277 }
6278 SmallInt | Integer | BigInt | Real | Double | Decimal | ForeignKey => {
6279 // Numeric literal — validated at derive time; emit unquoted.
6280 col.default.clone()
6281 }
6282 _ => format!("'{}'", col.default.replace('\'', "''")),
6283 }
6284}
6285
6286/// Map a user-supplied boolean default string (`"true"` / `"false"`
6287/// / `"1"` / `"0"`, case-insensitive) to the SQLite integer literal
6288/// the column expects. Anything unrecognised falls through to `0`
6289/// — a developer-visible miss (default is wrong, not stored as
6290/// text) is friendlier than the runtime decode error the textual
6291/// path produces.
6292fn sqlite_bool_default(raw: &str) -> i32 {
6293 match raw.trim().to_ascii_lowercase().as_str() {
6294 "true" | "1" | "t" | "yes" => 1,
6295 _ => 0,
6296 }
6297}
6298
6299/// IMP-3: lower `#[umbral(min = N)]` / `#[umbral(max = N)]` to a
6300/// DDL CHECK clause. Returns `None` when the column declares
6301/// neither bound. The rendered SQL works on both SQLite and
6302/// Postgres (`"<col>" >= N`, `"<col>" <= N`, joined by `AND`).
6303/// Only applied to numeric columns — applying it to text would
6304/// compare strings lexicographically and surprise everyone.
6305fn check_min_max_sql(col: &Column) -> Option<String> {
6306 if col.min.is_none() && col.max.is_none() {
6307 return None;
6308 }
6309 if !matches!(
6310 col.ty,
6311 SqlType::SmallInt | SqlType::Integer | SqlType::BigInt | SqlType::Real | SqlType::Double
6312 ) {
6313 return None;
6314 }
6315 let name = col.name.replace('"', "\"\"");
6316 let mut parts = Vec::with_capacity(2);
6317 if let Some(n) = col.min {
6318 parts.push(format!("\"{name}\" >= {n}"));
6319 }
6320 if let Some(n) = col.max {
6321 parts.push(format!("\"{name}\" <= {n}"));
6322 }
6323 Some(format!("CHECK ({})", parts.join(" AND ")))
6324}
6325
6326/// Build a Postgres `ColumnDef`. Integer primary keys use the
6327/// standard `auto_increment()` flag — sea-query's `PostgresQueryBuilder`
6328/// lowers that to `BIGSERIAL` for `BigInt` and `SERIAL` for `Integer`.
6329/// No SQLite-style INTEGER-type override needed; Postgres has proper
6330/// `BIGSERIAL` / identity columns and respects the declared width.
6331///
6332/// For `SqlType::ForeignKey` columns: rendered as `BIGINT` with a
6333/// `REFERENCES "<target>"("id")` suffix. The target table name comes
6334/// from `col.fk_target`.
6335fn build_column_def_postgres(col: &Column) -> sea_query::ColumnDef {
6336 use sea_query::{Alias, ColumnDef};
6337
6338 // ForeignKey gets a special path: column type + inline REFERENCES
6339 // clause both derived from the target model's PK.
6340 if matches!(col.ty, SqlType::ForeignKey) {
6341 let fk_target = col
6342 .fk_target
6343 .as_deref()
6344 .unwrap_or("_unknown_")
6345 .replace('"', "\"\"");
6346 let (pk_col_name, pk_col_type) = fk_target_pk(&fk_target);
6347 // sea-query's ColumnType variants are dialect-agnostic; the
6348 // same value works for both SQLite and Postgres builders here.
6349 let mut def = ColumnDef::new_with_type(Alias::new(&col.name), pk_col_type);
6350 if !col.nullable {
6351 def.not_null();
6352 }
6353 // BUG-15: `#[umbral(unique)]` on a FK column is the
6354 // OneToOne idiom — emit UNIQUE inline so the
6355 // referencing-row uniqueness is enforced at the DB.
6356 // The FK branch used to skip this because it returned
6357 // before the non-FK unique branch ran.
6358 if col.unique {
6359 def.unique_key();
6360 }
6361 // gaps2 #22: skip the physical `REFERENCES` clause when the FK
6362 // opted out of the DB constraint (cross-database FK). The
6363 // logical column + `fk_target` stay intact.
6364 if col.db_constraint {
6365 def.extra(format!(
6366 "REFERENCES \"{fk_target}\"(\"{pk_col_name}\"){}{PG_FK_DEFERRABLE}",
6367 fk_action_suffix(col),
6368 ));
6369 }
6370 return def;
6371 }
6372
6373 let column_type = crate::backend::PostgresBackend.map_column(col);
6374
6375 let mut def = ColumnDef::new_with_type(Alias::new(&col.name), column_type);
6376 if !col.nullable {
6377 def.not_null();
6378 }
6379 if col.primary_key {
6380 def.primary_key();
6381 if matches!(
6382 col.ty,
6383 SqlType::Integer | SqlType::BigInt | SqlType::SmallInt
6384 ) {
6385 def.auto_increment();
6386 }
6387 }
6388 // `#[umbral(unique)]` lifts to a column-level UNIQUE clause on
6389 // Postgres too. Skipped for PK columns (already unique).
6390 if col.unique && !col.primary_key {
6391 def.unique_key();
6392 }
6393 // IMP-3: numeric bounds CHECK. Mirrors the SQLite branch.
6394 if let Some(check) = check_min_max_sql(col) {
6395 def.extra(check);
6396 }
6397 // Single-valued Choices: emit a CHECK constraint so a third-party
6398 // process writing directly to the DB can't insert a value the Rust
6399 // enum can't model. MultiChoice carries the same choices/labels
6400 // metadata but the stored value is a CSV — a single-value `IN (...)`
6401 // constraint would reject every legal CSV. Validating "every CSV
6402 // piece is a known variant" needs a regex with per-variant
6403 // escaping, which we leave to the sqlx Decode path at v1.
6404 if !col.choices.is_empty() && !col.is_multichoice {
6405 let col_name_escaped = col.name.replace('"', "\"\"");
6406 let values_sql = col
6407 .choices
6408 .iter()
6409 .map(|v| format!("'{}'", v.replace('\'', "''")))
6410 .collect::<Vec<_>>()
6411 .join(", ");
6412 def.extra(format!("CHECK (\"{col_name_escaped}\" IN ({values_sql}))"));
6413 }
6414 // User-declared `#[umbral(default = "...")]` lifts to a DDL DEFAULT
6415 // clause. Required for `ALTER TABLE ADD COLUMN` of a NOT NULL
6416 // column against a non-empty table — Postgres needs either a
6417 // default or a separate backfill.
6418 if !col.default.is_empty() {
6419 def.default(col.default.clone());
6420 } else if (col.auto_now || col.auto_now_add)
6421 && matches!(
6422 col.ty,
6423 SqlType::Timestamptz | SqlType::Timestamp | SqlType::Date | SqlType::Time
6424 )
6425 {
6426 // Mirror of the SQLite branch above. Without a DEFAULT
6427 // Postgres rejects `ALTER TABLE ADD COLUMN ... NOT NULL`
6428 // on a populated table. `now()` evaluates per-row during
6429 // the backfill so every existing row gets a sane value;
6430 // future INSERTs override via the macro-emitted Rust path.
6431 def.default(sea_query::Expr::cust("now()"));
6432 }
6433 def
6434}
6435
6436#[cfg(test)]
6437mod tests {
6438 use super::*;
6439 use std::collections::HashSet;
6440
6441 // ---- gaps2 #100: squash_plan (pure squash-vs-originals decision) ----
6442
6443 /// A minimal migration file for `plugin`/`id`; `replaces` lists the ids it
6444 /// squashes (all within `plugin`). Empty `replaces` = an ordinary migration.
6445 fn mf(plugin: &str, id: &str, replaces: &[&str]) -> MigrationFile {
6446 MigrationFile {
6447 id: id.to_string(),
6448 plugin: plugin.to_string(),
6449 depends_on: Vec::new(),
6450 operations: Vec::new(),
6451 snapshot_after: Snapshot::default(),
6452 replaces: replaces
6453 .iter()
6454 .map(|m| MigrationRef {
6455 plugin: plugin.to_string(),
6456 migration: m.to_string(),
6457 })
6458 .collect(),
6459 }
6460 }
6461
6462 fn applied_set(plugin: &str, ids: &[&str]) -> HashSet<(String, String)> {
6463 ids.iter()
6464 .map(|id| (plugin.to_string(), id.to_string()))
6465 .collect()
6466 }
6467
6468 #[test]
6469 fn squash_plan_plain_migrations_apply_when_unrecorded_skip_when_applied() {
6470 let files = vec![mf("app", "0001", &[]), mf("app", "0002", &[])];
6471 let applied = applied_set("app", &["0001"]);
6472 let plan = squash_plan(&files, &applied).unwrap();
6473 assert_eq!(plan, vec![ApplyDecision::Skip, ApplyDecision::Apply]);
6474 }
6475
6476 #[test]
6477 fn squash_plan_fresh_db_applies_squash_and_shadows_its_originals() {
6478 // Both the squash and its originals are on disk (Django keeps both).
6479 // Nothing applied → run the squash once, skip every original.
6480 let files = vec![
6481 mf("app", "0001", &[]),
6482 mf("app", "0002", &[]),
6483 mf("app", "0001_squashed_0002", &["0001", "0002"]),
6484 ];
6485 let applied = applied_set("app", &[]);
6486 let plan = squash_plan(&files, &applied).unwrap();
6487 assert_eq!(
6488 plan,
6489 vec![
6490 ApplyDecision::Skip, // 0001 shadowed by the squash
6491 ApplyDecision::Skip, // 0002 shadowed by the squash
6492 ApplyDecision::Apply // the squash builds the whole schema
6493 ]
6494 );
6495 }
6496
6497 #[test]
6498 fn squash_plan_existing_db_with_full_history_record_only() {
6499 // Both originals already applied → the squash records itself without
6500 // running (its schema already exists), originals skipped.
6501 let files = vec![
6502 mf("app", "0001", &[]),
6503 mf("app", "0002", &[]),
6504 mf("app", "0001_squashed_0002", &["0001", "0002"]),
6505 ];
6506 let applied = applied_set("app", &["0001", "0002"]);
6507 let plan = squash_plan(&files, &applied).unwrap();
6508 assert_eq!(
6509 plan,
6510 vec![
6511 ApplyDecision::Skip,
6512 ApplyDecision::Skip,
6513 ApplyDecision::RecordOnly
6514 ]
6515 );
6516 }
6517
6518 #[test]
6519 fn squash_plan_already_recorded_squash_is_skipped() {
6520 // The squash ran on a previous migrate; it's in the tracking table.
6521 let files = vec![
6522 mf("app", "0001", &[]),
6523 mf("app", "0002", &[]),
6524 mf("app", "0001_squashed_0002", &["0001", "0002"]),
6525 ];
6526 let applied = applied_set("app", &["0001_squashed_0002"]);
6527 let plan = squash_plan(&files, &applied).unwrap();
6528 // originals never applied but shadowed by the (already-applied) squash.
6529 assert_eq!(
6530 plan,
6531 vec![
6532 ApplyDecision::Skip,
6533 ApplyDecision::Skip,
6534 ApplyDecision::Skip
6535 ]
6536 );
6537 }
6538
6539 #[test]
6540 fn squash_plan_partial_transition_falls_back_to_surviving_originals() {
6541 // 0001 applied, 0002 not; originals still on disk → ignore the squash,
6542 // finish 0002 individually.
6543 let files = vec![
6544 mf("app", "0001", &[]),
6545 mf("app", "0002", &[]),
6546 mf("app", "0001_squashed_0002", &["0001", "0002"]),
6547 ];
6548 let applied = applied_set("app", &["0001"]);
6549 let plan = squash_plan(&files, &applied).unwrap();
6550 assert_eq!(
6551 plan,
6552 vec![
6553 ApplyDecision::Skip, // 0001 already applied
6554 ApplyDecision::Apply, // 0002 finishes individually
6555 ApplyDecision::Skip // squash ignored during the transition
6556 ]
6557 );
6558 }
6559
6560 #[test]
6561 fn squash_plan_partial_transition_with_missing_original_errors() {
6562 // 0001 applied, 0002 not, and 0002's file is GONE → can't reconcile.
6563 let files = vec![
6564 mf("app", "0001", &[]),
6565 mf("app", "0001_squashed_0002", &["0001", "0002"]),
6566 ];
6567 let applied = applied_set("app", &["0001"]);
6568 let err = squash_plan(&files, &applied).unwrap_err();
6569 match err {
6570 MigrateError::SquashInconsistent {
6571 plugin,
6572 squash,
6573 missing,
6574 } => {
6575 assert_eq!(plugin, "app");
6576 assert_eq!(squash, "0001_squashed_0002");
6577 assert_eq!(missing, vec!["app/0002".to_string()]);
6578 }
6579 other => panic!("expected SquashInconsistent, got {other:?}"),
6580 }
6581 }
6582
6583 /// audit_2 core-migrate #7 — the advisory-lock key must be deterministic
6584 /// (every process computes the same key for the same target, so they
6585 /// mutually exclude) and distinct per discriminator (different aliases /
6586 /// schemas migrate concurrently). Pins the FNV constants so a refactor that
6587 /// changes the hash — silently breaking cross-process exclusion — fails.
6588 #[test]
6589 fn pg_migration_lock_key_is_deterministic_and_distinct() {
6590 // Deterministic: same input → same key, run to run, process to process.
6591 assert_eq!(
6592 pg_migration_lock_key("default"),
6593 pg_migration_lock_key("default"),
6594 );
6595 // Distinct: different aliases/schemas get different keys.
6596 assert_ne!(
6597 pg_migration_lock_key("default"),
6598 pg_migration_lock_key("replica"),
6599 );
6600 assert_ne!(
6601 pg_migration_lock_key("tenant_a"),
6602 pg_migration_lock_key("tenant_b"),
6603 );
6604 // Pin the exact value so the hash can't drift unnoticed (two binaries on
6605 // different umbral versions must still agree on the key).
6606 assert_eq!(
6607 pg_migration_lock_key("default"),
6608 {
6609 let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
6610 for b in b"umbral_migrations\0"
6611 .iter()
6612 .copied()
6613 .chain(b"default".iter().copied())
6614 {
6615 hash ^= b as u64;
6616 hash = hash.wrapping_mul(0x0000_0100_0000_01b3);
6617 }
6618 hash as i64
6619 },
6620 "the lock-key hash changed — this breaks cross-process exclusion \
6621 between an old and a new migrator; bump deliberately if intended",
6622 );
6623 }
6624
6625 /// M8 — `plugin_order()` falls back to `registered_plugins()` when
6626 /// no topological order has been published. The fallback keeps the
6627 /// engine usable from low-level paths that drive `init_plugins`
6628 /// directly (the M5 / M6 tests that pre-date phase 1.5 of
6629 /// `App::build()`).
6630 ///
6631 /// Runs in the lib's unit-test binary, which is wholly separate
6632 /// from the integration test binaries and so owns its own copies
6633 /// of `REGISTRY` and `PLUGIN_ORDER`. This test seeds `REGISTRY` via
6634 /// `init_plugins`, never touches `init_plugin_order`, and pins the
6635 /// fallback to the sorted-by-name `registered_plugins()` output.
6636 /// As the only test that touches either OnceLock in this binary,
6637 /// it has them to itself.
6638 #[test]
6639 fn plugin_order_falls_back_to_registered_plugins_when_unpublished() {
6640 let mut per_plugin: std::collections::HashMap<String, Vec<ModelMeta>> =
6641 std::collections::HashMap::new();
6642 per_plugin.insert(
6643 "zeta".to_string(),
6644 vec![ModelMeta {
6645 view: None,
6646 materialized: false,
6647 name: "ZetaModel".to_string(),
6648 table: "zeta".to_string(),
6649 fields: Vec::new(),
6650 display: "ZetaModel".to_string(),
6651 icon: "database".to_string(),
6652 database: None,
6653 singleton: false,
6654 unique_together: Vec::new(),
6655 indexes: Vec::new(),
6656 ordering: Vec::new(),
6657 m2m_relations: Vec::new(),
6658 soft_delete: false,
6659 audited: false,
6660 app_label: "app".to_string(),
6661 }],
6662 );
6663 per_plugin.insert(
6664 "alpha".to_string(),
6665 vec![ModelMeta {
6666 view: None,
6667 materialized: false,
6668 name: "AlphaModel".to_string(),
6669 table: "alpha".to_string(),
6670 fields: Vec::new(),
6671 display: "AlphaModel".to_string(),
6672 icon: "database".to_string(),
6673 database: None,
6674 singleton: false,
6675 unique_together: Vec::new(),
6676 indexes: Vec::new(),
6677 ordering: Vec::new(),
6678 m2m_relations: Vec::new(),
6679 soft_delete: false,
6680 audited: false,
6681 app_label: "app".to_string(),
6682 }],
6683 );
6684 init_plugins(per_plugin);
6685
6686 // `init_plugin_order` was never called, so `plugin_order` must
6687 // return the sorted-by-name fallback.
6688 let order = plugin_order();
6689 assert_eq!(
6690 order,
6691 vec!["alpha".to_string(), "zeta".to_string()],
6692 "fallback should sort by name; got {order:?}",
6693 );
6694 assert_eq!(
6695 order,
6696 registered_plugins(),
6697 "fallback should exactly equal registered_plugins()",
6698 );
6699 }
6700
6701 /// Gap #65: `#[umbral(unique)]` lifts to a column-level UNIQUE in
6702 /// CREATE TABLE DDL on both backends. PK columns skip the clause
6703 /// because they're already unique by virtue of being the PK.
6704 #[test]
6705 fn unique_column_emits_unique_keyword_on_both_backends() {
6706 use sea_query::{Alias, PostgresQueryBuilder, SqliteQueryBuilder, Table};
6707
6708 let id = Column {
6709 name: "id".into(),
6710 ty: SqlType::BigInt,
6711 primary_key: true,
6712 nullable: false,
6713 fk_target: None,
6714 noform: false,
6715 privileged: false,
6716 private: false,
6717 secret: false,
6718 db_constraint: true,
6719 noedit: false,
6720 auto_user_add: false,
6721 auto_user: false,
6722 is_string_repr: false,
6723 max_length: 0,
6724 choices: vec![],
6725 choice_labels: vec![],
6726 default: String::new(),
6727 is_multichoice: false,
6728 // Set even though it's a PK so we can assert below that
6729 // the emit path drops the redundant clause.
6730 unique: true,
6731 on_delete: crate::orm::FkAction::NoAction,
6732 on_update: crate::orm::FkAction::NoAction,
6733 index: false,
6734 auto_now_add: false,
6735 auto_uuid: false,
6736 auto_now: false,
6737 trim: false,
6738 lowercase: false,
6739 case_insensitive: false,
6740 help: String::new(),
6741 example: String::new(),
6742 widget: None,
6743 supported_backends: Vec::new(),
6744 min: None,
6745 max: None,
6746 text_format: ::core::option::Option::None,
6747 slug_from: ::core::option::Option::None,
6748 };
6749 let username = Column {
6750 name: "username".into(),
6751 ty: SqlType::Text,
6752 primary_key: false,
6753 nullable: false,
6754 fk_target: None,
6755 noform: false,
6756 privileged: false,
6757 private: false,
6758 secret: false,
6759 db_constraint: true,
6760 noedit: false,
6761 auto_user_add: false,
6762 auto_user: false,
6763 is_string_repr: false,
6764 max_length: 0,
6765 choices: vec![],
6766 choice_labels: vec![],
6767 default: String::new(),
6768 is_multichoice: false,
6769 unique: true,
6770 on_delete: crate::orm::FkAction::NoAction,
6771 on_update: crate::orm::FkAction::NoAction,
6772 index: false,
6773 auto_now_add: false,
6774 auto_uuid: false,
6775 auto_now: false,
6776 trim: false,
6777 lowercase: false,
6778 case_insensitive: false,
6779 help: String::new(),
6780 example: String::new(),
6781 widget: None,
6782 supported_backends: Vec::new(),
6783 min: None,
6784 max: None,
6785 text_format: ::core::option::Option::None,
6786 slug_from: ::core::option::Option::None,
6787 };
6788 let email = Column {
6789 name: "email".into(),
6790 ty: SqlType::Text,
6791 primary_key: false,
6792 nullable: false,
6793 fk_target: None,
6794 noform: false,
6795 privileged: false,
6796 private: false,
6797 secret: false,
6798 db_constraint: true,
6799 noedit: false,
6800 auto_user_add: false,
6801 auto_user: false,
6802 is_string_repr: false,
6803 max_length: 0,
6804 choices: vec![],
6805 choice_labels: vec![],
6806 default: String::new(),
6807 is_multichoice: false,
6808 unique: false,
6809 on_delete: crate::orm::FkAction::NoAction,
6810 on_update: crate::orm::FkAction::NoAction,
6811 index: false,
6812 auto_now_add: false,
6813 auto_uuid: false,
6814 auto_now: false,
6815 trim: false,
6816 lowercase: false,
6817 case_insensitive: false,
6818 help: String::new(),
6819 example: String::new(),
6820 widget: None,
6821 supported_backends: Vec::new(),
6822 min: None,
6823 max: None,
6824 text_format: ::core::option::Option::None,
6825 slug_from: ::core::option::Option::None,
6826 };
6827
6828 for backend in ["sqlite", "postgres"] {
6829 let mut stmt = Table::create();
6830 stmt.table(Alias::new("u"));
6831 for col in [&id, &username, &email] {
6832 let mut def = if backend == "sqlite" {
6833 build_column_def_sqlite(col)
6834 } else {
6835 build_column_def_postgres(col)
6836 };
6837 stmt.col(&mut def);
6838 }
6839 let sql = if backend == "sqlite" {
6840 stmt.to_string(SqliteQueryBuilder)
6841 } else {
6842 stmt.to_string(PostgresQueryBuilder)
6843 };
6844
6845 // UNIQUE on the explicitly-marked non-PK column.
6846 assert!(
6847 sql.contains("\"username\"") && sql.to_uppercase().contains("UNIQUE"),
6848 "{backend}: expected UNIQUE on username; got: {sql}",
6849 );
6850 // No UNIQUE on `email` (flag false).
6851 let email_clause = sql
6852 .split("\"email\"")
6853 .nth(1)
6854 .unwrap_or_default()
6855 .split(',')
6856 .next()
6857 .unwrap_or_default();
6858 assert!(
6859 !email_clause.to_uppercase().contains("UNIQUE"),
6860 "{backend}: email should not be UNIQUE; clause: {email_clause}",
6861 );
6862 // PK still PK; the redundant UNIQUE flag is dropped so we
6863 // don't double up the constraint.
6864 let id_clause = sql
6865 .split("\"id\"")
6866 .nth(1)
6867 .unwrap_or_default()
6868 .split(',')
6869 .next()
6870 .unwrap_or_default();
6871 assert!(
6872 id_clause.to_uppercase().contains("PRIMARY KEY"),
6873 "{backend}: id should still be PRIMARY KEY; clause: {id_clause}",
6874 );
6875 assert!(
6876 !id_clause.to_uppercase().contains("UNIQUE"),
6877 "{backend}: PK column should not also carry UNIQUE; clause: {id_clause}",
6878 );
6879 }
6880 }
6881
6882 /// Gap #68: `on_delete` / `on_update` lift to the `REFERENCES`
6883 /// tail in DDL. `NoAction` emits no clause (the SQL default);
6884 /// any other variant emits `ON DELETE <kw>` / `ON UPDATE <kw>`
6885 /// on both backends. The clause goes inside the same `extra(...)`
6886 /// string that already carries `REFERENCES "<target>"("id")` —
6887 /// the test asserts the full tail shape so a future refactor
6888 /// that splits the FK rendering won't silently regress.
6889 #[test]
6890 fn fk_action_lifts_to_references_clause_on_both_backends() {
6891 use sea_query::{Alias, PostgresQueryBuilder, SqliteQueryBuilder, Table};
6892
6893 // Need an FK target table; the DDL renderer looks up the
6894 // PK column type for `auth_user` via `fk_target_pk`.
6895 // Using "post" since it's already registered as a real
6896 // Model in the lib (resolves to BigInt id).
6897 let plain_fk = Column {
6898 name: "owner_id".into(),
6899 ty: SqlType::ForeignKey,
6900 primary_key: false,
6901 nullable: false,
6902 fk_target: Some("post".into()),
6903 noform: false,
6904 privileged: false,
6905 private: false,
6906 secret: false,
6907 db_constraint: true,
6908 noedit: false,
6909 auto_user_add: false,
6910 auto_user: false,
6911 is_string_repr: false,
6912 max_length: 0,
6913 choices: vec![],
6914 choice_labels: vec![],
6915 default: String::new(),
6916 is_multichoice: false,
6917 unique: false,
6918 on_delete: crate::orm::FkAction::NoAction,
6919 on_update: crate::orm::FkAction::NoAction,
6920 index: false,
6921 auto_now_add: false,
6922 auto_uuid: false,
6923 auto_now: false,
6924 trim: false,
6925 lowercase: false,
6926 case_insensitive: false,
6927 help: String::new(),
6928 example: String::new(),
6929 widget: None,
6930 supported_backends: Vec::new(),
6931 min: None,
6932 max: None,
6933 text_format: ::core::option::Option::None,
6934 slug_from: ::core::option::Option::None,
6935 };
6936 let cascade_fk = Column {
6937 on_delete: crate::orm::FkAction::Cascade,
6938 on_update: crate::orm::FkAction::Cascade,
6939 index: false,
6940 auto_now_add: false,
6941 auto_uuid: false,
6942 auto_now: false,
6943 trim: false,
6944 lowercase: false,
6945 case_insensitive: false,
6946 help: String::new(),
6947 example: String::new(),
6948 widget: None,
6949 supported_backends: Vec::new(),
6950 ..plain_fk.clone()
6951 };
6952 let restrict_fk = Column {
6953 on_delete: crate::orm::FkAction::Restrict,
6954 ..plain_fk.clone()
6955 };
6956 let set_null_fk = Column {
6957 nullable: true,
6958 on_delete: crate::orm::FkAction::SetNull,
6959 ..plain_fk.clone()
6960 };
6961
6962 for backend in ["sqlite", "postgres"] {
6963 let render_one = |col: &Column| -> String {
6964 let mut stmt = Table::create();
6965 stmt.table(Alias::new("t"));
6966 let mut def = if backend == "sqlite" {
6967 build_column_def_sqlite(col)
6968 } else {
6969 build_column_def_postgres(col)
6970 };
6971 stmt.col(&mut def);
6972 if backend == "sqlite" {
6973 stmt.to_string(SqliteQueryBuilder)
6974 } else {
6975 stmt.to_string(PostgresQueryBuilder)
6976 }
6977 };
6978
6979 // NoAction → REFERENCES with no tail clauses.
6980 let sql = render_one(&plain_fk);
6981 assert!(
6982 sql.contains("REFERENCES")
6983 && !sql.to_uppercase().contains("ON DELETE")
6984 && !sql.to_uppercase().contains("ON UPDATE"),
6985 "{backend}: NoAction should emit REFERENCES alone; got: {sql}",
6986 );
6987
6988 // Cascade on both ON DELETE and ON UPDATE.
6989 let sql = render_one(&cascade_fk);
6990 assert!(
6991 sql.to_uppercase().contains("ON DELETE CASCADE")
6992 && sql.to_uppercase().contains("ON UPDATE CASCADE"),
6993 "{backend}: Cascade should emit both clauses; got: {sql}",
6994 );
6995
6996 // Restrict on ON DELETE only; ON UPDATE is NoAction so
6997 // no clause appears.
6998 let sql = render_one(&restrict_fk);
6999 assert!(
7000 sql.to_uppercase().contains("ON DELETE RESTRICT"),
7001 "{backend}: Restrict missing; got: {sql}",
7002 );
7003 assert!(
7004 !sql.to_uppercase().contains("ON UPDATE"),
7005 "{backend}: ON UPDATE shouldn't appear for NoAction; got: {sql}",
7006 );
7007
7008 // SET NULL renders verbatim (two-word keyword).
7009 let sql = render_one(&set_null_fk);
7010 assert!(
7011 sql.to_uppercase().contains("ON DELETE SET NULL"),
7012 "{backend}: SET NULL missing; got: {sql}",
7013 );
7014 }
7015 }
7016
7017 /// Gap #65 follow-up: the diff engine detects changes to *every*
7018 /// schema-meaningful field, not just `ty` and `nullable`. Each
7019 /// branch builds a baseline column, mutates one field, runs
7020 /// `diff_columns`, and asserts an `AlterColumn` op is produced.
7021 /// Catches the regression where toggling `unique` or `on_delete`
7022 /// would silently leave the table unchanged.
7023 #[test]
7024 fn diff_detects_all_schema_meaningful_field_changes() {
7025 fn baseline() -> Column {
7026 Column {
7027 name: "x".into(),
7028 ty: SqlType::Text,
7029 primary_key: false,
7030 nullable: false,
7031 fk_target: None,
7032 noform: false,
7033 privileged: false,
7034 private: false,
7035 secret: false,
7036 db_constraint: true,
7037 noedit: false,
7038 auto_user_add: false,
7039 auto_user: false,
7040 is_string_repr: false,
7041 max_length: 0,
7042 choices: vec![],
7043 choice_labels: vec![],
7044 default: String::new(),
7045 is_multichoice: false,
7046 unique: false,
7047 on_delete: crate::orm::FkAction::NoAction,
7048 on_update: crate::orm::FkAction::NoAction,
7049 index: false,
7050 auto_now_add: false,
7051 auto_uuid: false,
7052 auto_now: false,
7053 trim: false,
7054 lowercase: false,
7055 case_insensitive: false,
7056 help: String::new(),
7057 example: String::new(),
7058 widget: None,
7059 supported_backends: Vec::new(),
7060 min: None,
7061 max: None,
7062 text_format: ::core::option::Option::None,
7063 slug_from: ::core::option::Option::None,
7064 }
7065 }
7066 fn meta_with(col: Column) -> ModelMeta {
7067 ModelMeta {
7068 view: None,
7069 materialized: false,
7070 name: "M".into(),
7071 table: "m".into(),
7072 fields: vec![col],
7073 display: "M".into(),
7074 icon: "database".into(),
7075 database: None,
7076 singleton: false,
7077 unique_together: Vec::new(),
7078 indexes: Vec::new(),
7079 ordering: Vec::new(),
7080 m2m_relations: Vec::new(),
7081 soft_delete: false,
7082 audited: false,
7083 app_label: "app".into(),
7084 }
7085 }
7086 let prev = meta_with(baseline());
7087 // Safe-to-alter changes: each must surface as an `AlterColumn`.
7088 // (`nullable` here is false→true — a *loosening*, which is safe;
7089 // the tightening direction is guarded separately below.)
7090 let safe_mutations: Vec<(&str, fn(&mut Column))> = vec![
7091 ("default", |c| c.default = "hello".into()),
7092 ("choices", |c| {
7093 c.choices = vec!["a".into(), "b".into()];
7094 c.choice_labels = vec!["A".into(), "B".into()];
7095 }),
7096 ("nullable", |c| c.nullable = true),
7097 ];
7098 for (label, mutate) in safe_mutations {
7099 let mut col = baseline();
7100 mutate(&mut col);
7101 let current = meta_with(col);
7102 let ops = diff_columns("M", &prev, ¤t).expect("diff should succeed");
7103 assert!(
7104 !ops.is_empty(),
7105 "{label}: diff should produce at least one op; got none",
7106 );
7107 assert!(
7108 ops.iter()
7109 .any(|op| matches!(op, Operation::AlterColumn { column, .. } if column == "x")),
7110 "{label}: expected AlterColumn on `x`; got: {ops:?}",
7111 );
7112 }
7113
7114 // Adding UNIQUE to an existing column is detected too, but as an
7115 // `UnsafeAlter` guard rather than a bare `AlterColumn`: dropping a
7116 // UNIQUE constraint onto a populated column aborts the migration
7117 // if duplicates already exist, so the engine refuses it with a
7118 // duplicate-pre-check message instead of silently emitting it.
7119 let mut col = baseline();
7120 col.unique = true;
7121 let current = meta_with(col);
7122 match diff_columns("M", &prev, ¤t) {
7123 Err(MigrateError::UnsafeAlter { column, reason, .. }) => {
7124 assert_eq!(column, "x");
7125 assert!(
7126 reason.contains("UNIQUE"),
7127 "unsafe-alter reason should mention UNIQUE; got: {reason}",
7128 );
7129 }
7130 other => panic!("unique add should be an UnsafeAlter guard; got: {other:?}"),
7131 }
7132 }
7133
7134 /// Gap #65 follow-up: the Postgres `AlterColumn` render handles
7135 /// the new diff types (unique, default, choices, FK actions)
7136 /// with native `ALTER TABLE ... ADD/DROP CONSTRAINT` /
7137 /// `SET/DROP DEFAULT` statements. SQLite is unchanged — the
7138 /// rebuild dance already swallows any column metadata change.
7139 #[test]
7140 fn postgres_alter_column_renders_constraint_changes() {
7141 let baseline = Column {
7142 name: "x".into(),
7143 ty: SqlType::Text,
7144 primary_key: false,
7145 nullable: false,
7146 fk_target: None,
7147 noform: false,
7148 privileged: false,
7149 private: false,
7150 secret: false,
7151 db_constraint: true,
7152 noedit: false,
7153 auto_user_add: false,
7154 auto_user: false,
7155 is_string_repr: false,
7156 max_length: 0,
7157 choices: vec![],
7158 choice_labels: vec![],
7159 default: String::new(),
7160 is_multichoice: false,
7161 unique: false,
7162 on_delete: crate::orm::FkAction::NoAction,
7163 on_update: crate::orm::FkAction::NoAction,
7164 index: false,
7165 auto_now_add: false,
7166 auto_uuid: false,
7167 auto_now: false,
7168 trim: false,
7169 lowercase: false,
7170 case_insensitive: false,
7171 help: String::new(),
7172 example: String::new(),
7173 widget: None,
7174 supported_backends: Vec::new(),
7175 min: None,
7176 max: None,
7177 text_format: ::core::option::Option::None,
7178 slug_from: ::core::option::Option::None,
7179 };
7180
7181 // unique false → true: emit ADD CONSTRAINT ... UNIQUE
7182 let mut new = baseline.clone();
7183 new.unique = true;
7184 let stmts = render_alter_column_postgres("m", "x", &[new], Some(&[baseline.clone()]));
7185 let joined = stmts.join("\n");
7186 assert!(
7187 joined.contains("ADD CONSTRAINT") && joined.contains("UNIQUE"),
7188 "unique add: expected ADD CONSTRAINT UNIQUE; got: {joined}",
7189 );
7190
7191 // unique true → false: emit DROP CONSTRAINT ... IF EXISTS
7192 let prev_unique = Column {
7193 unique: true,
7194 ..baseline.clone()
7195 };
7196 let stmts =
7197 render_alter_column_postgres("m", "x", &[baseline.clone()], Some(&[prev_unique]));
7198 let joined = stmts.join("\n");
7199 assert!(
7200 joined.contains("DROP CONSTRAINT IF EXISTS"),
7201 "unique drop: expected DROP CONSTRAINT IF EXISTS; got: {joined}",
7202 );
7203
7204 // default empty → "hello": SET DEFAULT 'hello'
7205 let mut new = baseline.clone();
7206 new.default = "hello".into();
7207 let stmts = render_alter_column_postgres("m", "x", &[new], Some(&[baseline.clone()]));
7208 let joined = stmts.join("\n");
7209 assert!(
7210 joined.contains("SET DEFAULT 'hello'"),
7211 "default set: expected SET DEFAULT; got: {joined}",
7212 );
7213
7214 // default "hello" → empty: DROP DEFAULT
7215 let prev_default = Column {
7216 default: "hello".into(),
7217 ..baseline.clone()
7218 };
7219 let stmts =
7220 render_alter_column_postgres("m", "x", &[baseline.clone()], Some(&[prev_default]));
7221 let joined = stmts.join("\n");
7222 assert!(
7223 joined.contains("DROP DEFAULT"),
7224 "default drop: expected DROP DEFAULT; got: {joined}",
7225 );
7226
7227 // FK on_delete change → DROP + readd FK with new clause
7228 let fk_baseline = Column {
7229 ty: SqlType::ForeignKey,
7230 fk_target: Some("other".into()),
7231 ..baseline.clone()
7232 };
7233 let fk_cascade = Column {
7234 on_delete: crate::orm::FkAction::Cascade,
7235 ..fk_baseline.clone()
7236 };
7237 let stmts = render_alter_column_postgres("m", "x", &[fk_cascade], Some(&[fk_baseline]));
7238 let joined = stmts.join("\n");
7239 assert!(
7240 joined.contains("DROP CONSTRAINT IF EXISTS")
7241 && joined.contains("FOREIGN KEY")
7242 && joined.contains("ON DELETE CASCADE"),
7243 "FK cascade add: expected drop+readd with ON DELETE CASCADE; got: {joined}",
7244 );
7245 }
7246
7247 /// IMP-2 from bugs/tests/testBugs.md: a `#[umbral(default = "true")]`
7248 /// on a boolean column used to land as `DEFAULT 'true'` on
7249 /// SQLite, which decode-fails on read (column type is INTEGER,
7250 /// the stored TEXT can't deserialize as `bool`). The SQLite
7251 /// renderer now maps the string to `1` / `0`.
7252 #[test]
7253 fn sqlite_bool_default_translates_to_integer_literal() {
7254 use sea_query::{Alias, SqliteQueryBuilder, Table};
7255
7256 let bool_col = Column {
7257 name: "is_active".into(),
7258 ty: SqlType::Boolean,
7259 primary_key: false,
7260 nullable: false,
7261 fk_target: None,
7262 noform: false,
7263 privileged: false,
7264 private: false,
7265 secret: false,
7266 db_constraint: true,
7267 noedit: false,
7268 auto_user_add: false,
7269 auto_user: false,
7270 is_string_repr: false,
7271 max_length: 0,
7272 choices: vec![],
7273 choice_labels: vec![],
7274 default: "true".into(),
7275 is_multichoice: false,
7276 unique: false,
7277 on_delete: crate::orm::FkAction::NoAction,
7278 on_update: crate::orm::FkAction::NoAction,
7279 index: false,
7280 auto_now_add: false,
7281 auto_uuid: false,
7282 auto_now: false,
7283 trim: false,
7284 lowercase: false,
7285 case_insensitive: false,
7286 help: String::new(),
7287 example: String::new(),
7288 widget: None,
7289 supported_backends: Vec::new(),
7290 min: None,
7291 max: None,
7292 text_format: ::core::option::Option::None,
7293 slug_from: ::core::option::Option::None,
7294 };
7295 let mut stmt = Table::create();
7296 stmt.table(Alias::new("t"));
7297 let mut def = build_column_def_sqlite(&bool_col);
7298 stmt.col(&mut def);
7299 let sql = stmt.to_string(SqliteQueryBuilder);
7300 assert!(
7301 sql.contains("DEFAULT 1") && !sql.contains("DEFAULT 'true'"),
7302 "bool default 'true' on sqlite should render as DEFAULT 1; got: {sql}",
7303 );
7304
7305 // "false" → 0
7306 let mut bool_col_false = bool_col.clone();
7307 bool_col_false.default = "false".into();
7308 let mut stmt = Table::create();
7309 stmt.table(Alias::new("t"));
7310 let mut def = build_column_def_sqlite(&bool_col_false);
7311 stmt.col(&mut def);
7312 let sql = stmt.to_string(SqliteQueryBuilder);
7313 assert!(
7314 sql.contains("DEFAULT 0") && !sql.contains("DEFAULT 'false'"),
7315 "bool default 'false' on sqlite should render as DEFAULT 0; got: {sql}",
7316 );
7317
7318 // Non-bool columns are untouched (text default stays
7319 // single-quoted literal).
7320 let text_col = Column {
7321 name: "label".into(),
7322 ty: SqlType::Text,
7323 default: "hello".into(),
7324 ..bool_col.clone()
7325 };
7326 let mut stmt = Table::create();
7327 stmt.table(Alias::new("t"));
7328 let mut def = build_column_def_sqlite(&text_col);
7329 stmt.col(&mut def);
7330 let sql = stmt.to_string(SqliteQueryBuilder);
7331 assert!(
7332 sql.contains("DEFAULT 'hello'"),
7333 "text default should stay quoted; got: {sql}",
7334 );
7335 }
7336
7337 /// BUG-4 from bugs/tests/testBugs.md: `#[umbral(index)]` lifts
7338 /// to a `CREATE INDEX IF NOT EXISTS idx_<table>_<col>` statement
7339 /// alongside the `CREATE TABLE`. The index is skipped on PK
7340 /// and UNIQUE columns (those are already indexed by the
7341 /// constraint).
7342 #[test]
7343 fn index_attribute_emits_create_index_alongside_create_table() {
7344 let id = Column {
7345 name: "id".into(),
7346 ty: SqlType::BigInt,
7347 primary_key: true,
7348 nullable: false,
7349 fk_target: None,
7350 noform: false,
7351 privileged: false,
7352 private: false,
7353 secret: false,
7354 db_constraint: true,
7355 noedit: false,
7356 auto_user_add: false,
7357 auto_user: false,
7358 is_string_repr: false,
7359 max_length: 0,
7360 choices: vec![],
7361 choice_labels: vec![],
7362 default: String::new(),
7363 is_multichoice: false,
7364 unique: false,
7365 on_delete: crate::orm::FkAction::NoAction,
7366 on_update: crate::orm::FkAction::NoAction,
7367 // PK with index=true; the renderer should skip the
7368 // extra CREATE INDEX because the PK constraint
7369 // already covers it.
7370 index: true,
7371 auto_now_add: false,
7372 auto_uuid: false,
7373 auto_now: false,
7374 trim: false,
7375 lowercase: false,
7376 case_insensitive: false,
7377 help: String::new(),
7378 example: String::new(),
7379 widget: None,
7380 supported_backends: Vec::new(),
7381 min: None,
7382 max: None,
7383 text_format: ::core::option::Option::None,
7384 slug_from: ::core::option::Option::None,
7385 };
7386 let slug = Column {
7387 name: "slug".into(),
7388 ty: SqlType::Text,
7389 primary_key: false,
7390 nullable: false,
7391 index: true,
7392 auto_now_add: false,
7393 auto_uuid: false,
7394 auto_now: false,
7395 trim: false,
7396 lowercase: false,
7397 case_insensitive: false,
7398 help: String::new(),
7399 example: String::new(),
7400 widget: None,
7401 supported_backends: Vec::new(),
7402 ..id.clone()
7403 };
7404 let title = Column {
7405 name: "title".into(),
7406 ty: SqlType::Text,
7407 primary_key: false,
7408 nullable: false,
7409 index: false,
7410 auto_now_add: false,
7411 auto_uuid: false,
7412 auto_now: false,
7413 trim: false,
7414 lowercase: false,
7415 case_insensitive: false,
7416 help: String::new(),
7417 example: String::new(),
7418 widget: None,
7419 supported_backends: Vec::new(),
7420 ..id.clone()
7421 };
7422 let op = Operation::CreateTable {
7423 table: "post".into(),
7424 columns: vec![id, slug, title],
7425 unique_together: Vec::new(),
7426 indexes: Vec::new(),
7427 };
7428
7429 for backend in ["sqlite", "postgres"] {
7430 let stmts = render_operation_for(&op, backend);
7431 assert!(
7432 stmts
7433 .iter()
7434 .any(|s| s.to_uppercase().contains("CREATE TABLE")),
7435 "{backend}: expected a CREATE TABLE; got: {stmts:?}",
7436 );
7437 let index_stmts: Vec<_> = stmts
7438 .iter()
7439 .filter(|s| s.to_uppercase().contains("CREATE INDEX"))
7440 .collect();
7441 assert_eq!(
7442 index_stmts.len(),
7443 1,
7444 "{backend}: expected exactly one CREATE INDEX (on `slug`); got {index_stmts:?}",
7445 );
7446 let ix = index_stmts[0];
7447 assert!(
7448 ix.contains("\"idx_post_slug\"") && ix.contains("(\"slug\")"),
7449 "{backend}: index should target post(slug); got: {ix}",
7450 );
7451 assert!(
7452 ix.to_uppercase().contains("IF NOT EXISTS"),
7453 "{backend}: should be idempotent via IF NOT EXISTS; got: {ix}",
7454 );
7455 }
7456 }
7457
7458 /// gaps3 #35: a `#[umbral(case_insensitive)]` text column renders
7459 /// per-backend — SQLite `COLLATE NOCASE`, Postgres `citext` with an
7460 /// idempotent `CREATE EXTENSION` emitted before the table.
7461 #[test]
7462 fn case_insensitive_column_renders_per_backend() {
7463 let id = Column {
7464 name: "id".into(),
7465 ty: SqlType::BigInt,
7466 primary_key: true,
7467 nullable: false,
7468 fk_target: None,
7469 noform: false,
7470 privileged: false,
7471 private: false,
7472 secret: false,
7473 db_constraint: true,
7474 noedit: false,
7475 auto_user_add: false,
7476 auto_user: false,
7477 is_string_repr: false,
7478 max_length: 0,
7479 choices: vec![],
7480 choice_labels: vec![],
7481 default: String::new(),
7482 is_multichoice: false,
7483 unique: false,
7484 on_delete: crate::orm::FkAction::NoAction,
7485 on_update: crate::orm::FkAction::NoAction,
7486 index: false,
7487 auto_now_add: false,
7488 auto_uuid: false,
7489 auto_now: false,
7490 trim: false,
7491 lowercase: false,
7492 case_insensitive: false,
7493 help: String::new(),
7494 example: String::new(),
7495 widget: None,
7496 supported_backends: Vec::new(),
7497 min: None,
7498 max: None,
7499 text_format: None,
7500 slug_from: None,
7501 };
7502 let name = Column {
7503 name: "name".into(),
7504 ty: SqlType::Text,
7505 primary_key: false,
7506 unique: true,
7507 case_insensitive: true,
7508 ..id.clone()
7509 };
7510 let op = Operation::CreateTable {
7511 table: "handle".into(),
7512 columns: vec![id, name],
7513 unique_together: Vec::new(),
7514 indexes: Vec::new(),
7515 };
7516
7517 // SQLite: COLLATE NOCASE on the column; no extension.
7518 let sqlite = render_operation_for(&op, "sqlite");
7519 let create = sqlite
7520 .iter()
7521 .find(|s| s.to_uppercase().contains("CREATE TABLE"))
7522 .expect("a CREATE TABLE");
7523 assert!(
7524 create.to_uppercase().contains("COLLATE NOCASE"),
7525 "sqlite case_insensitive column must carry COLLATE NOCASE; got: {create}"
7526 );
7527 assert!(
7528 !sqlite
7529 .iter()
7530 .any(|s| s.to_uppercase().contains("EXTENSION")),
7531 "sqlite must not emit a citext extension; got: {sqlite:?}"
7532 );
7533
7534 // Postgres: citext type + CREATE EXTENSION IF NOT EXISTS citext, and the
7535 // extension statement precedes the CREATE TABLE.
7536 let pg = render_operation_for(&op, "postgres");
7537 let ext_idx = pg
7538 .iter()
7539 .position(|s| s.to_uppercase().contains("CREATE EXTENSION") && s.contains("citext"))
7540 .expect("a CREATE EXTENSION citext statement");
7541 let create_idx = pg
7542 .iter()
7543 .position(|s| s.to_uppercase().contains("CREATE TABLE"))
7544 .expect("a CREATE TABLE");
7545 assert!(
7546 ext_idx < create_idx,
7547 "the citext extension must be created before the table; got: {pg:?}"
7548 );
7549 assert!(
7550 pg[create_idx].to_lowercase().contains("citext"),
7551 "postgres case_insensitive column must render as citext; got: {}",
7552 pg[create_idx]
7553 );
7554 assert!(
7555 pg[ext_idx].to_uppercase().contains("IF NOT EXISTS"),
7556 "the extension create must be idempotent; got: {}",
7557 pg[ext_idx]
7558 );
7559 }
7560
7561 /// Regression: adding an `auto_now` / `auto_now_add` column to an
7562 /// existing populated table.
7563 ///
7564 /// - SQLite: a 2-statement sequence (nullable ADD + UPDATE
7565 /// backfill) since SQLite refuses non-constant defaults in
7566 /// ALTER. The column ends up nullable at the DB level;
7567 /// Rust still enforces non-null at the type level.
7568 /// - Postgres: a single ALTER with `DEFAULT now()` — Postgres
7569 /// allows the non-constant default and backfills inline.
7570 #[test]
7571 fn auto_now_add_column_renders_safe_backfill_per_backend() {
7572 for (label, auto_now, auto_now_add) in
7573 [("auto_now", true, false), ("auto_now_add", false, true)]
7574 {
7575 let col = Column {
7576 name: "updated_at".to_string(),
7577 ty: SqlType::Timestamptz,
7578 primary_key: false,
7579 nullable: false,
7580 fk_target: None,
7581 noform: false,
7582 privileged: false,
7583 private: false,
7584 secret: false,
7585 db_constraint: true,
7586 noedit: false,
7587 auto_user_add: false,
7588 auto_user: false,
7589 is_string_repr: false,
7590 max_length: 0,
7591 choices: Vec::new(),
7592 choice_labels: Vec::new(),
7593 default: String::new(),
7594 is_multichoice: false,
7595 unique: false,
7596 on_delete: crate::orm::FkAction::NoAction,
7597 on_update: crate::orm::FkAction::NoAction,
7598 index: false,
7599 auto_now_add,
7600 auto_uuid: false,
7601 auto_now,
7602 trim: false,
7603 lowercase: false,
7604 case_insensitive: false,
7605 help: String::new(),
7606 example: String::new(),
7607 widget: None,
7608 supported_backends: Vec::new(),
7609 min: None,
7610 max: None,
7611 text_format: None,
7612 slug_from: None,
7613 };
7614
7615 // SQLite: the AddColumn op must produce TWO statements:
7616 // an ADD COLUMN nullable + an UPDATE backfill. The ADD
7617 // must NOT carry `NOT NULL` (otherwise SQLite rejects
7618 // it on the populated rows), and must NOT carry a
7619 // DEFAULT (otherwise SQLite rejects the non-constant).
7620 let op = Operation::AddColumn {
7621 table: "customer".to_string(),
7622 column: col.clone(),
7623 };
7624 let stmts = render_operation_sqlite(&op);
7625 assert_eq!(
7626 stmts.len(),
7627 2,
7628 "{label} SQLite: must emit ADD + UPDATE, got: {stmts:?}",
7629 );
7630 let add_sql = stmts[0].to_uppercase();
7631 assert!(
7632 add_sql.contains("ADD COLUMN"),
7633 "{label} SQLite: first stmt must be ADD COLUMN, got: {}",
7634 stmts[0],
7635 );
7636 assert!(
7637 !add_sql.contains("NOT NULL"),
7638 "{label} SQLite: ADD COLUMN must be nullable (NOT NULL = SQLite reject), got: {}",
7639 stmts[0],
7640 );
7641 assert!(
7642 !add_sql.contains("DEFAULT"),
7643 "{label} SQLite: ADD COLUMN must omit DEFAULT (non-constant = SQLite reject), got: {}",
7644 stmts[0],
7645 );
7646 let backfill_sql = &stmts[1];
7647 assert!(
7648 backfill_sql.contains("UPDATE") && backfill_sql.contains("datetime('now')"),
7649 "{label} SQLite: second stmt must be backfill UPDATE, got: {backfill_sql}",
7650 );
7651
7652 // Postgres: single ALTER with NOT NULL + DEFAULT now().
7653 let pstmts = render_operation_postgres(&op);
7654 assert_eq!(
7655 pstmts.len(),
7656 1,
7657 "{label} Postgres: single statement suffices, got: {pstmts:?}",
7658 );
7659 let p = &pstmts[0];
7660 assert!(
7661 p.to_lowercase().contains("default now()"),
7662 "{label} Postgres: expected DEFAULT now() in ALTER, got: {p}",
7663 );
7664 assert!(
7665 p.to_uppercase().contains("NOT NULL"),
7666 "{label} Postgres: keeps NOT NULL (Postgres allows non-constant defaults), got: {p}",
7667 );
7668 }
7669 }
7670
7671 /// Audit core-migrate #14 — raw DDL that interpolates
7672 /// developer-supplied identifiers must escape inner double quotes by
7673 /// doubling them (the quoting idiom used everywhere else), not strip
7674 /// or pass them through verbatim. A `"` in a table name previously
7675 /// produced malformed DDL in the multi-column index helper (ON-clause
7676 /// table was quote-stripped) and the M2M junction DDL (five raw
7677 /// interpolations).
7678 #[test]
7679 fn raw_ddl_escapes_quoted_identifiers() {
7680 // Multi-column index: the ON-clause table reference must carry
7681 // the doubled quote, not a stripped one.
7682 let idx = create_multi_index_stmt("we\"ird", &["a\"b".to_string(), "c".to_string()]);
7683 assert!(
7684 idx.contains("ON \"we\"\"ird\""),
7685 "multi-index ON clause must escape the quote (doubled); got: {idx}",
7686 );
7687 assert!(
7688 idx.contains("\"a\"\"b\""),
7689 "multi-index column list must escape the quote; got: {idx}",
7690 );
7691
7692 // M2M junction DDL: every interpolated identifier escapes its
7693 // inner quote. Check both backends.
7694 let op = Operation::CreateM2MTable {
7695 junction_table: "j\"t".to_string(),
7696 parent_table: "p\"t".to_string(),
7697 parent_col: "p\"c".to_string(),
7698 child_table: "c\"t".to_string(),
7699 child_col: "c\"c".to_string(),
7700 parent_ty: SqlType::BigInt,
7701 child_ty: SqlType::Text,
7702 };
7703 for backend in ["sqlite", "postgres"] {
7704 let sql = render_operation_for(&op, backend).join("\n");
7705 for (raw, escaped) in [
7706 ("j\"t", "\"j\"\"t\""),
7707 ("p\"t", "\"p\"\"t\""),
7708 ("p\"c", "\"p\"\"c\""),
7709 ("c\"t", "\"c\"\"t\""),
7710 ("c\"c", "\"c\"\"c\""),
7711 ] {
7712 assert!(
7713 sql.contains(escaped),
7714 "{backend}: identifier `{raw}` must render escaped as {escaped}; got: {sql}",
7715 );
7716 }
7717 }
7718 }
7719}
7720
7721/// A `#[derive(Model)]` type's self-registration (gaps3 #46).
7722///
7723/// The derive submits one of these into a link-time slice, so
7724/// `AppBuilder::auto_models()` can discover every model the binary links without
7725/// the app naming each one. It stores a *function* rather than a `ModelMeta`
7726/// because `ModelMeta::for_::<T>()` isn't const — the meta is built on demand at
7727/// boot.
7728pub struct ModelRegistration {
7729 /// Builds this model's meta.
7730 pub meta: fn() -> ModelMeta,
7731 /// `module_path!()` at the derive site (gaps4 #40) — lets
7732 /// [`models_registered_in`] filter the slice down to one crate's models,
7733 /// so a plugin's `models()` can be `umbral::discovered_models!()` instead
7734 /// of a hand-maintained `ModelMeta::for_::<T>()` list.
7735 pub module_path: &'static str,
7736}
7737
7738inventory::collect!(ModelRegistration);
7739
7740/// Every model that self-registered via `#[derive(Model)]` in the linked binary.
7741pub fn link_registered_models() -> Vec<ModelMeta> {
7742 inventory::iter::<ModelRegistration>
7743 .into_iter()
7744 .map(|r| (r.meta)())
7745 .collect()
7746}
7747
7748/// The self-registered models whose defining CRATE matches `caller_path`'s
7749/// crate (gaps4 #40) — the engine behind `umbral::discovered_models!()`.
7750///
7751/// Matching is by the first `::` segment of `module_path!()`, i.e. the crate
7752/// name, so models in submodules (`my_plugin::models::Post`) are found from a
7753/// `models()` impl at the crate root. Inherits `auto_models()`' documented
7754/// linker caveat: a model object file nothing references can be dropped and
7755/// would then be missing here — if a table must NEVER silently go missing,
7756/// list it explicitly.
7757pub fn models_registered_in(caller_path: &str) -> Vec<ModelMeta> {
7758 let crate_root = caller_path.split("::").next().unwrap_or(caller_path);
7759 inventory::iter::<ModelRegistration>
7760 .into_iter()
7761 .filter(|r| r.module_path.split("::").next() == Some(crate_root))
7762 .map(|r| (r.meta)())
7763 .collect()
7764}