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