znippy_plugin_git/arms.rs
1//! **Which implementation of each trait a store is built on** — one selector,
2//! read once, never per operation.
3//!
4//! Three traits in this crate have more than one implementation, every one of
5//! them written and benchmarked (PLAN §7, §8):
6//!
7//! | trait | arms | what varies |
8//! |---|---|---|
9//! | [`ArchiveWrite`] | [`WriterArm::Fast`] / [`Safe`](WriterArm::Safe) / [`Uring`](WriterArm::Uring) | **the durability contract**, and nothing else |
10//! | [`ObjectIndex`](crate::index_layout::ObjectIndex) | [`IndexArm::OneTableFourColumns`] / [`FourTables`](IndexArm::FourTables) / [`PackedPayload`](IndexArm::PackedPayload) | the payload layout behind the same `stree` |
11//! | [`Gc`] | [`GcArm::NewGeneration`] / [`CompactInPlace`](GcArm::CompactInPlace) | whether the old generation survives until the new one is proven |
12//!
13//! Until this module existed, `GitStore` named one of each by hand, so none of
14//! them could be A/B'd through a server or a bench without editing the source.
15//! [`StoreConfig`] is that choice made data.
16//!
17//! # The default is the shipping combination and it did not move
18//!
19//! [`StoreConfig::DEFAULT`] is `SafeWriter` + `ObjectReadStack<OneTableFourColumns>`
20//! + `NewGeneration`, which is exactly what
21//! [`GitStore::open`](crate::git_ops::GitStore::open) built before this module
22//! and is exactly what it builds now.
23//! **`open` and `open_with` do not read the environment at all** — an operator
24//! who exports a variable cannot silently change the durability contract under a
25//! caller that never asked for a selector. Choosing an arm is something a caller
26//! does on purpose, through
27//! [`open_with_arms`](crate::git_ops::GitStore::open_with_arms) or
28//! [`open_from_env`](crate::git_ops::open_from_env).
29//!
30//! # Read once, at construction. Never per operation.
31//!
32//! Every `getenv` this crate performs goes through [`read_env`] and happens
33//! either inside [`StoreConfig::from_env`] (which a store calls **once**, while
34//! it is being opened — the three arms and, since 2026-08-21, the cache ceiling
35//! with them) or behind a process-wide `OnceLock` for the knobs that never vary
36//! between two stores in one process. A `std::env::var` in a copy loop would be
37//! one syscall per object served, and that exact defect was found and fixed in
38//! gunnar the day before this was written. It is not a style rule here, it is
39//! asserted: [`env_reads_here`] counts every read this module makes on the
40//! calling thread, and
41//! `the_selector_is_read_once_at_construction_and_never_per_operation` pushes
42//! and serves a whole pack across an unchanged counter.
43//!
44//! Five reads at construction — three arms, the cache ceiling and the explode
45//! policy — and zero per operation. [`env_reads`] is the same count
46//! process-wide; see [`env_reads_here`] for why the guard asserts on the
47//! thread-local one. The complete roster of keys is [`ALL_ENV`].
48//!
49//! # The names, so an operator can type them
50//!
51//! ```text
52//! ZNIPPY_GIT_WRITER = fast | safe | uring (default: safe)
53//! ZNIPPY_GIT_INDEX = one-table | four-tables | packed (default: one-table)
54//! ZNIPPY_GIT_GC = new-generation | in-place (default: new-generation)
55//!
56//! ZNIPPY_GIT_REDB_CACHE_BYTES = <bytes> (default: 67108864)
57//!
58//! ZNIPPY_GIT_EXPLODE = off | graph | full (default: see exploded_arrow.rs)
59//! ZNIPPY_GIT_BOUNDARY_DELTA = 0 | off | false to disable (default: on)
60//! ZNIPPY_GIT_REACH_COMMITS = <commits> (default: 512)
61//! ZNIPPY_GIT_EMIT_WORKERS = <workers>, 0 = all cores (default: 4)
62//! ```
63//!
64//! The first four are per store and are what [`StoreConfig`] holds and prints;
65//! the last four are per process. [`ALL_ENV`] is all thirteen (eight + the five memory/batch knobs of 2026-08-30), and a consumer that
66//! has to carry any of them across a boundary asserts its list against it.
67//!
68//! An unset variable takes the default. A variable set to something else is an
69//! **error**, named and listing what is accepted — a typo that silently fell
70//! back to the default would make an operator believe a measurement came from an
71//! arm that never ran, which is worse than a failed open.
72//!
73//! ⚠ The fourth one is **not an arm**, and it is the only variable here that
74//! [`GitStore::open`](crate::git_ops::GitStore::open) reads. It selects no
75//! implementation and changes no contract — it is a memory ceiling, and redb's
76//! own default for it is 1 GiB *per database*. See [`redb_cache_bytes`].
77
78use std::path::{Path, PathBuf};
79use std::sync::atomic::{AtomicU64, Ordering};
80
81use anyhow::{Result, bail};
82
83use crate::archive_write::{ArchiveWrite, FastWriter, SafeWriter};
84use crate::gc::{CompactInPlace, Gc, NewGeneration};
85
86/// `getenv` calls this module has made since the process started.
87///
88/// The counter behind the read-once law. Every path that reads the environment
89/// in this crate goes through [`read_env`], which bumps it; nothing else does.
90/// A caller — or a guard — can therefore prove that serving N objects cost zero
91/// environment reads, which is the only way to state the law as an assertion
92/// rather than as a comment.
93pub fn env_reads() -> u64 {
94 ENV_READS.load(Ordering::Relaxed)
95}
96
97/// The same count, **for the calling thread only**.
98///
99/// [`env_reads`] is process-wide, which makes it the right number to report and
100/// the wrong number to *assert on*: any other thread opening a store moves it,
101/// so a guard written against it is a race dressed as a test. It was only ever
102/// accidentally deterministic — until 2026-08-10 the sole reader was
103/// [`StoreConfig::from_env`], every caller of which took `ENV_LOCK`, so the
104/// tests that counted were the only tests that counted. That stopped being true
105/// when [`redb_cache_bytes`] joined it: **every** store open reads the
106/// environment now, including [`GitStore::open`](crate::git_ops::GitStore::open),
107/// and most of the suite opens stores without the lock.
108///
109/// So the law is asserted per thread, where a call path actually lives. It is
110/// also the stronger statement: "this thread served 2687 objects and touched the
111/// environment zero times" is what the law says, and it is now true regardless
112/// of what else the process is doing.
113pub fn env_reads_here() -> u64 {
114 ENV_READS_HERE.with(|c| c.get())
115}
116
117static ENV_READS: AtomicU64 = AtomicU64::new(0);
118
119thread_local! {
120 static ENV_READS_HERE: std::cell::Cell<u64> = const { std::cell::Cell::new(0) };
121}
122
123/// The one place this crate touches the environment.
124///
125/// `pub(crate)` so the per-process knobs in `delta.rs`, `exploded_arrow.rs`
126/// and `git_ops.rs` read through the same counted door rather than through
127/// their own `std::env::var` — every read this crate makes moves
128/// [`env_reads`], or the read-once law is a comment.
129pub(crate) fn read_env(key: &str) -> Option<String> {
130 ENV_READS.fetch_add(1, Ordering::Relaxed);
131 ENV_READS_HERE.with(|c| c.set(c.get() + 1));
132 match std::env::var(key) {
133 Ok(v) if !v.trim().is_empty() => Some(v.trim().to_string()),
134 _ => None,
135 }
136}
137
138// ── the writer arm ────────────────────────────────────────────────────────────
139
140/// Which [`ArchiveWrite`] the push path appends through.
141///
142/// **The arm changes what `append` promises and nothing else.** All three write
143/// the pushed pack's bytes verbatim, at the same offset, into the same file; a
144/// store built on any of them holds byte-identical payload. What differs is how
145/// much of the write is durable when `append` returns, and that is the whole
146/// axis.
147#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
148pub enum WriterArm {
149 /// [`FastWriter`] — **one `pwrite`, no fsync, no journal.**
150 ///
151 /// `append` returns as soon as the bytes are in the kernel's page cache.
152 /// Stated plainly, because an operator selecting this is choosing it:
153 ///
154 /// * **A machine crash (power loss, panic, a hard reset) after `append`
155 /// returns loses the bytes.** Not "may lose" — nothing has told the device
156 /// about them.
157 /// * There is **no journal**, so a store on this arm has no durable record
158 /// that a pack was ever acked. It therefore has **no crash recovery**: a
159 /// reopen cannot re-queue an interrupted pack because there is nothing to
160 /// diff the index against, and pack ordinals restart at 0
161 /// (`indexer::packs_already_acked`).
162 /// * A *process* crash alone keeps the bytes: page cache survives `exit`.
163 /// That is the only crash it survives.
164 ///
165 /// It is here because it is the **ceiling the other two are measured
166 /// against** — at 8 KiB it acks in 3.9 µs against `SafeWriter`'s 132 µs
167 /// (`examples/push_path_bench.rs`, oden 2026-08-07) — and because there are
168 /// real workloads whose durability contract is not git's: a rebuildable
169 /// mirror, a bulk import that is re-run on failure, a benchmark. Selecting
170 /// it is not refused and not warned about. It is the operator's call and
171 /// the contract above is what they are choosing.
172 Fast,
173 /// [`SafeWriter`] — blob `fsync`, **then** the journal row, then the
174 /// journal's `fsync`. **The default, and git's contract**: a push that was
175 /// acked survives a crash. Four syscalls per append.
176 #[default]
177 Safe,
178 /// [`UringWriter`](crate::uring_write::UringWriter) — the **same ordering**
179 /// as [`Safe`](WriterArm::Safe), enforced by the kernel through an
180 /// `IOSQE_IO_LINK` chain instead of by the caller blocking between four
181 /// syscalls. One `io_uring_enter` carrying four linked ops.
182 ///
183 /// MEASURED 2026-08-07 (PLAN §8.3): **it did not win** — within noise of
184 /// `SafeWriter` at every pack size, because the cost is the two device
185 /// flushes and not the syscall count. Kept as a selectable arm precisely so
186 /// that finding can be re-run rather than remembered. Linux only.
187 ///
188 /// **What this arm costs that the other two do not: pinned memory, and a
189 /// ceiling on how many stores a process can hold.** Its ring registers a
190 /// staging buffer, and `IORING_REGISTER_BUFFERS` pins pages against
191 /// `RLIMIT_MEMLOCK` — a limit the kernel counts on the `user_struct`, so it
192 /// is shared by every process the uid is running. One writer per store means
193 /// the number of repositories this arm can serve is
194 /// `RLIMIT_MEMLOCK / page`, and `Fast` and `Safe` have no such ceiling
195 /// because they register nothing. It was **~107 stores** on a stock 8 MiB
196 /// limit until 2026-08-14, when the registration was cut from 64 KiB to the
197 /// one page a journal row actually needs; see
198 /// [`uring_write`](crate::uring_write)'s module docs for the arithmetic, the
199 /// measurement and the failure it produced in gunnar's sweep.
200 ///
201 /// It used to differ from `Safe` in one more way, and that is now closed:
202 /// [`UringWriter::create`](crate::uring_write::UringWriter::create) opened
203 /// its journal with `File::create`, so a **reopen truncated the journal**
204 /// where `SafeWriter` appends to it — the durability of `Safe` within a
205 /// process and the crash recovery of `Fast` across one. Both durable arms
206 /// now open the same log the same way, which is what "one journal format,
207 /// two transports" (LAW 5) always claimed of them.
208 Uring,
209}
210
211impl WriterArm {
212 /// Every arm, for a bench that sweeps them.
213 pub const ALL: [WriterArm; 3] = [WriterArm::Fast, WriterArm::Safe, WriterArm::Uring];
214
215 /// The spelling an operator types.
216 pub fn as_str(self) -> &'static str {
217 match self {
218 WriterArm::Fast => "fast",
219 WriterArm::Safe => "safe",
220 WriterArm::Uring => "uring",
221 }
222 }
223
224 /// One line on what this arm's `append` promises. Same string the trait's
225 /// own [`ArchiveWrite::durability`] returns, so a bench row and a config
226 /// dump cannot disagree.
227 pub fn durability(self) -> &'static str {
228 match self {
229 WriterArm::Fast => {
230 "none — page cache only; a machine crash after return loses the bytes, and there \
231 is no journal, so no crash recovery and no stable pack ordinals"
232 }
233 WriterArm::Safe | WriterArm::Uring => {
234 "full — blob fsynced, then a journal row fsynced; crash after return keeps both"
235 }
236 }
237 }
238
239 /// Whether this arm keeps the durable extent log that
240 /// [`GitStore::open_with_arms`](crate::git_ops::GitStore::open_with_arms)
241 /// derives §13.12's `indexed` bit from.
242 ///
243 /// `None` is [`Fast`](WriterArm::Fast) and it is load-bearing rather than
244 /// cosmetic: with no journal there is no durable record of an ack, so the
245 /// crash-recovery diff has nothing to run against and a reopen re-queues
246 /// nothing. Returning the path of a journal this arm does not write would
247 /// make a reopened store diff against a **stale** log and re-absorb packs
248 /// that a different arm acked.
249 pub fn journal(self, blobs: &Path) -> Option<PathBuf> {
250 match self {
251 WriterArm::Fast => None,
252 WriterArm::Safe | WriterArm::Uring => Some(crate::archive_write::journal_path(blobs)),
253 }
254 }
255
256 /// Build the writer. `blobs` is the file the pushed packs are appended to.
257 pub fn create(self, blobs: &Path) -> Result<Box<dyn ArchiveWrite>> {
258 Ok(match self {
259 WriterArm::Fast => Box::new(FastWriter::create(blobs)?),
260 WriterArm::Safe => Box::new(SafeWriter::create(blobs)?),
261 #[cfg(target_os = "linux")]
262 WriterArm::Uring => Box::new(crate::uring_write::UringWriter::create(blobs)?),
263 #[cfg(not(target_os = "linux"))]
264 WriterArm::Uring => bail!(
265 "the `uring` writer arm is Linux-only — this target has no io_uring, and this \
266 crate does not substitute a pwrite path under an io_uring name"
267 ),
268 })
269 }
270
271 /// Parse the spelling an operator types. Errors name what is accepted.
272 pub fn parse(s: &str) -> Result<Self> {
273 Ok(match s.trim().to_ascii_lowercase().as_str() {
274 "fast" | "fastwriter" => WriterArm::Fast,
275 "safe" | "safewriter" => WriterArm::Safe,
276 "uring" | "uringwriter" | "io_uring" => WriterArm::Uring,
277 other => bail!(
278 "'{other}' is not a writer arm — expected one of fast, safe, uring \
279 ({}={other})",
280 ENV_WRITER
281 ),
282 })
283 }
284}
285
286// ── the index arm ─────────────────────────────────────────────────────────────
287
288/// Which payload layout sits behind the read stack's `stree`.
289///
290/// **This one is a *type*, not a value**, and that is deliberate: it is the hot
291/// path. `GitStore` is generic over it
292/// (`GitStore<S = OneTableFourColumns>`), so a caller that knows its arm at
293/// compile time — which includes every existing caller, through the default —
294/// pays no dispatch at all. This enum exists so a *runtime* selector can pick
295/// one; see [`crate::git_ops::open_from_env`], which monomorphises all three and
296/// hands back a `Box<dyn GitOps>`.
297///
298/// MEASURED 2026-08-07 (PLAN §13): a packed 25-byte column beats four columns by
299/// 19–76% on the payload gather but loses column scans by 3.3x–14.4x, and no git
300/// operation does a full-row fetch. That is why `OneTableFourColumns` is the
301/// default — and why the losing arms stay selectable, because that conclusion is
302/// a measurement and measurements get re-run.
303#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
304pub enum IndexArm {
305 /// [`OneTableFourColumns`](crate::index_layout::OneTableFourColumns) — the
306 /// default. One Arrow IPC section, four columns.
307 #[default]
308 OneTableFourColumns,
309 /// [`FourTables`](crate::index_layout::FourTables) — four independent IPC
310 /// sections joined on the ordinal.
311 FourTables,
312 /// [`PackedPayload`](crate::index_layout::PackedPayload) — one 25-byte
313 /// fixed-size-binary column holding the whole row.
314 PackedPayload,
315}
316
317impl IndexArm {
318 pub const ALL: [IndexArm; 3] = [
319 IndexArm::OneTableFourColumns,
320 IndexArm::FourTables,
321 IndexArm::PackedPayload,
322 ];
323
324 /// The spelling an operator types.
325 pub fn as_str(self) -> &'static str {
326 match self {
327 IndexArm::OneTableFourColumns => "one-table",
328 IndexArm::FourTables => "four-tables",
329 IndexArm::PackedPayload => "packed",
330 }
331 }
332
333 /// The name [`ObjectIndex::name`](crate::index_layout::ObjectIndex::name)
334 /// reports for the projection this arm builds. Distinct from
335 /// [`as_str`](IndexArm::as_str) on purpose: one is what an operator types,
336 /// the other is what the built object calls itself, and a guard that
337 /// compares them is comparing the selector against applied output.
338 pub fn projection_name(self) -> &'static str {
339 match self {
340 IndexArm::OneTableFourColumns => "OneTableFourColumns",
341 IndexArm::FourTables => "FourTables",
342 IndexArm::PackedPayload => "PackedPayload",
343 }
344 }
345
346 pub fn parse(s: &str) -> Result<Self> {
347 Ok(match s.trim().to_ascii_lowercase().as_str() {
348 "one-table" | "one_table" | "onetablefourcolumns" | "one" => {
349 IndexArm::OneTableFourColumns
350 }
351 "four-tables" | "four_tables" | "fourtables" | "four" => IndexArm::FourTables,
352 "packed" | "packedpayload" => IndexArm::PackedPayload,
353 other => bail!(
354 "'{other}' is not an index arm — expected one of one-table, four-tables, packed \
355 ({ENV_INDEX}={other})"
356 ),
357 })
358 }
359}
360
361// ── the gc arm ────────────────────────────────────────────────────────────────
362
363/// Which [`Gc`] [`GitOps::gc`](crate::git_ops::GitOps::gc) runs as its third
364/// step, after reachability and after the dead index rows are dropped.
365#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
366pub enum GcArm {
367 /// [`NewGeneration`] — the default. Hard-link, compact the link, **verify
368 /// the result by reading every entry back**, rename to `x.gN.znippy`,
369 /// unlink the old generation last. A death at any point leaves at least one
370 /// complete readable archive.
371 #[default]
372 NewGeneration,
373 /// [`CompactInPlace`] — base znippy's `compact_archive` against the archive
374 /// under its own name: stage beside it, rename over it. Verification is
375 /// **post-hoc**: by the time there is something to read back, the original
376 /// is already gone. One archive on disk instead of two, and no generation
377 /// suffix — which is the reason to pick it — at the price of that window.
378 CompactInPlace,
379}
380
381impl GcArm {
382 pub const ALL: [GcArm; 2] = [GcArm::NewGeneration, GcArm::CompactInPlace];
383
384 pub fn as_str(self) -> &'static str {
385 match self {
386 GcArm::NewGeneration => "new-generation",
387 GcArm::CompactInPlace => "in-place",
388 }
389 }
390
391 /// The name the built implementation reports through [`Gc::name`], which is
392 /// also what lands in [`GcReport::strategy`](crate::gc::GcReport::strategy)
393 /// — the applied output a guard reads.
394 pub fn strategy(self) -> &'static str {
395 match self {
396 GcArm::NewGeneration => "NewGeneration",
397 GcArm::CompactInPlace => "CompactInPlace",
398 }
399 }
400
401 /// Build it. `Send + Sync` because a `GitStore` is shared across the
402 /// indexer's threads.
403 pub fn create(self) -> Box<dyn Gc + Send + Sync> {
404 match self {
405 GcArm::NewGeneration => Box::new(NewGeneration::new()),
406 GcArm::CompactInPlace => Box::new(CompactInPlace::new()),
407 }
408 }
409
410 pub fn parse(s: &str) -> Result<Self> {
411 Ok(match s.trim().to_ascii_lowercase().as_str() {
412 "new-generation" | "new_generation" | "newgeneration" | "new" | "generation" => {
413 GcArm::NewGeneration
414 }
415 "in-place" | "in_place" | "inplace" | "compactinplace" | "compact" => {
416 GcArm::CompactInPlace
417 }
418 other => bail!(
419 "'{other}' is not a gc arm — expected one of new-generation, in-place \
420 ({ENV_GC}={other})"
421 ),
422 })
423 }
424}
425
426// ── the config ────────────────────────────────────────────────────────────────
427
428/// Environment variable naming the [`ArchiveWrite`] arm.
429pub const ENV_WRITER: &str = "ZNIPPY_GIT_WRITER";
430/// Environment variable naming the
431/// [`ObjectIndex`](crate::index_layout::ObjectIndex) arm.
432pub const ENV_INDEX: &str = "ZNIPPY_GIT_INDEX";
433/// Environment variable naming the [`Gc`] arm.
434pub const ENV_GC: &str = "ZNIPPY_GIT_GC";
435/// Environment variable bounding **redb's page cache**, in bytes. See
436/// [`redb_cache_bytes`].
437pub const ENV_REDB_CACHE: &str = "ZNIPPY_GIT_REDB_CACHE_BYTES";
438/// Environment variable switching the **boundary re-delta** off for an A/B
439/// (`0` / `off` / `false`; anything else, or unset, is on). Read once per
440/// process by [`crate::delta::enabled`].
441pub const ENV_BOUNDARY_DELTA: &str = "ZNIPPY_GIT_BOUNDARY_DELTA";
442/// Environment variable naming the **explode policy** of the
443/// `objects.exploded` table (`off` · `graph` · `full`). Read once per store
444/// open by `ExplodePolicy::from_env`.
445pub const ENV_EXPLODE: &str = "ZNIPPY_GIT_EXPLODE";
446/// Environment variable setting the **explode cache budget** — the byte cap on
447/// the `objects.exploded` pools, `0`/`unlimited` for none. Read once per store
448/// open by `crate::exploded_cache::budget_from_env`, which supplies the
449/// filesystem-derived default when it is unset. The tier it bounds MEASURED
450/// 141.7 GiB for an 8.12 GiB pack before the cap existed.
451pub const ENV_EXPLODE_BUDGET: &str = "ZNIPPY_GIT_EXPLODE_BUDGET";
452/// Environment variable switching the **persisted commit-graph cache**
453/// (`objects.graph`) off: `off` / `0` / `false` neither reads nor writes the file,
454/// and removes one left by an earlier open; `on` / `1` / `true`, or unset, is the
455/// default (read when its token matches, re-persist after a fold). Read once per
456/// store open by [`graph_cache_from_env`]. For a store that serves no reads — a
457/// replica that only holds packs (gunnar's twin) — where the file is a derived
458/// memo that must not exist on its disk.
459pub const ENV_GRAPH_CACHE: &str = "ZNIPPY_GIT_GRAPH_CACHE";
460
461/// **Is the `objects.graph` memo on for this store?** [`ENV_GRAPH_CACHE`], through
462/// the counted door. An unparsable value is an error, never a silent default.
463pub fn graph_cache_from_env() -> Result<bool> {
464 match read_env(ENV_GRAPH_CACHE) {
465 None => Ok(GRAPH_CACHE_DEFAULT.load(Ordering::Relaxed)),
466 Some(v) => parse_graph_cache(&v),
467 }
468}
469
470/// What [`graph_cache_from_env`] answers when [`ENV_GRAPH_CACHE`] is unset. `true` unless an
471/// embedder said otherwise.
472static GRAPH_CACHE_DEFAULT: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(true);
473
474/// **Set the process-wide default of the graph cache** for every store opened AFTER this call
475/// (an explicit [`ENV_GRAPH_CACHE`] still wins). For an embedder that learns its role at runtime
476/// and cannot set the environment — edition 2024 makes `set_var` unsafe, and gunnar forbids
477/// `unsafe` — such as gunnar's twin, which holds packs and serves no reads.
478///
479/// **It has four siblings** (0.1.5), one per remaining token such an embedder needs:
480/// [`crate::set_explode_default`], [`crate::set_explode_budget_default`],
481/// [`crate::set_resolve_workers_default`] and [`crate::set_resolve_lookahead_default`]. All
482/// five keep this same precedence rule, and `tests/twin_runtime_limits_env_wins.rs` asserts it
483/// of all five at once so the rule cannot drift apart between them.
484pub fn set_graph_cache_default(on: bool) {
485 GRAPH_CACHE_DEFAULT.store(on, Ordering::Relaxed);
486}
487
488fn parse_graph_cache(v: &str) -> Result<bool> {
489 match v.to_ascii_lowercase().as_str() {
490 "on" | "1" | "true" => Ok(true),
491 "off" | "0" | "false" => Ok(false),
492 other => Err(anyhow::anyhow!(
493 "{ENV_GRAPH_CACHE}={other:?}: expected on/off (1/0, true/false)"
494 )),
495 }
496}
497/// Environment variable capping the **live reachability walk** in commits. An
498/// operator knob, not an arm — see `live_reach_policy` in `git_ops.rs`. Read
499/// once per process.
500pub const ENV_REACH_COMMITS: &str = "ZNIPPY_GIT_REACH_COMMITS";
501/// Environment variable setting the **phase-1 emit workers per request**. An
502/// operator knob, not an arm — see `emit_workers` in `git_ops.rs`. Read once
503/// per process.
504pub const ENV_EMIT_WORKERS: &str = "ZNIPPY_GIT_EMIT_WORKERS";
505/// Environment variable setting the **memory watchdog's floor** as a percent
506/// of `MemTotal` (default 20): below it the indexer flushes and waits. See
507/// [`crate::memguard`]. Read once per process.
508pub const ENV_MEM_BUDGET_PCT: &str = "ZNIPPY_GIT_MEM_BUDGET_PCT";
509/// Environment variable setting the same floor in **absolute bytes**, which
510/// wins over the percent when both are set. Read once per process.
511pub const ENV_MEM_BUDGET_BYTES: &str = "ZNIPPY_GIT_MEM_BUDGET_BYTES";
512/// Environment variable naming the file the watchdog reads instead of
513/// `/proc/meminfo` — a test's fake. Read once per process.
514pub const ENV_MEMINFO: &str = "ZNIPPY_GIT_MEMINFO";
515/// Environment variable bounding the **held delta-base set** of the resolver in
516/// bytes (default 25 % of `MemTotal`): over it the base whose next use is
517/// farthest is dropped and re-inflated from the mmapped pack when needed. See
518/// [`crate::resolve`]. Read once per process.
519pub const ENV_BASE_BUDGET_BYTES: &str = "ZNIPPY_GIT_BASE_BUDGET_BYTES";
520/// Environment variable setting how many **index rows are flushed per batch**
521/// by the absorber (default 262 144) instead of the whole pack's rows at once.
522/// Read once per process.
523pub const ENV_INDEX_BATCH_ROWS: &str = "ZNIPPY_GIT_INDEX_BATCH_ROWS";
524/// Environment variable setting the **absolute resume margin** in bytes
525/// (default 64 MiB): a pause ends once `MemAvailable` is the floor plus this.
526/// It used to be `budget / 8`, which on a 16 GB box demanded 410 MB of headroom
527/// to leave a pause that a 75 MB shortfall opened — the livelock lane P2
528/// measured. See [`crate::memguard`]. Read once per process.
529pub const ENV_MEM_RESUME_MARGIN_BYTES: &str = "ZNIPPY_GIT_MEM_RESUME_MARGIN_BYTES";
530/// Environment variable setting **how many subtrees the resolver walks at
531/// once** (default: one per core). `1` is the serial resolver, and the stream
532/// out is the same at every value. See [`crate::resolve::workers_from_env`].
533/// Read once per process.
534pub const ENV_RESOLVE_WORKERS: &str = "ZNIPPY_GIT_RESOLVE_WORKERS";
535/// Environment variable bounding **how far the resolver's workers may run
536/// ahead of its ordered output**, in bytes (default 128 MiB). Read once per
537/// process.
538pub const ENV_RESOLVE_LOOKAHEAD_BYTES: &str = "ZNIPPY_GIT_RESOLVE_LOOKAHEAD_BYTES";
539/// Environment variable bounding **how long one pause may last**, in
540/// milliseconds (default 300 000 = 5 minutes). Past it the worker stops
541/// waiting and the pack is failed BY NAME rather than parked forever. See
542/// [`crate::memguard::PauseEnd`]. Read once per process.
543pub const ENV_MEM_PAUSE_MAX_MS: &str = "ZNIPPY_GIT_MEM_PAUSE_MAX_MS";
544
545/// **Every environment variable this crate reads outside its tests**, in one
546/// place, so a consumer that carries them across a process or container
547/// boundary can assert its list against this one instead of against a count
548/// it remembers.
549///
550/// # Why a list, and why here
551///
552/// gunnar's forge runs `gunnar serve` inside a podman container, and podman
553/// passes nothing implicitly: a variable set on the host reaches the server
554/// only if the forge names it in its crossing list. On 2026-08-11 five separate
555/// wrong conclusions were drawn on one box from knobs that were "set" and never
556/// arrived — the server ran on its default and the probe reported "no effect",
557/// a null result shaped exactly like a real one. The forge's own list was three
558/// entries long against a crate that read four keys, and **nothing went red
559/// when this crate grew a key**, because the only list that looked was the
560/// consumer's.
561///
562/// This is the producer's list. It is the one a consumer asserts against — "my
563/// crossing list plus my named exclusions equals `ALL_ENV`" — so the day this
564/// crate grows a ninth key, the consumer's guard is what says so, and the
565/// failure names the key rather than producing a quiet default.
566///
567/// # It is asserted complete, not trusted
568///
569/// `every_environment_read_in_this_crate_is_named_in_all_env` (in this module's
570/// tests) walks the crate's own source and checks that every `std::env::var`
571/// call outside a `#[cfg(test)]` region names a key on this list — by literal
572/// or by one of the `ENV_*` constants above. Grow a read without growing the
573/// list and that test is what goes red.
574///
575/// Three of the thirteen are **arms** (they select an implementation and change
576/// what the store promises); the other five are knobs and ceilings that select
577/// nothing. [`StoreConfig`] holds the arms and the cache ceiling, and its
578/// `Display` prints those four; the remaining four are per-process knobs that
579/// never vary between two stores in one process.
580pub const ALL_ENV: &[&str] = &[
581 ENV_WRITER,
582 ENV_INDEX,
583 ENV_GC,
584 ENV_REDB_CACHE,
585 ENV_BOUNDARY_DELTA,
586 ENV_EXPLODE,
587 ENV_EXPLODE_BUDGET,
588 ENV_GRAPH_CACHE,
589 ENV_REACH_COMMITS,
590 ENV_EMIT_WORKERS,
591 ENV_MEM_BUDGET_PCT,
592 ENV_MEM_BUDGET_BYTES,
593 ENV_MEMINFO,
594 ENV_BASE_BUDGET_BYTES,
595 ENV_INDEX_BATCH_ROWS,
596 ENV_MEM_RESUME_MARGIN_BYTES,
597 ENV_MEM_PAUSE_MAX_MS,
598 ENV_RESOLVE_WORKERS,
599 ENV_RESOLVE_LOOKAHEAD_BYTES,
600];
601
602/// **The page-cache ceiling for one repository's redb database**, in bytes.
603///
604/// # Why this exists at all
605///
606/// `redb::Database::create(path)` is `Builder::new().create(path)`, and
607/// `Builder::new` ends with `set_cache_size(1024 * 1024 * 1024)` — redb-2.6.3
608/// `db.rs:1140`. **One GiB, per database, by default**, split 90 % read /
609/// 10 % write by `set_cache_size` (`db.rs:1184`), so the stock ceiling is 1 GiB
610/// *per repository* and a long-lived server holding N repositories has N times
611/// that.
612///
613/// # "Two databases per repository" was stale, and it cost a wrong diagnosis
614///
615/// This paragraph said "a store opens two of them, so the stock ceiling is
616/// 2 GiB per repository" until 2026-08-31. **A store opens one.**
617/// [`GitStore::open_with_arms`](crate::git_ops::GitStore::open_with_arms)
618/// constructs exactly one [`ObjectReadStack`](crate::read_stack::ObjectReadStack)
619/// — `objects.tail` — and `tests/fd_per_store.rs` counts exactly one
620/// `objects.tail` descriptor per store and asserts the total is three. The
621/// second database was the exploded table's, and it stopped being a redb when
622/// §14 moved that tier to append-only Arrow IPC
623/// ([`crate::exploded_arrow`]); the sentence outlived it.
624///
625/// It is recorded rather than quietly deleted because a stale doc is a second
626/// truth and this one was believed: `gunnar/.nornir/stinky.md` S-079 doubled
627/// this figure to 128 MiB per store on the strength of it, and built an
628/// explanation of a live appliance's 10 G on top. See S-082 for the
629/// measurement that refuted it.
630///
631/// It is a ceiling and not a reservation, so a small repository never noticed.
632/// MEASURED 2026-08-10 (t14s, `h2h-linear-sha1-2048c-1024f-16k`, a 1 078 472 704-byte
633/// `objects.exploded`): a single clone parked **594 MB resident** and kept it for
634/// the life of the process, because the pages a full-file traversal touched all
635/// fit under the ceiling and nothing evicted them.
636///
637/// # How the default was chosen
638///
639/// A B-tree read cache earns its keep on the **interior** nodes, which every
640/// lookup re-touches, and earns nothing on a single sequential pass over the
641/// leaves, which is what a `refold` is. The interior levels of a 4 KiB-page
642/// B-tree over that 1.078 GB table are ~7 MB; 64 MiB holds all of them nine
643/// times over and leaves 6.4 MiB of write cache, which is more than one absorb
644/// batch dirties. Measured against the 1 GiB default on the clone path it costs
645/// nothing and returns most of the resident set — see the sweep in
646/// `agentAA-report.md`.
647///
648/// # This is the one environment read that is not an arm
649///
650/// Every other variable in this module selects an *implementation* and therefore
651/// changes what the store promises, which is why
652/// [`GitStore::open`](crate::git_ops::GitStore::open) deliberately reads none of
653/// them. This one selects nothing: it is a memory ceiling, every arm behaves
654/// identically under any value of it, and the bytes on disk are the same either
655/// way. So it *is* read on the `open` path — an operator who has to cap a
656/// server's footprint cannot be told to use a different constructor.
657///
658/// An unparseable or zero value is an **error**, for the same reason a
659/// misspelled arm is: a silent fallback would let an operator believe a
660/// measurement came from a ceiling that was never applied.
661pub fn redb_cache_bytes() -> Result<usize> {
662 let Some(raw) = read_env(ENV_REDB_CACHE) else {
663 return Ok(DEFAULT_REDB_CACHE_BYTES);
664 };
665 match raw.parse::<usize>() {
666 Ok(0) | Err(_) => bail!(
667 "{ENV_REDB_CACHE}={raw:?} is not a positive byte count; it bounds redb's page \
668 cache per database and the default is {DEFAULT_REDB_CACHE_BYTES}"
669 ),
670 Ok(n) => Ok(n),
671 }
672}
673
674/// 64 MiB. See [`redb_cache_bytes`] for how that number was arrived at.
675pub const DEFAULT_REDB_CACHE_BYTES: usize = 64 * 1024 * 1024;
676
677/// One implementation of each trait, chosen.
678///
679/// A plain `Copy` value with no interior state: it is *read* once, at
680/// construction, and from then on the store holds the built objects rather than
681/// this. Handing it around after that is a description of what was built, not a
682/// switch anything consults.
683#[derive(Debug, Clone, Copy, PartialEq, Eq)]
684pub struct StoreConfig {
685 pub writer: WriterArm,
686 pub index: IndexArm,
687 pub gc: GcArm,
688 /// The redb page-cache ceiling the store's `objects.tail` database is opened
689 /// under. One database per store — see [`redb_cache_bytes`].
690 /// **Not an arm** — it selects nothing — but it is part of what was built,
691 /// and a `Display` of the config that left it out would have hidden the
692 /// one key that was "falsified" on 2026-08-11 by never reaching the
693 /// container at all. [`StoreConfig::from_env`] fills it from
694 /// [`redb_cache_bytes`]; [`DEFAULT`](Self::DEFAULT) is
695 /// [`DEFAULT_REDB_CACHE_BYTES`].
696 pub redb_cache_bytes: usize,
697}
698
699impl Default for StoreConfig {
700 /// [`StoreConfig::DEFAULT`], spelled out. A derived `Default` would put a
701 /// zero-byte page cache on the shipping combination.
702 fn default() -> Self {
703 Self::DEFAULT
704 }
705}
706
707impl StoreConfig {
708 /// **What shipped before this module existed, and what still ships.**
709 ///
710 /// `SafeWriter` + `ObjectReadStack<OneTableFourColumns>` + `NewGeneration`.
711 /// [`GitStore::open`](crate::git_ops::GitStore::open) builds exactly this
712 /// and consults no environment to do it.
713 pub const DEFAULT: StoreConfig = StoreConfig {
714 writer: WriterArm::Safe,
715 index: IndexArm::OneTableFourColumns,
716 gc: GcArm::NewGeneration,
717 redb_cache_bytes: DEFAULT_REDB_CACHE_BYTES,
718 };
719
720 /// Read the three arms and the cache ceiling. **Four `getenv`s, once, here.**
721 ///
722 /// It is called from a store's constructor and from nowhere on a serving
723 /// path; [`env_reads`] is what makes that assertable rather than asserted-
724 /// by-comment. An unset or empty variable takes the default; a variable set
725 /// to an unknown value is an error rather than a silent fallback, so a typo
726 /// cannot make a measurement look like it came from an arm that never ran.
727 ///
728 /// The ceiling rides along because it is read on the same construction path
729 /// and printed by the same `Display`: a caller that logs `from_env()`'s
730 /// result once at store open has then logged **every** per-store variable
731 /// this crate consults, and `ZNIPPY_GIT_REDB_CACHE_BYTES` is no longer the
732 /// one that can be set on a host, stop at a container boundary, and leave
733 /// no trace of having done so.
734 pub fn from_env() -> Result<Self> {
735 let mut cfg = StoreConfig::DEFAULT;
736 if let Some(v) = read_env(ENV_WRITER) {
737 cfg.writer = WriterArm::parse(&v)?;
738 }
739 if let Some(v) = read_env(ENV_INDEX) {
740 cfg.index = IndexArm::parse(&v)?;
741 }
742 if let Some(v) = read_env(ENV_GC) {
743 cfg.gc = GcArm::parse(&v)?;
744 }
745 cfg.redb_cache_bytes = redb_cache_bytes()?;
746 Ok(cfg)
747 }
748
749 /// Same three arms with a different writer. For a bench that sweeps one axis.
750 pub fn with_writer(mut self, w: WriterArm) -> Self {
751 self.writer = w;
752 self
753 }
754
755 pub fn with_index(mut self, i: IndexArm) -> Self {
756 self.index = i;
757 self
758 }
759
760 pub fn with_gc(mut self, g: GcArm) -> Self {
761 self.gc = g;
762 self
763 }
764
765 /// Same arms, a different redb page-cache ceiling. What
766 /// [`GitStore::open`](crate::git_ops::GitStore::open) uses to put the
767 /// environment's ceiling on the default arms without reading any arm.
768 pub fn with_redb_cache_bytes(mut self, bytes: usize) -> Self {
769 self.redb_cache_bytes = bytes;
770 self
771 }
772}
773
774/// `KEY=value` for every per-store variable, in the spelling an operator
775/// would type — so one log line at store open is a complete, reproducible
776/// statement of what the environment selected, including the ceiling.
777impl std::fmt::Display for StoreConfig {
778 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
779 write!(
780 f,
781 "{}={} {}={} {}={} {}={}",
782 ENV_WRITER,
783 self.writer.as_str(),
784 ENV_INDEX,
785 self.index.as_str(),
786 ENV_GC,
787 self.gc.as_str(),
788 ENV_REDB_CACHE,
789 self.redb_cache_bytes
790 )
791 }
792}
793
794#[cfg(test)]
795mod tests {
796 use super::*;
797 use crate::archive_write::read_journal;
798 use crate::git_ops::{GitOps, GitStore, open_from_env, open_selected};
799 use crate::index_layout::{OneTableFourColumns, PackedPayload};
800 use crate::object::GitHashKind;
801 use crate::store::tests::{corpus_pack, tmpdir};
802 use std::sync::Mutex;
803
804 #[test]
805 fn the_graph_cache_switch_parses_on_and_off_and_refuses_anything_else() {
806 for v in ["on", "1", "true", "ON"] {
807 assert!(parse_graph_cache(v).unwrap(), "{v}");
808 }
809 for v in ["off", "0", "false", "Off"] {
810 assert!(!parse_graph_cache(v).unwrap(), "{v}");
811 }
812 let err = parse_graph_cache("sometimes").unwrap_err().to_string();
813 assert!(err.contains(ENV_GRAPH_CACHE), "{err}");
814 }
815
816 fn loadavg() -> String {
817 std::fs::read_to_string("/proc/loadavg")
818 .unwrap_or_default()
819 .split_whitespace()
820 .take(3)
821 .collect::<Vec<_>>()
822 .join(" ")
823 }
824
825 /// The environment is process-global and this crate's tests run in parallel
826 /// threads, so every guard that sets or counts an environment read takes
827 /// this first. Without it [`env_reads`] would be a shared counter two tests
828 /// moved at once, and the read-once guard below would be a coin toss.
829 static ENV_LOCK: Mutex<()> = Mutex::new(());
830
831 /// Run `f` with this module's variables set to `vars`, restoring whatever
832 /// was there before.
833 ///
834 /// `ENV_REDB_CACHE` is cleared along with the three arms even though no test
835 /// sets it: it is read on the same construction path, and a value inherited
836 /// from the surrounding shell would otherwise change the footprint a test
837 /// measures without appearing anywhere in the test.
838 fn with_env<R>(vars: &[(&str, &str)], f: impl FnOnce() -> R) -> R {
839 let _g = ENV_LOCK.lock().unwrap_or_else(|e| e.into_inner());
840 let saved: Vec<(String, Option<String>)> = [ENV_WRITER, ENV_INDEX, ENV_GC, ENV_REDB_CACHE]
841 .iter()
842 .map(|k| (k.to_string(), std::env::var(k).ok()))
843 .collect();
844 // SAFETY (edition 2024): the process-wide environment is mutated only
845 // under ENV_LOCK, and no thread this crate spawns reads it.
846 unsafe {
847 for (k, _) in &saved {
848 std::env::remove_var(k);
849 }
850 for (k, v) in vars {
851 std::env::set_var(k, v);
852 }
853 }
854 let out = f();
855 unsafe {
856 for (k, v) in &saved {
857 match v {
858 Some(v) => std::env::set_var(k, v),
859 None => std::env::remove_var(k),
860 }
861 }
862 }
863 out
864 }
865
866 /// **The writer arm really is the writer that runs — asserted on the files
867 /// on disk, not on a name.**
868 ///
869 /// The two arms differ in exactly one observable thing and it is a file: a
870 /// durable arm writes an Arrow IPC journal row naming the extent before
871 /// `append` returns, and [`WriterArm::Fast`] writes no journal at all. So
872 /// the same pack is pushed into two stores that differ only in
873 /// [`StoreConfig::writer`], and what is asserted is the **payload byte for
874 /// byte in both** (the arm must not change what is stored) and the
875 /// **journal present in one and absent in the other** (the arm must change
876 /// what is promised).
877 ///
878 /// Seen RED by ignoring the selection in `GitStore::open_with_arms` —
879 /// `arms.writer.create(&blobs)?` → `WriterArm::Safe.create(&blobs)?`, which
880 /// is the hardcoded `SafeWriter::create(&blobs)?` this whole change removes:
881 /// "the fast arm left a journal at
882 /// /tmp/…-arm-writer-fast-…/objects.pack.journal — a writer that was not
883 /// selected ran". Restored.
884 ///
885 /// `writer_name()` is asserted **after** the two directory listings, on
886 /// purpose: with it first the same break failed on a `&'static str`
887 /// (`left: "SafeWriter", right: "FastWriter"`), which proves a label was
888 /// copied and not that a byte moved.
889 ///
890 /// **What it does NOT catch, found by trying.** Making `WriterArm::journal`
891 /// return `Some` for `Fast` as well leaves this guard **green**: no
892 /// `SafeWriter` runs, so no journal file is ever created, and
893 /// `!fast_journal.exists()` still holds. That break is caught one level down
894 /// by [`only_the_arms_that_write_a_journal_name_one`], which asserts on the
895 /// path itself rather than on a file — recorded here rather than quietly
896 /// claimed, because a guard's blind spots are the part worth writing down.
897 #[test]
898 fn each_writer_arm_leaves_its_own_durability_on_disk() {
899 let (pack, rows) = corpus_pack();
900
901 // ── the fast arm: bytes, and nothing that says they were acked ───────
902 let fast_dir = tmpdir("arm-writer-fast");
903 let fast = GitStore::<OneTableFourColumns>::open_with_arms(
904 &fast_dir,
905 "alice",
906 GitHashKind::Sha1,
907 StoreConfig::DEFAULT.with_writer(WriterArm::Fast),
908 )
909 .unwrap();
910 let fast_tx = fast.put_pack(&pack).unwrap();
911 let (fo, fl) = fast_tx.extent.unwrap();
912 let fast_blobs = fast_dir.join("objects.pack");
913 let fast_journal = crate::archive_write::journal_path(&fast_blobs);
914 assert_eq!(
915 &std::fs::read(&fast_blobs).unwrap()[fo as usize..(fo + fl) as usize],
916 &pack[..],
917 "the fast arm did not store the pack verbatim — the arm may change the promise, \
918 never the payload"
919 );
920 assert!(
921 !fast_journal.exists(),
922 "the fast arm left a journal at {} — a writer that was not selected ran",
923 fast_journal.display()
924 );
925
926 // ── the safe arm: the same bytes, plus the durable row that claims them
927 let safe_dir = tmpdir("arm-writer-safe");
928 let safe = GitStore::<OneTableFourColumns>::open_with_arms(
929 &safe_dir,
930 "alice",
931 GitHashKind::Sha1,
932 StoreConfig::DEFAULT.with_writer(WriterArm::Safe),
933 )
934 .unwrap();
935 let safe_tx = safe.put_pack(&pack).unwrap();
936 let (so, sl) = safe_tx.extent.unwrap();
937 let safe_blobs = safe_dir.join("objects.pack");
938 let safe_journal = crate::archive_write::journal_path(&safe_blobs);
939 assert_eq!(
940 &std::fs::read(&safe_blobs).unwrap()[so as usize..(so + sl) as usize],
941 &pack[..],
942 "the safe arm did not store the pack verbatim"
943 );
944 assert!(
945 safe_journal.exists(),
946 "the safe arm wrote no journal — the durable arm did not run"
947 );
948 assert_eq!(
949 read_journal(&safe_journal).unwrap(),
950 vec![(so, sl)],
951 "the journal does not name the extent that was acked"
952 );
953
954 // Both arms store the same bytes at the same offset: the selection
955 // moves the durability contract and nothing else.
956 assert_eq!((fo, fl), (so, sl), "the two arms disagree about the extent");
957 // The labels, asserted **after** the evidence — a name that agreed with
958 // a directory listing that did not would be the wrong thing to fail on.
959 assert_eq!(fast.writer_name(), "FastWriter");
960 assert_eq!(safe.writer_name(), "SafeWriter");
961 fast.wait_indexed();
962 safe.wait_indexed();
963 assert_eq!(fast.object_count(), rows.len());
964 assert_eq!(safe.object_count(), rows.len());
965 eprintln!(
966 "load {}; {} objects: fast arm {} journal row(s), safe arm {} — {}",
967 loadavg(),
968 rows.len(),
969 if fast_journal.exists() { 1 } else { 0 },
970 read_journal(&safe_journal).unwrap().len(),
971 WriterArm::Fast.durability(),
972 );
973 }
974
975 /// **The third writer arm is selectable through a store too, and it lands
976 /// the same durable journal.**
977 ///
978 /// The `fast`/`safe` guard above is the contrast; this one is the coverage.
979 /// It asserts on applied output — the pack verbatim on disk and one journal
980 /// row naming its extent — which is the same pair
981 /// [`each_writer_arm_leaves_its_own_durability_on_disk`] asserts for `safe`,
982 /// because the two durable arms write **one** journal format through two
983 /// transports (LAW 5) and a store must not be able to tell them apart.
984 ///
985 /// A kernel that cannot run the chain is a **failure, not a skip**: the same
986 /// policy `tests/archive_write.rs` already takes, because a silently skipped
987 /// arm is an arm nobody notices stopped working.
988 ///
989 /// Seen RED by pointing the arm at the wrong writer —
990 /// `WriterArm::Uring => Box::new(SafeWriter::create(blobs)?)` in
991 /// `WriterArm::create`: "the uring arm ran a different writer — left:
992 /// \"SafeWriter\", right: \"UringWriter\"". Restored.
993 ///
994 /// **That red lands on a name, and here that is the honest ceiling.** The
995 /// two durable arms are *required* to be indistinguishable on disk — one
996 /// journal format, one encoder, two transports — so there is no applied
997 /// output that separates them, and inventing one would mean breaking LAW 5
998 /// to make a guard feel better. The applied-output assertions above prove
999 /// the chain did the job; the name is the only thing that can say which
1000 /// transport did it.
1001 #[cfg(target_os = "linux")]
1002 #[test]
1003 fn the_uring_arm_is_selectable_through_a_store_and_lands_the_same_journal() {
1004 let dir = tmpdir("arm-writer-uring");
1005 let store = GitStore::<OneTableFourColumns>::open_with_arms(
1006 &dir,
1007 "alice",
1008 GitHashKind::Sha1,
1009 StoreConfig::DEFAULT.with_writer(WriterArm::Uring),
1010 )
1011 .expect(
1012 "the io_uring arm could not be built on this kernel — not skipped, this is a real \
1013 failure",
1014 );
1015 let (pack, oid) = crate::store::tests::one_blob_pack(b"a push down the io_uring chain");
1016 let tx = store.put_pack(&pack).unwrap();
1017 let (o, l) = tx.extent.unwrap();
1018
1019 let blobs = dir.join("objects.pack");
1020 assert_eq!(
1021 &std::fs::read(&blobs).unwrap()[o as usize..(o + l) as usize],
1022 &pack[..],
1023 "the uring arm did not store the pack verbatim"
1024 );
1025 assert_eq!(
1026 read_journal(&crate::archive_write::journal_path(&blobs)).unwrap(),
1027 vec![(o, l)],
1028 "the uring arm's journal does not name the extent it acked — the two durable arms \
1029 must write one journal format"
1030 );
1031 assert_eq!(store.writer_name(), "UringWriter", "the uring arm ran a different writer");
1032 store.wait_indexed();
1033 assert!(store.has(&oid).unwrap(), "the pushed object never reached the index");
1034 }
1035
1036 /// **The gc arm really is the gc that runs — asserted on which files exist
1037 /// afterwards.**
1038 ///
1039 /// [`GcArm::NewGeneration`] renames the compacted archive to `x.g1.znippy`
1040 /// and unlinks the original **last**; [`GcArm::CompactInPlace`] renames over
1041 /// the original name and produces no generation. Those are two different
1042 /// sets of files on disk for the same input, which is what is asserted —
1043 /// `GcReport::strategy` is checked too, but it is a label and the
1044 /// directory listing is the evidence.
1045 ///
1046 /// Seen RED by ignoring the selection in `GitStore::open_with_arms` —
1047 /// `gc: arms.gc.create()` → `gc: GcArm::NewGeneration.create()`: "the
1048 /// in-place arm produced a new generation at
1049 /// /tmp/…-arm-gc-in-place-…/repository.g1.znippy — a gc that was not
1050 /// selected ran". Restored.
1051 ///
1052 /// `report.strategy` is asserted **after** the directory listing, on
1053 /// purpose: with the strategy check first the same break failed on a
1054 /// `&'static str` (`left: "NewGeneration", right: "CompactInPlace"`), which
1055 /// is a label agreeing with itself and says nothing about what happened to
1056 /// the archive.
1057 #[test]
1058 fn each_gc_arm_leaves_its_own_generation_on_disk() {
1059 for arm in GcArm::ALL {
1060 let dir = tmpdir(&format!("arm-gc-{}", arm.as_str()));
1061 let store = GitStore::<OneTableFourColumns>::open_with_arms(
1062 &dir,
1063 "alice",
1064 GitHashKind::Sha1,
1065 StoreConfig::DEFAULT.with_gc(arm),
1066 )
1067 .unwrap();
1068 let (pack, _) = corpus_pack();
1069 store.put(&pack, &[]).unwrap();
1070 store.absorb_pending().unwrap();
1071
1072 let root = store
1073 .graph_snapshot()
1074 .into_iter()
1075 .find(|c| c.generation == 1)
1076 .expect("a root commit");
1077 let root_raw = hex::decode(&root.oid).unwrap();
1078 store
1079 .update_ref("refs/heads/root", None, Some(&root_raw))
1080 .unwrap();
1081
1082 // A real znippy archive for the compaction step to work on — the
1083 // same fixture `store::tests`' own gc guard builds.
1084 let files = vec![
1085 ("pack-0.pack".to_string(), pack.clone()),
1086 ("pack-1.pack".to_string(), pack.clone()),
1087 ];
1088 znippy_common::create_archive(store.archive_path(), &files, 3).unwrap();
1089 let original = store.archive_path().to_path_buf();
1090 let generation = crate::gc::next_generation(&original).unwrap();
1091
1092 let report = store.gc().unwrap();
1093 match arm {
1094 GcArm::NewGeneration => {
1095 assert_eq!(report.archive, generation);
1096 assert!(
1097 generation.exists(),
1098 "the new-generation arm produced no {}",
1099 generation.display()
1100 );
1101 assert!(
1102 !original.exists(),
1103 "the new-generation arm kept the old generation at {}",
1104 original.display()
1105 );
1106 assert_eq!(report.retired.as_deref(), Some(original.as_path()));
1107 assert!(report.verified, "the new generation was not read back");
1108 }
1109 GcArm::CompactInPlace => {
1110 assert!(
1111 !generation.exists(),
1112 "the in-place arm produced a new generation at {} — a gc that was not \
1113 selected ran",
1114 generation.display()
1115 );
1116 assert!(
1117 original.exists(),
1118 "the in-place arm removed the archive it compacts into"
1119 );
1120 assert_eq!(report.archive, original);
1121 assert_eq!(report.retired, None);
1122 }
1123 }
1124 // The label, after the evidence.
1125 assert_eq!(report.strategy, arm.strategy());
1126 assert!(
1127 report.bytes_after <= report.bytes_before,
1128 "{} grew the archive: {} → {}",
1129 arm.strategy(),
1130 report.bytes_before,
1131 report.bytes_after
1132 );
1133 eprintln!(
1134 "load {}; gc arm {}: {} → {} bytes, archive now {}",
1135 loadavg(),
1136 arm.strategy(),
1137 report.bytes_before,
1138 report.bytes_after,
1139 report.archive.display()
1140 );
1141 }
1142 }
1143
1144 /// **The index arm really is the layout that gets built — asserted on the
1145 /// Arrow bytes it materialised, and on all three agreeing about every
1146 /// object.**
1147 ///
1148 /// A type name would prove nothing here (all three stacks report
1149 /// `ObjectReadStack`, because that is what they are). What distinguishes the
1150 /// arms is the projection they actually built:
1151 /// [`ObjectIndex::ipc_bytes`] is the size of the Arrow IPC payload held
1152 /// resident, and one packed 25-byte column, four columns in one section and
1153 /// four independent sections are three different numbers for the same 1352
1154 /// objects. Pairwise-distinct is the assertion.
1155 ///
1156 /// The other half matters more: the arms are a **layout** choice, so all
1157 /// three must answer identically. Every oid's full row is compared across
1158 /// the three, so an arm that decoded its own payload wrongly fails here
1159 /// rather than in production.
1160 ///
1161 /// Driven through [`open_selected`] — the runtime door — so the guard fails
1162 /// when the *selector* stops selecting rather than only when a layout is
1163 /// broken.
1164 ///
1165 /// Seen RED by pinning the type in `open_selected`, every match arm building
1166 /// `SelectedStore::OneTableFourColumns(GitStore::<OneTableFourColumns>…)`,
1167 /// which is the state this change starts from: "one-table and four-tables
1168 /// materialised the same 145544 IPC bytes — the index arm was not
1169 /// selected". Restored.
1170 ///
1171 /// **`store.arms().index` stayed green under that break**, and that is the
1172 /// reason it is not the assertion: `arms` is the value that was *requested*,
1173 /// so it agrees with the caller no matter which type got built. The byte
1174 /// count is what the store actually did.
1175 #[test]
1176 fn each_index_arm_builds_its_own_projection_and_all_three_agree() {
1177 let (pack, rows) = corpus_pack();
1178
1179 let mut built = Vec::new();
1180 for arm in IndexArm::ALL {
1181 let store = open_selected(
1182 &tmpdir(&format!("arm-index-{}", arm.as_str())),
1183 "alice",
1184 GitHashKind::Sha1,
1185 StoreConfig::DEFAULT.with_index(arm),
1186 )
1187 .unwrap();
1188 store.put_pack(&pack).unwrap();
1189 store.wait_indexed();
1190 // A rebuild, so the projection under test — not the redb tail — is
1191 // what answers below.
1192 store.rebuild_projection().unwrap();
1193 assert_eq!(store.arms().index, arm);
1194 // The name the BUILT projection reports, against the name this arm
1195 // says it builds. `arms().index` above is the request echoed back
1196 // and stays green when the selector selects nothing;
1197 // `index_name()` goes through `ObjectReadStack::projection_name`
1198 // to `S::name()`, so it can only answer what was monomorphised.
1199 //
1200 // It is what a server logs to say which layout it is running —
1201 // gunnar's `store.znippy_arms_selected` — and the reason
1202 // `ObjectIndex::name` on the stack itself is no use for that: the
1203 // stack answers "ObjectReadStack" for all three, because that is
1204 // what the stack is.
1205 assert_eq!(
1206 store.index_name(),
1207 arm.projection_name(),
1208 "the {} arm built a projection calling itself {}",
1209 arm.as_str(),
1210 store.index_name()
1211 );
1212 built.push((arm, store));
1213 }
1214
1215 // Applied output: three layouts, three different quantities of Arrow IPC
1216 // materialised for the same objects.
1217 for (i, (a, sa)) in built.iter().enumerate() {
1218 for (b, sb) in &built[i + 1..] {
1219 assert_ne!(
1220 sa.index_ipc_bytes(),
1221 sb.index_ipc_bytes(),
1222 "{} and {} materialised the same {} IPC bytes — the index arm was not selected",
1223 a.as_str(),
1224 b.as_str(),
1225 sa.index_ipc_bytes()
1226 );
1227 }
1228 }
1229
1230 // …and they are three spellings of one answer. `get` goes through the
1231 // whole stack — index row, extent, bytes off disk — so a layout that
1232 // decoded its own payload wrongly cannot agree here by accident.
1233 for r in rows.iter().take(512) {
1234 let first = built[0].1.get(&r.oid).unwrap();
1235 assert!(first.is_some(), "{} is missing from one-table", hex::encode(&r.oid));
1236 for (arm, store) in &built[1..] {
1237 assert_eq!(
1238 store.get(&r.oid).unwrap(),
1239 first,
1240 "{} disagrees with one-table about {}",
1241 arm.as_str(),
1242 hex::encode(&r.oid)
1243 );
1244 }
1245 }
1246 eprintln!(
1247 "load {}; {} objects: {}",
1248 loadavg(),
1249 rows.len(),
1250 built
1251 .iter()
1252 .map(|(a, s)| format!("{} {} B", a.as_str(), s.index_ipc_bytes()))
1253 .collect::<Vec<_>>()
1254 .join(", ")
1255 );
1256 }
1257
1258 /// **The default did not move, and no environment variable can move it.**
1259 ///
1260 /// The compatibility half of this change. `GitStore::open` built
1261 /// `SafeWriter` + `ObjectReadStack<OneTableFourColumns>` + `NewGeneration`
1262 /// before the selector existed and must build exactly that after it —
1263 /// including in a process whose environment names all three *other* arms,
1264 /// because a caller that never asked for a selector must not have its
1265 /// durability contract changed by somebody's shell.
1266 ///
1267 /// Asserted on applied output rather than on `arms()`: the journal row that
1268 /// only a durable writer produces, and the generation file that only
1269 /// `NewGeneration` produces.
1270 ///
1271 /// Seen RED by making the old constructor read the environment —
1272 /// `Self::open_with_arms(root, account, hash, StoreConfig::DEFAULT)` →
1273 /// `Self::open_with_arms(root, account, hash, StoreConfig::from_env()?)` in
1274 /// `GitStore::open_with`: "GitStore::open wrote no journal — an environment
1275 /// variable moved the default durability contract". Restored.
1276 ///
1277 /// `arms()` is asserted **after** the journal, for the third time in this
1278 /// module and for the same reason: with it first the break failed on
1279 /// `StoreConfig { writer: Fast, … }` vs `StoreConfig { writer: Safe, … }`,
1280 /// which is the request echoed back rather than the contract that was kept.
1281 #[test]
1282 fn the_default_is_unchanged_and_the_environment_cannot_move_it() {
1283 with_env(
1284 &[
1285 (ENV_WRITER, "fast"),
1286 (ENV_INDEX, "packed"),
1287 (ENV_GC, "in-place"),
1288 ],
1289 || {
1290 let dir = tmpdir("arm-default");
1291 let store = GitStore::open(&dir, "alice").unwrap();
1292 let (pack, rows) = corpus_pack();
1293 let tx = store.put(&pack, &[]).unwrap();
1294 let journal = crate::archive_write::journal_path(&dir.join("objects.pack"));
1295 assert!(
1296 journal.exists(),
1297 "GitStore::open wrote no journal — an environment variable moved the default \
1298 durability contract"
1299 );
1300 assert_eq!(read_journal(&journal).unwrap(), vec![tx.extent.unwrap()]);
1301 store.wait_indexed();
1302 assert_eq!(store.object_count(), rows.len());
1303 // The labels, after the evidence.
1304 assert_eq!(
1305 store.arms(),
1306 StoreConfig::DEFAULT,
1307 "GitStore::open did not build the shipping arms"
1308 );
1309 assert_eq!(store.writer_name(), "SafeWriter");
1310
1311 // And the gc arm is still the one that keeps the old generation
1312 // until the new one is proven.
1313 let root = store
1314 .graph_snapshot()
1315 .into_iter()
1316 .find(|c| c.generation == 1)
1317 .expect("a root commit");
1318 let root_raw = hex::decode(&root.oid).unwrap();
1319 store
1320 .update_ref("refs/heads/root", None, Some(&root_raw))
1321 .unwrap();
1322 let files = vec![("pack-0.pack".to_string(), pack.clone())];
1323 znippy_common::create_archive(store.archive_path(), &files, 3).unwrap();
1324 let generation = crate::gc::next_generation(store.archive_path()).unwrap();
1325 let report = store.gc().unwrap();
1326 assert_eq!(report.strategy, "NewGeneration");
1327 assert!(
1328 generation.exists(),
1329 "GitStore::open's gc arm is not NewGeneration — the environment moved it"
1330 );
1331 },
1332 );
1333 }
1334
1335 /// **The selector is read once, at construction, and never per operation.**
1336 ///
1337 /// gunnar's own rule, stated in its source: *read once, here, and not per
1338 /// entry — a getenv inside the copy loop would be one syscall per object
1339 /// served*. A per-object `getenv` was a real defect found and fixed there,
1340 /// so this is asserted with a counter rather than trusted to review.
1341 ///
1342 /// [`env_reads`] counts every environment read this crate makes. The whole
1343 /// corpus pack is then pushed and served — one `put`, 1352 `has`, a batch
1344 /// `extents`, `get`, `size`, `refs` — across a **counter that does not
1345 /// move**.
1346 ///
1347 /// **Five reads at construction since 2026-08-21** (four since 2026-08-10,
1348 /// three before). Three are [`StoreConfig::from_env`]'s arms; the fourth is
1349 /// [`redb_cache_bytes`], which `from_env` now reads alongside them so the
1350 /// ceiling is part of the config a caller logs; the fifth is the explode
1351 /// policy `ExplodedArchive::open` reads through the same counted door.
1352 /// None of the extra two is an arm — they select no implementation — but
1353 /// each is an environment read, so each is counted, and each is read
1354 /// exactly once per store rather than once per database or once per lookup.
1355 ///
1356 /// Seen RED by re-reading the selector on the lookup path — adding
1357 /// `let _ = crate::arms::StoreConfig::from_env()?;` at the top of
1358 /// `GitStore::lookup_one`: "the selector was read **8070** times while
1359 /// serving 2692 operations — it must be read once, at construction; left:
1360 /// 8070, right: 3". Restored.
1361 ///
1362 /// Seen RED a second time, and this one is what keeps the guard from being
1363 /// vacuous: `StoreConfig::from_env` stubbed to `return
1364 /// Ok(StoreConfig::DEFAULT)` before it reads anything gives "opening a store
1365 /// read the environment **0** times, not once per variable". A selector that
1366 /// is never read at all also never moves the counter during serving, so
1367 /// without the construction-time line the guard would pass over a selector
1368 /// that does nothing. Restored.
1369 #[test]
1370 fn the_selector_is_read_once_at_construction_and_never_per_operation() {
1371 with_env(
1372 &[
1373 (ENV_WRITER, "safe"),
1374 (ENV_INDEX, "four-tables"),
1375 (ENV_GC, "in-place"),
1376 ],
1377 || {
1378 let dir = tmpdir("arm-read-once");
1379 let (pack, rows) = corpus_pack();
1380
1381 // **The once-per-PROCESS knobs are warmed first, and without
1382 // this the guard is a coin toss.** It measured 12 reads run
1383 // alone and 7 run inside the suite, because the memory
1384 // watchdog's five keys and the resolver's four are read
1385 // through `OnceLock`s that some earlier test usually
1386 // initialised: the number this asserts depended on the test
1387 // ORDER, so it was green in the suite and red on its own —
1388 // which is how it sat red on master from 2026-09-15 without
1389 // anybody being able to read it. Warmed here, the window below
1390 // holds the per-store-open reads and nothing else, in both
1391 // orders.
1392 let _ = crate::memguard::global();
1393 let _ = crate::resolve::base_budget_from_env();
1394 let _ = crate::resolve::row_batch_from_env();
1395 let _ = crate::resolve::workers_from_env();
1396 let _ = crate::resolve::lookahead_from_env();
1397
1398 // `env_reads_here`, not `env_reads`: the process-wide counter
1399 // is moved by every other test thread that opens a store, and
1400 // since `redb_cache_bytes` joined the readers that is most of
1401 // them. Per thread it is exact.
1402 let before = env_reads_here();
1403 let store = open_from_env(&dir, "alice", GitHashKind::Sha1).unwrap();
1404 let at_open = env_reads_here();
1405 // Three arms, the redb cache ceiling, the explode policy, the
1406 // explode budget and the graph-cache switch. Once each — a
1407 // store opens TWO redb databases and still reads the ceiling
1408 // once.
1409 //
1410 // It said SIX from 2026-08-30 until 2026-09-21, and the
1411 // seventh is [`ENV_GRAPH_CACHE`]: 5ed59f2 (2026-09-15) added
1412 // `graph_cache_from_env` to `GitStore::open_with_arms` and put
1413 // the key in [`ALL_ENV`], but did not move this number — so
1414 // this guard has been RED on master ever since, which is a
1415 // guard nobody can read. The read is legitimate and wanted;
1416 // the count was stale. If an eighth appears, the question to
1417 // ask is whether the new key is read once per store open or
1418 // once per operation, not whether to raise the number again.
1419 assert_eq!(
1420 at_open - before,
1421 7,
1422 "opening a store read the environment {} times, not once per variable",
1423 at_open - before
1424 );
1425
1426 // Everything below this line is serving. Nothing here may read
1427 // the environment even once.
1428 let mut ops = 0u64;
1429 store.put(&pack, &[]).unwrap();
1430 ops += 1;
1431 let oids: Vec<&[u8]> = rows.iter().map(|r| r.oid.as_slice()).collect();
1432 for oid in &oids {
1433 assert!(store.has(oid).unwrap());
1434 ops += 1;
1435 }
1436 let ext = store.extents(&oids).unwrap();
1437 ops += 1;
1438 assert_eq!(ext.len(), oids.len());
1439 assert!(ext.iter().all(Option::is_some));
1440 assert!(store.get(oids[0]).unwrap().is_some());
1441 assert!(store.size(oids[0]).unwrap().is_some());
1442 store.refs().unwrap();
1443 ops += 3;
1444
1445 let after = env_reads_here();
1446 assert_eq!(
1447 after, at_open,
1448 "the selector was read {} times while serving {ops} operations — it must be \
1449 read once, at construction",
1450 after - before
1451 );
1452 eprintln!(
1453 "load {}; {ops} operations over {} objects: {} environment read(s), all of \
1454 them at construction",
1455 loadavg(),
1456 rows.len(),
1457 at_open - before,
1458 );
1459 },
1460 );
1461 }
1462
1463 /// **A bad value is an error, not a silent default** — through the real
1464 /// front door, so the refusal cannot be a parser test that nothing calls.
1465 #[test]
1466 fn an_unknown_arm_refuses_to_open_rather_than_falling_back() {
1467 with_env(&[(ENV_WRITER, "safest")], || {
1468 let dir = tmpdir("arm-typo");
1469 let e = match open_from_env(&dir, "alice", GitHashKind::Sha1) {
1470 Ok(_) => panic!("a typo opened a store on the default arm"),
1471 Err(e) => e,
1472 };
1473 let msg = format!("{e:#}");
1474 assert!(
1475 msg.contains("'safest' is not a writer arm") && msg.contains(ENV_WRITER),
1476 "the refusal must name the bad value and the variable it came from: {msg}"
1477 );
1478 });
1479 // …and an empty environment is the shipping default, not an error.
1480 with_env(&[], || {
1481 assert_eq!(StoreConfig::from_env().unwrap(), StoreConfig::DEFAULT);
1482 });
1483 }
1484
1485 /// **The runtime door and the typed door build the same store.**
1486 ///
1487 /// [`open_selected`] exists so a value can choose the index arm, which is a
1488 /// type; the risk it carries is that the dispatched path drifts from the
1489 /// monomorphised one. Asserted on applied output: the same pack pushed
1490 /// through both answers the same for every oid.
1491 #[test]
1492 fn the_selected_store_and_the_typed_store_answer_alike() {
1493 let (pack, rows) = corpus_pack();
1494 let arms = StoreConfig::DEFAULT.with_index(IndexArm::PackedPayload);
1495
1496 let boxed = open_selected(
1497 &tmpdir("arm-selected"),
1498 "alice",
1499 GitHashKind::Sha1,
1500 arms,
1501 )
1502 .unwrap();
1503 let typed = GitStore::<PackedPayload>::open_with_arms(
1504 &tmpdir("arm-typed"),
1505 "alice",
1506 GitHashKind::Sha1,
1507 arms,
1508 )
1509 .unwrap();
1510
1511 boxed.put(&pack, &[]).unwrap();
1512 typed.put(&pack, &[]).unwrap();
1513 for r in rows.iter().take(512) {
1514 let a = boxed.get(&r.oid).unwrap();
1515 let b = typed.get(&r.oid).unwrap();
1516 assert_eq!(a, b, "the boxed and typed stores disagree about {}", hex::encode(&r.oid));
1517 assert!(a.is_some());
1518 }
1519 }
1520
1521 /// Round-trips through every spelling, and a typo is an error rather than a
1522 /// default.
1523 #[test]
1524 fn every_arm_parses_from_the_name_an_operator_types() {
1525 for a in WriterArm::ALL {
1526 assert_eq!(WriterArm::parse(a.as_str()).unwrap(), a);
1527 assert_eq!(WriterArm::parse(&a.as_str().to_uppercase()).unwrap(), a);
1528 }
1529 for a in IndexArm::ALL {
1530 assert_eq!(IndexArm::parse(a.as_str()).unwrap(), a);
1531 }
1532 for a in GcArm::ALL {
1533 assert_eq!(GcArm::parse(a.as_str()).unwrap(), a);
1534 }
1535 // A typo must not become the default: an operator who mistyped `safe`
1536 // and got a measurement labelled `safe` would be reading a lie.
1537 for bad in ["saef", "", "fastwriter2", "none"] {
1538 assert!(
1539 WriterArm::parse(bad).is_err(),
1540 "'{bad}' parsed as a writer arm"
1541 );
1542 }
1543 assert!(IndexArm::parse("one-tabel").is_err());
1544 assert!(GcArm::parse("newgen").is_err());
1545 }
1546
1547 /// **[`ALL_ENV`] is complete, checked against the crate's own source.**
1548 ///
1549 /// Walks every `.rs` under this crate's `src/`, cuts each file at its first
1550 /// `#[cfg(test)]` (the convention here: one tests module, at the bottom),
1551 /// and finds every environment read in what is left — `std::env::var(…)`,
1552 /// `env::var_os(…)` and [`read_env`]`(…)`. The key each one names, whether
1553 /// a string literal or one of the `ENV_*` constants declared in this file,
1554 /// must be on [`ALL_ENV`]. The only `std::env::var` allowed outside that
1555 /// rule is the one inside [`read_env`] itself, which takes its key as a
1556 /// parameter.
1557 ///
1558 /// This is what goes red the day the crate grows a key without naming it —
1559 /// the defect gunnar's forge could not see from its side (it asserted its
1560 /// crossing list against a count it remembered, and the count was wrong
1561 /// for four months of one day). A source scan is the honest shape for it:
1562 /// the keys are string literals and the reads are call sites, and no
1563 /// runtime counter can tell which *names* were read.
1564 ///
1565 /// Seen RED by adding `std::env::var("ZNIPPY_GIT_NEW_KNOB")` to
1566 /// `git_ops.rs` outside its tests: "git_ops.rs reads ZNIPPY_GIT_NEW_KNOB
1567 /// and ALL_ENV does not name it". Restored.
1568 #[test]
1569 fn every_environment_read_in_this_crate_is_named_in_all_env() {
1570 let src = Path::new(env!("CARGO_MANIFEST_DIR")).join("src");
1571 let this_file = std::fs::read_to_string(src.join("arms.rs")).unwrap();
1572
1573 // `pub const ENV_X: &str = "…";` → ENV_X ↦ "…", from this file only.
1574 let mut consts: std::collections::BTreeMap<String, String> = Default::default();
1575 for line in this_file.lines() {
1576 let t = line.trim();
1577 let Some(rest) = t.strip_prefix("pub const ENV_") else { continue };
1578 let Some((name, val)) = rest.split_once(": &str = \"") else { continue };
1579 let val = val.split('"').next().unwrap();
1580 consts.insert(format!("ENV_{name}"), val.to_string());
1581 }
1582 assert!(consts.len() >= 8, "fewer ENV_* constants than expected: {consts:?}");
1583
1584 let mut seen: Vec<(String, String)> = Vec::new();
1585 let mut files = std::fs::read_dir(&src)
1586 .unwrap()
1587 .map(|e| e.unwrap().path())
1588 .filter(|p| p.extension().is_some_and(|e| e == "rs"))
1589 .collect::<Vec<_>>();
1590 files.sort();
1591 assert!(!files.is_empty());
1592 for path in files {
1593 let file = path.file_name().unwrap().to_string_lossy().into_owned();
1594 let text = std::fs::read_to_string(&path).unwrap();
1595 let non_test = match text.find("\n#[cfg(test)]") {
1596 Some(at) => &text[..at],
1597 None => &text[..],
1598 };
1599 for needle in ["env::var(", "env::var_os(", "read_env("] {
1600 let mut from = 0;
1601 while let Some(at) = non_test[from..].find(needle) {
1602 let call_at = from + at;
1603 from = call_at + needle.len();
1604 // The definition of `read_env` itself, and doc/comment lines
1605 // that merely mention a call, are not reads.
1606 let line_start = non_test[..call_at].rfind('\n').map_or(0, |i| i + 1);
1607 let line = non_test[line_start..].lines().next().unwrap_or("").trim_start();
1608 if line.starts_with("//") || line.starts_with("pub(crate) fn read_env") {
1609 continue;
1610 }
1611 // The `env::var(key)` inside `read_env` — the one door.
1612 if file == "arms.rs" && needle == "env::var(" && line.contains("var(key)") {
1613 continue;
1614 }
1615 let arg = non_test[from..].trim_start();
1616 let key = if let Some(lit) = arg.strip_prefix('"') {
1617 lit.split('"').next().unwrap().to_string()
1618 } else {
1619 let ident: String = arg
1620 .chars()
1621 .take_while(|c| c.is_ascii_alphanumeric() || *c == '_' || *c == ':')
1622 .collect();
1623 let short = ident.rsplit("::").next().unwrap_or(&ident).to_string();
1624 consts.get(&short).cloned().unwrap_or_else(|| {
1625 panic!(
1626 "{file} reads the environment through `{needle}{ident}…`, which \
1627 is neither a string literal nor an ENV_* constant this test \
1628 can resolve — name the key as a `pub const ENV_*` in arms.rs"
1629 )
1630 })
1631 };
1632 seen.push((file.clone(), key));
1633 }
1634 }
1635 }
1636 assert!(
1637 seen.len() >= 8,
1638 "the scan found only {} environment reads; it is supposed to find at least the \
1639 eight in ALL_ENV ({seen:?})",
1640 seen.len()
1641 );
1642 for (file, key) in &seen {
1643 assert!(
1644 ALL_ENV.contains(&key.as_str()),
1645 "{file} reads {key} and ALL_ENV does not name it"
1646 );
1647 }
1648 // And the other direction: a key on the list that nothing reads is a
1649 // stale entry a consumer would be carrying for nothing.
1650 for key in ALL_ENV {
1651 assert!(
1652 seen.iter().any(|(_, k)| k == key),
1653 "ALL_ENV names {key} but no non-test code in this crate reads it"
1654 );
1655 }
1656 }
1657
1658 /// The default is the shipping combination, spelled out rather than derived
1659 /// — a derived `Default` that drifted would move the durability contract.
1660 #[test]
1661 fn the_default_config_is_the_shipping_combination() {
1662 assert_eq!(StoreConfig::default(), StoreConfig::DEFAULT);
1663 assert_eq!(StoreConfig::DEFAULT.writer, WriterArm::Safe);
1664 assert_eq!(StoreConfig::DEFAULT.index, IndexArm::OneTableFourColumns);
1665 assert_eq!(StoreConfig::DEFAULT.gc, GcArm::NewGeneration);
1666 assert_eq!(
1667 StoreConfig::DEFAULT.to_string(),
1668 "ZNIPPY_GIT_WRITER=safe ZNIPPY_GIT_INDEX=one-table ZNIPPY_GIT_GC=new-generation \
1669 ZNIPPY_GIT_REDB_CACHE_BYTES=67108864"
1670 );
1671 }
1672
1673 /// **Only the durable arms name a journal**, and the fast one names none.
1674 ///
1675 /// Not cosmetic: the path returned here is what a reopen derives §13.12's
1676 /// `indexed` bit from, so an arm that writes no journal must not hand back
1677 /// the name of one — a store that ran on `SafeWriter` and is reopened on
1678 /// `FastWriter` would then diff against a log this writer is not appending
1679 /// to, and resume pack ordinals from it.
1680 ///
1681 /// Seen RED by `WriterArm::Fast => Some(journal_path(blobs))`: "left:
1682 /// Some(\"/nonexistent/objects.pack.journal\") right: None". Restored.
1683 /// This is the guard that catches that break — the disk-level one above
1684 /// stays green for it, which is why both exist.
1685 #[test]
1686 fn only_the_arms_that_write_a_journal_name_one() {
1687 let blobs = Path::new("/nonexistent/objects.pack");
1688 assert_eq!(WriterArm::Fast.journal(blobs), None);
1689 assert_eq!(
1690 WriterArm::Safe.journal(blobs),
1691 Some(PathBuf::from("/nonexistent/objects.pack.journal"))
1692 );
1693 assert_eq!(
1694 WriterArm::Uring.journal(blobs),
1695 WriterArm::Safe.journal(blobs),
1696 "the two durable arms share one journal format (LAW 5) and must share its name"
1697 );
1698 }
1699
1700 /// The arm's stated durability is the writer's own, not a second copy of it
1701 /// that could drift.
1702 #[test]
1703 fn the_arms_durability_line_matches_the_writer_it_builds() {
1704 let dir = crate::store::tests::tmpdir("arm-durability");
1705 for arm in [WriterArm::Fast, WriterArm::Safe] {
1706 let w = arm.create(&dir.join(format!("{}.pack", arm.as_str()))).unwrap();
1707 match arm {
1708 WriterArm::Fast => {
1709 assert_eq!(w.name(), "FastWriter");
1710 assert!(
1711 w.durability().starts_with("none"),
1712 "FastWriter must say plainly that it promises nothing: {}",
1713 w.durability()
1714 );
1715 assert!(arm.durability().starts_with("none"));
1716 }
1717 WriterArm::Safe => {
1718 assert_eq!(w.name(), "SafeWriter");
1719 assert_eq!(w.durability(), arm.durability());
1720 }
1721 WriterArm::Uring => unreachable!(),
1722 }
1723 }
1724 }
1725}