fix(recovery): a torn generation-log tail is a terminal verdict, never a wait
Some checks failed
CI / Node 22 (push) Successful in 12m22s
CI / Node 24 (push) Successful in 12m21s
CI / Integration + conformance (Node 22) (push) Failing after 16m58s
CI / Bun (latest) (push) Successful in 12m23s

Two halves of one defect, found by a seeded-SIGKILL crash lane.

THE FALSE POSITIVE. stampEntityTree() recorded generationStore.generation()
— the ALLOCATED counter, a number a write in flight has claimed and may
never commit — while the JSDoc beside it already said the source is the
committed generation. Every crash inside a write window therefore produced
a spurious verdict at the next open: either 'sourceGeneration N is ahead of
the log head N-1' (the allocated generation died with the process) or
'rollup invariant nounCount: stamped X, observed Y' (the recovery fold
folded facts the stamp's counts predate). Both told the operator to run
repairIndex() — a whole-store recount — for a store that was coherent.
Measured before this commit: 4 of 11 SIGKILL cycles on a healthy store
raised one of the two. The stamp and the open now both read
committedGeneration(), which is what every other open-time watermark in the
class already reasons about.

THE TERMINAL VERDICT. A stamp still ahead of committed truth after the
recovery fold witnesses a generation that is not in the log — the stamp's
fsync outlived the tail's, and there is nothing to arrive. That is its own
verdict state now ('torn'), never folded in with 'incoherent': the two have
opposite cures. A writer open demotes it — the unusable stamped surface is
re-derived at the committed generation from the live counters, O(1),
straight-line, no loop and no await on external progress, narrated with
both count sets, the stamp's path and its committedAt. A read-only open
cannot re-stamp, so it says so and names the cure instead of guessing, and
still serves. Neither branch waits, and neither locks an owner out of a
canonical tree the stamp only describes.

Pins: the verifier returns the torn verdict with both generations; a
fabricated head-behind-source store narrates precisely, demotes inside a
bounded open, serves its rows, and is quiet at the next open (the demotion
converges); a read-only open narrates the same verdict and leaves the bytes
untouched.
This commit is contained in:
David Snelling 2026-08-31 09:07:18 -07:00
parent b8475cc86a
commit 298cb6daca
3 changed files with 241 additions and 16 deletions

View file

@ -12497,6 +12497,18 @@ export class Brainy<T = any> implements BrainyInterface<T> {
* healed by `repairIndex()`, whose unconditional recount rebuilds the
* rollups from a canonical walk and re-stamps. Best-effort: a stamp-write
* fault warns loudly but never fails the flush that carried real data.
*
* THE SOURCE IS `committedGeneration()`, NEVER `generation()`. The latter is
* the ALLOCATED counter a number a write in flight has claimed and may
* never commit. Stamping it made the stamp's generation label a claim about
* counts it was not taken at, and every crash inside a write window then
* produced a spurious verdict at the next open: either `sourceGeneration N
* is ahead of the log head N-1` (the allocated generation died with the
* process) or `rollup invariant 'nounCount': stamped X, observed Y` (the
* recovery fold folded facts the stamp's counts predate). MEASURED on the
* crash-consistency lane before this line changed: 4 of 11 SIGKILL cycles on
* a coherent store raised one of those two verdicts, each of them naming
* `repairIndex()` a whole-store recount as the cure for nothing.
*/
private async stampEntityTree(): Promise<void> {
if (this.isReadOnly) return
@ -12507,7 +12519,7 @@ export class Brainy<T = any> implements BrainyInterface<T> {
])
await writeFamilyStamp(this.storage, ENTITY_TREE_STAMP_PATH, {
family: 'entity-tree',
sourceGeneration: this.generationStore.generation(),
sourceGeneration: this.generationStore.committedGeneration(),
members: { mode: 'rollup', invariants: { nounCount, verbCount } }
})
} catch (error) {
@ -12520,16 +12532,24 @@ export class Brainy<T = any> implements BrainyInterface<T> {
/**
* @description Open-time coherence check for the entity tree's family stamp:
* compare `sourceGeneration` against the log head and the stamped rollup
* invariants against the live counters. Verdicts:
* compare `sourceGeneration` against the store's COMMITTED generation and
* the stamped rollup invariants against the live counters. Verdicts:
* - `coherent` / `absent` (legacy store; first flush stamps) silent.
* - `behind` benign for the tree (it is written BY the commit; only the
* stamp is stale a crash landed between commit and flush). Refreshed at
* the next flush.
* - `torn` a TORN GENERATION-LOG TAIL, handled by
* {@link demoteTornEntityTreeStamp}: terminal, never a wait.
* - `incoherent` LOUD: the tree or its counters diverged from what was
* stamped `repairIndex()` recounts from canonical and re-stamps.
* Never blocks open; a fault reading the stamp is surfaced as unverifiable,
* never conflated with absence.
*
* THE COMPARISON IS AGAINST `committedGeneration()`, matching what
* {@link stampEntityTree} writes and what every other open-time watermark in
* this class already reasons about (the fact-scan capability, the metadata /
* graph / HNSW watermark verdicts). Comparing against the allocated counter
* was the one place that disagreed, and disagreeing was the whole defect.
*/
private async verifyEntityTreeStamp(): Promise<void> {
let stamp: FamilyStamp | null
@ -12546,11 +12566,16 @@ export class Brainy<T = any> implements BrainyInterface<T> {
this.storage.getNounCount(),
this.storage.getVerbCount()
])
const verdict = verifyFamilyStamp(stamp, this.generationStore.generation(), {
const verdict = verifyFamilyStamp(stamp, this.generationStore.committedGeneration(), {
nounCount,
verbCount
})
if (verdict.state === 'incoherent') {
if (verdict.state === 'torn') {
await this.demoteTornEntityTreeStamp(stamp as FamilyStamp, verdict.stampSource, verdict.head, {
nounCount,
verbCount
})
} else if (verdict.state === 'incoherent') {
prodLog.warn(
`[Brainy] entity-tree stamp INCOHERENT at open: ${verdict.failures.join('; ')}. ` +
`The canonical tree or its counters diverged from the stamped state — run ` +
@ -12564,6 +12589,92 @@ export class Brainy<T = any> implements BrainyInterface<T> {
}
}
/**
* @description THE TERMINAL VERDICT for a torn generation-log tail.
*
* A stamp whose `sourceGeneration` sits ABOVE the store's committed
* watermark witnesses a generation that is not in the log: the stamp's fsync
* outlived the tail's. By the time this runs, log-authority recovery has
* already folded every intact fact above the manifest and advanced the
* watermark to cover them so if the stamp is STILL ahead, the generation
* it names is not merely late, it is GONE. There is nothing to wait for.
*
* That is the whole point of this method. A field report of this class
* (single-process store, abrupt termination mid-fold) described a reopen
* that narrated the tear and then held 100% CPU with zero log growth for
* eight minutes before an operator wiped the directory. A recovery that
* cannot say what it is waiting for has no business spinning; the honest
* answer here is a verdict, taken now, at O(1) cost.
*
* WHAT THE VERDICT DOES the stamped surface is UNUSABLE, so it is
* discarded rather than believed: the stamped counts describe a generation
* that never became durable, and comparing them against live counters can
* only produce noise. The tree itself is not in question (it IS canonical
* every commit writes it, and the fold re-applied every after-image the log
* still holds), so the demotion is a re-derivation of this family's verified
* surface at the generation the store can actually show:
*
* - WRITER open re-stamp at `committedGeneration()` from the live
* counters exactly what the next flush would write, taken now so the
* tear cannot re-narrate on every subsequent open. Both count sets are
* logged so an operator can see whether anything really moved.
* - READER open a reader cannot re-stamp. Narrate the same terminal
* verdict with the named cure and carry on serving; a read-only inspector
* is never locked out of a store, and never left waiting either.
*
* BOUNDEDNESS: straight-line code. No loop, no retry, no await on any
* external progress signal the two counter reads and one stamp write are
* the entire cost, and none of them scales with the store.
*/
private async demoteTornEntityTreeStamp(
stamp: FamilyStamp,
stampSource: number,
head: number,
observed: { nounCount: number; verbCount: number }
): Promise<void> {
const stamped = stamp.members.mode === 'rollup' ? stamp.members.invariants : {}
const detail =
`[Brainy] TORN GENERATION-LOG TAIL at open: ${ENTITY_TREE_STAMP_PATH} witnesses source ` +
`generation ${stampSource} (stamped ${stamp.committedAt}), but the store's committed ` +
`generation is ${head} after crash recovery — the stamp's fsync outlived the log tail's, ` +
`and generation ${stampSource} is not in the log to arrive. Stamped rollups ` +
`${JSON.stringify(stamped)}; observed ${JSON.stringify(observed)}.`
if (this.isReadOnly) {
prodLog.warn(
`${detail} This open is READ-ONLY, so the stamp cannot be re-derived: the entity-tree ` +
`family stays UNVERIFIED for this session (reads are unaffected — the canonical tree ` +
`is the truth this stamp only describes). Cure: open the store with a writer, or run ` +
`brain.repairIndex() there, to recount from canonical and re-stamp.`
)
return
}
const startedAt = Date.now()
try {
await writeFamilyStamp(this.storage, ENTITY_TREE_STAMP_PATH, {
family: 'entity-tree',
sourceGeneration: head,
members: {
mode: 'rollup',
invariants: { nounCount: observed.nounCount, verbCount: observed.verbCount }
}
})
prodLog.warn(
`${detail} DEMOTED: the unusable stamp was re-derived at committed generation ${head} ` +
`from the live counters in ${Date.now() - startedAt}ms — terminal, not a wait. If the ` +
`observed counts above look wrong for your data, run brain.repairIndex() to recount ` +
`from canonical.`
)
} catch (error) {
prodLog.warn(
`${detail} The demotion's re-stamp FAILED (${(error as Error).message}) — the tear will ` +
`narrate again at the next open, which is the honest outcome; the store still serves ` +
`from canonical. Cure: run brain.repairIndex() to recount from canonical and re-stamp.`
)
}
}
/**
* Ask the writer process serving this data directory to flush its in-memory
* indexes to disk, so a read-only inspector can observe fresh state.

View file

@ -12,9 +12,11 @@
* the verified surface is a small set of rollup invariants (entity/
* relationship counts) plus `sourceGeneration`.
*
* `sourceGeneration` is the generation of the source-of-truth log this
* projection reflects open-time coherence becomes a COMPARISON (stamp vs
* log head), not a walk:
* `sourceGeneration` is the COMMITTED generation of the source-of-truth log
* this projection reflects never the allocated counter, which names a
* generation that may never commit (see {@link StampVerdict.torn}) so
* open-time coherence becomes a COMPARISON (stamp vs committed head), not a
* walk:
*
* - equal + invariants hold coherent, serve.
* - behind the projection missed the tail (crash between commit and stamp);
@ -24,6 +26,9 @@
* - invariants FAIL at equal generation genuine incoherence: loud, and the
* repair ritual (`repairIndex()`, whose recount rebuilds the rollups from a
* canonical walk) heals it.
* - AHEAD a torn generation-log tail: the stamp's fsync outlived the log
* tail's. TERMINAL, never a wait the generation the stamp names does not
* exist to arrive.
*
* Stamps are JSON on purpose every incident gets debugged by reading a
* stamp in a terminal.
@ -70,6 +75,12 @@ export type StampVerdict =
| { state: 'coherent' }
| { state: 'absent' } // legacy store — first stamp writes at the next flush
| { state: 'behind'; stampSource: number; head: number }
/**
* TORN GENERATION-LOG TAIL: the stamp witnesses a source generation the
* store's committed watermark can no longer show. TERMINAL there is no
* generation to wait for, so the open demotes (or refuses) and never spins.
*/
| { state: 'torn'; stampSource: number; head: number }
| { state: 'incoherent'; failures: string[] }
| { state: 'unverifiable'; reason: string } // a FAULT reading the stamp — never conflated with absence
@ -118,12 +129,15 @@ export function verifyFamilyStamp(
): StampVerdict {
if (stamp === null) return { state: 'absent' }
if (stamp.sourceGeneration > head) {
// A stamp AHEAD of the log claims state that never committed — the
// projection was stamped against truth that a crash rolled back.
return {
state: 'incoherent',
failures: [`sourceGeneration ${stamp.sourceGeneration} is ahead of the log head ${head}`]
}
// A stamp AHEAD of committed truth witnesses a generation the store can no
// longer show: the stamp's fsync survived a crash that the log tail did
// not. This is the TORN GENERATION-LOG TAIL — its own class, never folded
// in with `incoherent` (a count that drifted at a generation both sides
// agree on), because the two have opposite cures: incoherence is recounted,
// a tear is DEMOTED. It is also terminal by construction — there is no
// generation the open can wait for, because the one the stamp names is
// gone.
return { state: 'torn', stampSource: stamp.sourceGeneration, head }
}
if (stamp.sourceGeneration < head) {
return { state: 'behind', stampSource: stamp.sourceGeneration, head }

View file

@ -57,7 +57,11 @@ describe('entity-tree family stamp', () => {
const invariants = (stamp.members as any).invariants
expect(invariants.nounCount).toBe(await brain.storage.getNounCount())
expect(invariants.verbCount).toBe(await brain.storage.getVerbCount())
expect(stamp.sourceGeneration).toBe(brain.generation())
// THE SOURCE IS COMMITTED TRUTH, never the allocated counter. Stamping the
// counter labelled the stamp with a generation a write in flight had merely
// claimed, so every crash inside a write window produced a spurious verdict
// at the next open (see the torn-tail pins below).
expect(stamp.sourceGeneration).toBe(brain.generationStore.committedGeneration())
expect(stamp.generation).toBeGreaterThanOrEqual(1)
})
@ -112,6 +116,96 @@ describe('entity-tree family stamp', () => {
expect(stillIncoherent).toEqual([])
})
/**
* Rewrite the on-disk stamp so its `sourceGeneration` sits ABOVE the store's
* committed watermark the durable shape a torn generation-log tail leaves
* behind (the stamp's fsync outlived the tail's). Fabricated rather than
* crash-produced so the pin is deterministic; the seeded-SIGKILL lane
* (`scripts/crash-consistency.mjs` in the engine repo) produces the same
* shape from a real abrupt termination.
*/
const fabricateTear = (ahead: number): FamilyStamp => {
const file = path.join(dir, `${ENTITY_TREE_STAMP_PATH}.gz`)
const zlib = require('node:zlib')
const raw = JSON.parse(zlib.gunzipSync(fs.readFileSync(file)).toString('utf-8')) as FamilyStamp
const torn: FamilyStamp = { ...raw, sourceGeneration: raw.sourceGeneration + ahead }
fs.writeFileSync(file, zlib.gzipSync(JSON.stringify(torn)))
return torn
}
it('a torn generation-log tail is a TERMINAL VERDICT at open: narrated, demoted, never a wait', async () => {
for (let i = 0; i < 3; i++)
await brain.add({ data: `torn${i}`, type: 'document', metadata: { i } })
await brain.close()
const torn = fabricateTear(5)
const warn = vi.spyOn(prodLog, 'warn')
const startedAt = Date.now()
brain = await open()
const openMs = Date.now() - startedAt
const tearLines = warn.mock.calls.filter((c) => String(c[0]).includes('TORN GENERATION-LOG TAIL'))
expect(tearLines.length).toBe(1)
const said = String(tearLines[0][0])
// Narrated PRECISELY: both generations, the file, and the named cure.
expect(said).toContain(`source generation ${torn.sourceGeneration}`)
expect(said).toContain(`committed generation ${brain.generationStore.committedGeneration()}`)
expect(said).toContain(ENTITY_TREE_STAMP_PATH)
expect(said).toContain('DEMOTED')
expect(said).toMatch(/repairIndex\(\)/)
// Terminal, not a wait: the demotion is O(1) straight-line work, so a tear
// cannot turn an open into the 8-minute spin this class was reported as.
expect(openMs).toBeLessThan(30_000)
// The store SERVES — a tear in a stamp never locks an owner out of the
// canonical tree the stamp merely describes.
expect((await brain.find({ type: 'document', limit: 100 })).length).toBe(3)
// The demotion CONVERGED: the stamp now names committed truth, and the
// next open is quiet. A verdict that re-narrates every open is a wait
// wearing a different hat.
const restamped = (await readFamilyStamp(brain.storage, ENTITY_TREE_STAMP_PATH)) as FamilyStamp
expect(restamped.sourceGeneration).toBe(brain.generationStore.committedGeneration())
await brain.close()
const warn2 = vi.spyOn(prodLog, 'warn')
brain = await open()
expect(warn2.mock.calls.filter((c) => String(c[0]).includes('TORN'))).toEqual([])
})
it('a READ-ONLY open on a torn tail refuses to guess: terminal verdict + named cure, no re-stamp', async () => {
await brain.add({ data: 'ro', type: 'document', metadata: {} })
await brain.close()
const torn = fabricateTear(3)
const warn = vi.spyOn(prodLog, 'warn')
const reader: any = await Brainy.openReadOnly({
requireSubtype: false,
storage: { type: 'filesystem', path: dir },
silent: true,
dimensions: 384
})
const tearLines = warn.mock.calls.filter((c) => String(c[0]).includes('TORN GENERATION-LOG TAIL'))
expect(tearLines.length).toBe(1)
const said = String(tearLines[0][0])
expect(said).toContain('READ-ONLY')
expect(said).toContain('UNVERIFIED')
expect(said).toMatch(/repairIndex\(\)/)
await reader.close()
// A reader never rewrites the store: read the bytes back off disk (not
// through a writer open, which would demote them) — the torn stamp is
// exactly as it was found.
const onDisk = JSON.parse(
require('node:zlib')
.gunzipSync(fs.readFileSync(path.join(dir, `${ENTITY_TREE_STAMP_PATH}.gz`)))
.toString('utf-8')
) as FamilyStamp
expect(onDisk.sourceGeneration).toBe(torn.sourceGeneration)
expect(onDisk.generation).toBe(torn.generation)
brain = await open()
})
it('the one verifier handles both member modes', () => {
const rollup: FamilyStamp = {
family: 'x',
@ -127,7 +221,13 @@ describe('entity-tree family stamp', () => {
stampSource: 5,
head: 9
})
expect(verifyFamilyStamp(rollup, 3, { nounCount: 10 }).state).toBe('incoherent') // ahead of head
// AHEAD is its own class — a torn generation-log tail, never folded in
// with `incoherent`: the two have opposite cures (recount vs demote).
expect(verifyFamilyStamp(rollup, 3, { nounCount: 10 })).toEqual({
state: 'torn',
stampSource: 5,
head: 3
})
expect(verifyFamilyStamp(null, 5, {})).toEqual({ state: 'absent' })
const enumerated: FamilyStamp = {