fix(storage): counts persistence is single-flight, coalesced, and never races its own temp file
Some checks failed
CI / Node 24 (push) Successful in 12m24s
CI / Node 22 (push) Successful in 12m35s
CI / Integration + conformance (Node 22) (push) Failing after 17m25s
CI / Bun (latest) (push) Successful in 12m19s

persistCounts() was write-through on every count change with no
serialization, and the atomic writer named its temp file with millisecond
granularity. Two persists inside one millisecond shared the temp path: both
wrote it, the first rename consumed it, the second rename found nothing —
ENOENT, roughly 1,500 times a day on a busy production brain, with a full
ledger write per change behind it. No data was lost (the surviving rename
carried a complete ledger and the next change re-persisted), but the race
was real and the write rate absurd.

flushCounts() now runs exactly one persist at a time; requests arriving
during it collapse into one trailing pass that carries the burst's final
state — N changes cost at most two writes. writeFileAtomic() adds a
per-process sequence to the temp name so no two writes can share a path.
Pinned: a 25-change burst → ≤2 ledger writes, zero errors, ledger equal to
memory; parallel real writes land complete; three same-instant atomic
writes own three distinct temp paths.
This commit is contained in:
David Snelling 2026-09-01 09:32:23 -07:00
parent 4014e0f125
commit 5e3b343a0e
3 changed files with 162 additions and 9 deletions

View file

@ -1089,6 +1089,10 @@ export abstract class BaseStorageAdapter implements StorageAdapter {
// Counts changed since the last persist? Drives the write-through flush.
protected pendingCountPersist = false
/** The one persist running right now, if any (single-flight law — see flushCounts). */
private countPersistInFlight: Promise<void> | null = null
/** The one trailing persist a burst has queued behind the in-flight one. */
private countPersistTrailing: Promise<void> | null = null
/**
* Get total noun count - O(1) operation
@ -1341,15 +1345,46 @@ export abstract class BaseStorageAdapter implements StorageAdapter {
return
}
try {
// Persist to storage (implemented by subclass)
await this.persistCounts()
this.pendingCountPersist = false
} catch (error) {
console.error('CRITICAL: Failed to flush counts to storage:', error)
// Keep pending flag set so we retry on next operation
throw error
// SINGLE-FLIGHT, COALESCED. Counts are write-through on every change, so
// a burst of writes used to launch one persist per change, all in flight
// together. Two of them inside the same millisecond shared the atomic
// writer's temp path (`.tmp-<pid>-<ms>`): both wrote it, the first rename
// consumed it, the second rename found nothing — ENOENT, ~1,500 times a
// day on a busy production brain, with a full ledger write per change
// behind it. Now exactly one persist runs at a time; requests that arrive
// while it runs collapse into ONE trailing persist that carries the final
// state. A burst of N changes costs at most two writes and never races
// itself.
if (this.countPersistInFlight) {
// The in-flight write may have already serialised a stale snapshot —
// ask for one more pass after it, and let every caller in this burst
// await that same pass.
if (!this.countPersistTrailing) {
this.countPersistTrailing = this.countPersistInFlight
.catch(() => undefined)
.then(() => {
this.countPersistTrailing = null
return this.flushCounts()
})
}
return this.countPersistTrailing
}
this.countPersistInFlight = (async () => {
try {
// Persist to storage (implemented by subclass)
this.pendingCountPersist = false
await this.persistCounts()
} catch (error) {
// Keep the flag set so the next operation retries.
this.pendingCountPersist = true
console.error('CRITICAL: Failed to flush counts to storage:', error)
throw error
} finally {
this.countPersistInFlight = null
}
})()
return this.countPersistInFlight
}
/**

View file

@ -2400,8 +2400,15 @@ export class FileSystemStorage extends BaseStorage {
* Atomic write via temp-file-then-rename so concurrent readers never see a
* half-written lock JSON. Reused by writer-lock writes + heartbeat.
*/
/** Monotonic per-process sequence so two atomic writes never share a temp path. */
private static atomicWriteSeq = 0
private async writeFileAtomic(filePath: string, contents: string): Promise<void> {
const tmp = `${filePath}.tmp-${process.pid}-${Date.now()}`
// pid + timestamp alone collided: two writers of the same target inside
// one millisecond shared this path, and the loser's rename found the
// winner had already moved it (ENOENT). The sequence makes every call's
// temp path its own.
const tmp = `${filePath}.tmp-${process.pid}-${Date.now()}-${++FileSystemStorage.atomicWriteSeq}`
await fs.promises.writeFile(tmp, contents)
await fs.promises.rename(tmp, filePath)
}

View file

@ -0,0 +1,111 @@
/**
* @module tests/integration/counts-persist-single-flight
* @description Regression for a production race in FileSystemStorage's
* counts ledger: `persistCounts()` was write-through on every count change
* with no serialization, and the atomic writer named its temp file with
* millisecond granularity (`.tmp-<pid>-<ms>`). Two persists inside one
* millisecond shared the temp path both wrote it, the first rename
* consumed it, the second rename found nothing: ENOENT, ~1,500 times a day
* on a busy production brain, with a full ledger write per change behind it.
*
* Under pin: persists are single-flight and coalesced one in flight, at
* most one trailing pass carrying the burst's final state and every atomic
* write owns a unique temp path. A burst of N count changes costs at most
* two ledger writes, never errors, and leaves a ledger equal to memory.
*/
import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest'
import * as fs from 'node:fs'
import * as os from 'node:os'
import * as path from 'node:path'
import { Brainy } from '../../src/brainy.js'
import { NounType } from '../../src/types/graphTypes.js'
describe('counts persistence is single-flight, coalesced, and never races its own temp file', () => {
let dir: string
let brain: any
beforeEach(async () => {
process.env.BRAINY_DETERMINISTIC_EMBEDDINGS = 'true'
dir = fs.mkdtempSync(path.join(os.tmpdir(), 'brainy-counts-race-'))
brain = new Brainy({
requireSubtype: false,
storage: { type: 'filesystem', path: dir },
dimensions: 384,
silent: true
})
await brain.init()
})
afterEach(async () => {
vi.restoreAllMocks()
await brain.close()
fs.rmSync(dir, { recursive: true, force: true })
})
it('a burst of concurrent count changes → at most two ledger writes, zero errors, ledger == memory', async () => {
const storage = brain.storage
const countsPath: string = storage.countsFilePath
expect(countsPath, 'the filesystem adapter persists a counts ledger').toBeTruthy()
// Let init's own persists settle so the burst is measured alone.
await storage.flushCounts?.()
const renameSpy = vi.spyOn(fs.promises, 'rename')
const errorSpy = vi.spyOn(console, 'error')
// Twenty-five concurrent count changes — the shape of a write burst; each
// used to launch its own persist.
const BURST = 25
await Promise.all(
Array.from({ length: BURST }, () => storage.scheduleCountPersist())
)
const ledgerRenames = renameSpy.mock.calls.filter(([, to]) => String(to) === countsPath)
expect(ledgerRenames.length, 'single-flight + one trailing pass').toBeLessThanOrEqual(2)
expect(ledgerRenames.length, 'the burst was persisted at all').toBeGreaterThanOrEqual(1)
const persistErrors = errorSpy.mock.calls.filter((args) => String(args[0]).includes('persisting counts'))
expect(persistErrors).toEqual([])
const ledger = JSON.parse(fs.readFileSync(countsPath, 'utf-8'))
expect(ledger.totalNounCount).toBe(storage.totalNounCount)
expect(ledger.totalVerbCount).toBe(storage.totalVerbCount)
})
it('real writes in parallel: the ledger lands complete and no persist error is logged', async () => {
const storage = brain.storage
const countsPath: string = storage.countsFilePath
const errorSpy = vi.spyOn(console, 'error')
await Promise.all(
Array.from({ length: 12 }, (_, i) =>
brain.add({ data: `burst row ${i}`, type: NounType.Thing })
)
)
await storage.flushCounts?.()
const persistErrors = errorSpy.mock.calls.filter((args) => String(args[0]).includes('persisting counts'))
expect(persistErrors).toEqual([])
const ledger = JSON.parse(fs.readFileSync(countsPath, 'utf-8'))
expect(ledger.totalNounCount).toBe(storage.totalNounCount)
expect(await brain.getNounCount()).toBe(ledger.totalNounCount)
})
it('every atomic write owns its own temp path — two writes in one millisecond never collide', async () => {
const storage = brain.storage
const tmpNames: string[] = []
vi.spyOn(fs.promises, 'writeFile').mockImplementation(async (p: any) => {
tmpNames.push(String(p))
})
vi.spyOn(fs.promises, 'rename').mockImplementation(async () => undefined)
const target = path.join(dir, 'probe.json')
await Promise.all([
storage.writeFileAtomic(target, '{"a":1}'),
storage.writeFileAtomic(target, '{"a":2}'),
storage.writeFileAtomic(target, '{"a":3}')
])
const probeTmps = tmpNames.filter((n) => n.startsWith(`${target}.tmp-`))
expect(probeTmps.length).toBe(3)
expect(new Set(probeTmps).size, 'no two writes shared a temp path').toBe(3)
})
})