mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-27 22:15:33 +00:00
* feat(minions): pool-recovery audit + reconnect reason-threading + shared drain helper (#1685 GAP B, 5A) - pool-recovery-audit.ts: reap_detected (CONNECTION_ENDED) vs reconnect_other; recovered/failed split - postgres-engine reconnect(ctx?) classifies the triggering error so only true pooler reaps are tagged (CODEX #8) - retry.ts reconnect callback widened to thread the error; retry-matcher isConnectionEndedError - runExtractAtomsDrainForSource shared helper (cycleLockIdFor + withRefreshingLock) — one drain path (5A) - supervisor-audit readRecentSupervisorEvents (current+prev ISO week, CODEX #7) - extract-atoms-drain PROTECTED; autopilot.auto_drain.* config keys * feat(doctor): worker_oom_loop + pool_reap_health checks + cause-ranked top_issues (#1685 GAP A/B/C) - computeWorkerOomLoopCheck: unions supervisor rss_watchdog + minion_jobs watchdog-abort (CODEX #5), cap fallback to resolveDefaultMaxRssMb (CODEX #6) - computePoolReapHealthCheck: reaps-not-recovering fail, thrash warn - doctor-cause-rank rankIssues: tier ordering + grounded downstream_of (CODEX #9) + drift guard (4A) - supervisor causeStr + queue_health cross-reference worker_oom_loop (DRY 1C) - register both checks in doctor-categories ops * feat(autopilot): per-source extract_atoms auto-drain + handler + dream --drain refactor (#1685 GAP D) - autopilot per-source gate: enabled + !packDeclares + backlog>threshold + daily cap; time-sloted idempotency key (CODEX #2) - extract-atoms-drain Minion handler (thin wrapper, LockUnavailableError -> deferred) - dream --drain routes through the shared helper (5A) * chore: bump version and changelog (v0.42.12.0) #1685 brain-health-as-solved-problem: cause-ranked doctor, worker_oom_loop line, per-source auto-drain, pool-reap health. Layers on #1678/#1735. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(todos): file #1685 GAP E + remote-path follow-ups (v0.42.12.0) * fix(#1685): pre-landing review — multi-source auto-drain, honest pool-reap signal, lock-renewal reap labeling - autopilot: drop maxWaiting (coalesces by name+queue not source → only one source drained + cap over-count); pre-check idempotency key so only genuinely-new sources submit+count - pool_reap_health: fail on reconnect FAILURES (the real signal), not reaps>0&&failures>0 (false causality when a recovered reap + unrelated failure co-occur) - lock-renewal-tick threads its triggering error to reconnect() so a CONNECTION_ENDED pooler reap is labeled reap_detected not reconnect_other (pool_reap_health now fires for the #1678 incident path) * chore: re-version v0.42.12.0 → v0.42.16.0 (#1685) Slot collision avoidance per queue. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs: restore slim CLAUDE.md + move #1685 entries to KEY_FILES.md (fix check:doc-history) The master merge wrongly kept the pre-restructure 577KB CLAUDE.md; the check:doc-history guard caps it at 60KB. Take master's slim CLAUDE.md and record the #1685 files (doctor-cause-rank, pool-recovery-audit, worker_oom_loop + pool_reap_health checks, auto-drain, 5A helper) as current-state prose in docs/architecture/KEY_FILES.md (no release markers). llms regenerated. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
102 lines
4.3 KiB
TypeScript
102 lines
4.3 KiB
TypeScript
// #1685 GAP A — worker_oom_loop doctor check.
|
|
//
|
|
// Hermetic: writes synthetic supervisor audit JSONL into a GBRAIN_AUDIT_DIR
|
|
// tempdir and drives computeWorkerOomLoopCheck with engine=null (supervised
|
|
// half + cap logic + thresholds; the minion_jobs bare-worker branch is
|
|
// Postgres-only and mirrors the queue_health subcheck-3 query covered by E2E).
|
|
// Also pins the cross-week supervisor reader (CODEX #7).
|
|
|
|
import { describe, expect, test, beforeEach, afterEach } from 'bun:test';
|
|
import * as fs from 'fs';
|
|
import * as os from 'os';
|
|
import * as path from 'path';
|
|
import { withEnv } from './helpers/with-env.ts';
|
|
import { computeWorkerOomLoopCheck } from '../src/commands/doctor.ts';
|
|
import {
|
|
computeSupervisorAuditFilename,
|
|
readRecentSupervisorEvents,
|
|
} from '../src/core/minions/handlers/supervisor-audit.ts';
|
|
|
|
let tmpDir: string;
|
|
beforeEach(() => {
|
|
tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'worker-oom-loop-'));
|
|
});
|
|
afterEach(() => {
|
|
try { fs.rmSync(tmpDir, { recursive: true, force: true }); } catch { /* best-effort */ }
|
|
});
|
|
|
|
function writeSupervisorRows(rows: object[], fileDate = new Date()): void {
|
|
const file = path.join(tmpDir, computeSupervisorAuditFilename(fileDate));
|
|
fs.writeFileSync(file, rows.map((r) => JSON.stringify(r)).join('\n') + '\n', 'utf8');
|
|
}
|
|
|
|
const nowIso = () => new Date().toISOString();
|
|
const rssKill = () => ({ event: 'worker_exited', ts: nowIso(), likely_cause: 'rss_watchdog', code: 12 });
|
|
const breaker = (cap: number) => ({ event: 'health_warn', ts: nowIso(), reason: 'rss_watchdog_loop', max_rss_mb: cap });
|
|
|
|
describe('computeWorkerOomLoopCheck', () => {
|
|
test('fail with breaker-stamped cap when the breaker tripped', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
writeSupervisorRows([rssKill(), rssKill(), rssKill(), rssKill(), rssKill(), rssKill(), breaker(2048)]);
|
|
const c = await computeWorkerOomLoopCheck(null);
|
|
expect(c?.status).toBe('fail');
|
|
expect(c?.name).toBe('worker_oom_loop');
|
|
expect(c?.message).toContain('cap=2048MB');
|
|
expect(c?.message).toContain('raise --max-rss');
|
|
expect(c?.details?.oom_kills).toBe(6);
|
|
expect(c?.details?.supervisor_kills).toBe(6);
|
|
expect(c?.details?.bare_worker_kills).toBe(0);
|
|
expect(c?.details?.cap_source).toBe('breaker');
|
|
expect(c?.details?.breaker_tripped).toBe(true);
|
|
});
|
|
});
|
|
|
|
test('warn with auto-sized cap fallback when no breaker event stamped a cap (CODEX #6)', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
writeSupervisorRows([rssKill(), rssKill()]);
|
|
const c = await computeWorkerOomLoopCheck(null);
|
|
expect(c?.status).toBe('warn');
|
|
expect(c?.details?.cap_source).toBe('default');
|
|
expect(c?.message).toContain('auto-sized default');
|
|
expect(c?.details?.oom_kills).toBe(2);
|
|
});
|
|
});
|
|
|
|
test('fail at >=5 kills even without a breaker event', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
writeSupervisorRows([rssKill(), rssKill(), rssKill(), rssKill(), rssKill()]);
|
|
const c = await computeWorkerOomLoopCheck(null);
|
|
expect(c?.status).toBe('fail');
|
|
expect(c?.details?.breaker_tripped).toBe(false);
|
|
});
|
|
});
|
|
|
|
test('null when the worker never OOM-looped', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
writeSupervisorRows([{ event: 'started', ts: nowIso() }]);
|
|
expect(await computeWorkerOomLoopCheck(null)).toBeNull();
|
|
});
|
|
});
|
|
|
|
test('null on an empty audit dir', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
expect(await computeWorkerOomLoopCheck(null)).toBeNull();
|
|
});
|
|
});
|
|
});
|
|
|
|
describe('readRecentSupervisorEvents — cross-week (CODEX #7)', () => {
|
|
test('reads a within-24h event from the PREVIOUS ISO-week file', async () => {
|
|
await withEnv({ GBRAIN_AUDIT_DIR: tmpDir }, async () => {
|
|
// Event timestamped within the 24h window but written into last week's
|
|
// file (the Monday-reads-Sunday case). The single-file reader would miss
|
|
// it; the cross-week reader must find it.
|
|
const prevWeekFileDate = new Date(Date.now() - 7 * 86400000);
|
|
writeSupervisorRows([rssKill()], prevWeekFileDate);
|
|
const events = readRecentSupervisorEvents(24);
|
|
expect(events.length).toBe(1);
|
|
expect(events[0].event).toBe('worker_exited');
|
|
});
|
|
});
|
|
});
|