glm-5.3-flash
a48c12745a
pg-fix-scheduler-tick-seam — the pg scheduler's transient-fault tick seam lands, retiring review 003 Finding 1 (MEDIUM: the tick retry loop was dead code under test — the last uncovered hardening arm of the wave-4 fix batch, the exact class review 002's live bugs came from). The seam follows the sqlite-commit-error-arm pattern of record: a cfg(test) per-store counting fault arm (crate::seam::tick_fault — Flag = Arc<AtomicU64>, born disarmed at 0, arm(n) adds, take() consumes one via CAS try_update; a counting arm because a boolean cannot reach the exhaustion arm — arm(TICK_RETRIES+1) faults the loop's whole budget while arm(1) leaves the later ticks real) feeding a fabricated pool/database-shaped opaque Database error into the REAL retry loop — both tests drive the full run_schedules → tick_with_retries → tick_once path end-to-end, no loop replication. Plumbing per the house pattern: cfg-gated tick_fault field on EngineCtx and PgStore (engine_ctx() carries the clone; production builds never materialize it), cfg-gated fault param on tick_once (the seam check sits before the real attempt — one attempt faults, never ticks) and tick_with_retries, cfg-branched attempt calls; PgStore::arm_tick_fault(faults) arming surface + tick_fault_flag() observation surface (a Flag clone so the consumption pin survives the store moving into a runner task). Two arms pinned: (a) one_faulted_tick_is_retried_and_the_job_fires — fault consumed on attempt 1, the 250 ms backoff sleeps, the retried attempt fires the backdated due boundary into the queue, exactly-one consumption, the runner survives the hiccup to a clean stop Ok(()); (b) tick_retry_exhaustion_exits_typed_and_leaves_the_lock_to_lapse — the exhaustion exits typed (opaque Database, never LeadershipLost; the tick-phase context rides the source chain with the fabricated fault at its bottom; wall-clock ≥ 1.7 s proving the real backoff window), the leadership lock row is left (not released) to lapse at its TTL (present, expiry bounded by the run's TTL horizon), exact consumption, and the store remains fully usable after — the "runner survives" pin as the module docs' contract states it. Replay-proofed live: with the seam's consumption disabled the exhaustion test cannot complete (the runner loops forever — no fault ever enters the budget; the mutated run's timeout is itself the proof); the test is then bounded (20 s) so a future regression of this shape fails, not hangs. LEADER_TTL_S widened private → pub(crate) for the lock-left-to-TTL expiry pin. Verified: full pg column live vs the harness (124 lib + 25 suite + 9 schema; new tests green solo ×3); cargo test --workspace green server-less, skips clean (core 25 + harness 3, sqlite 191 + 25); clippy --all-targets -D warnings clean; fmt clean; cargo-llvm-cov confirms the Finding-1 sites (the retry/backoff/exhaustion-arm region + the tick_once seam arm) carry zero missed lines; taskgraph validate green (61 tasks)
2026-10-10 11:00:26 +00:00
glm-5.3-flash
57451cef46
pg-fix-scheduler-fire-wake — the scheduler tick's fire wake lands, closing the recorded cross-engine fire-wake latency-parity gap (the wave-5 gate's one standing item, review-wave-4-fixes' recorded-for-later note): scheduler.rs::tick now issues one coalesced pg_notify per firing schedule per tick after the fire loop, still inside the tick's in_tx frame — channel = the fired queue's name (schedule() rejects reserved names per ADR-021 §3, so the plain queue name is always the right channel), empty payload (Wake { channel } only, ADR-008 §3), best-effort via the shared tx::wake_tx (widened private → pub(crate) per the recorded one-call shape; enqueue_row stays wake-free — no double-wake; auto-commit Queue::enqueue and the tx producer paths untouched). NOTIFY's native transactional delivery makes the wake commit-atomic: a rolled-back tick (crash mid-tick ⇒ boundary refires) discards the wake with its fire rows — the tx/no-ghosts discipline the tx producer paths already pin; coalescing default one-per-firing-schedule-per-tick (not per-fire) matching the SQLite watcher's per-committed-tick cadence, decision recorded in the task's Notes. New four-arm pinning test (scheduler_tests.rs, tx_producer_wakes_are_commit_atomic pattern): pre-commit silence inside the runner's own in_tx frame driving the engine's own tick (rogue-ticks shape), deterministic delivery at/after commit, coalescing (≥2 boundaries → one wake), nothing-due silence (same-tick not-due schedule + a later all-not-due tick, in-frame and post-commit), and the end-to-end runner leg (a live run_schedules leader's wake reaches a registered listener); local listener_hears helper per the per-file helper convention; rollback arm native (NOTIFY transactional delivery — no tick-fault seam built, per the task pin). Docs: tx.rs 'The tx wakes' section + wake_tx doc name the scheduler fire path as a shared caller; scheduler.rs module docs + tick doc pin the coalesced commit-atomic semantics. Record updates: review-wave-4-fixes' recorded-for-later bullet, suite-scheduler-rows' gate disposition note + Notes bullet, review-wave-5 §Notes 4's audit row — all retired with pointers; implementation.md gains the review-rounds line (fires visible to registered listeners ahead of the mem engine's posture being written). Also carried: review-wave-5 §6 flake-ledger update — row_lock_ttl_expiry_and_reacquisition failed once more in a full pg run this session (green on the next two full runs + six focused; unrelated mechanism to this change), meeting the ledger's own 'fails again' trigger with the dedicated investigative session owed by the wave-7 watch carriage. Verified: full pg lib suite 122/122 vs the harness (new test green ×3 solo); pg contract suite 25/25 + schema 9/9; workspace cargo test green server-less (pg skips clean incl. the new test); clippy --all-targets -D warnings clean; fmt clean
2026-10-10 10:27:35 +00:00
glm-5.3-flash
5c03fb2130
pg suite-infra hardening (task pg-suite-infra-hardening): the wave-4 review's F-1 defense-in-depth candidates — must_recv_event re-shaped from the 20 ms try_recv polling loop to a parked recv().await under the 15 s timeout wrapper (stream_tests.rs, the sole pg copy — SQLite twin untouched; the poll loop's wake-subscription interaction surface the F-1 flake suspected is out of the hot path, and the deadline assert still bounds the property). The parked form's terminal arms are explicit: Some(Err(e)) keeps the errored-instead-of-idling panic, None gets a receiver-closed panic the poll form never saw. The deadline miss self-diagnoses (the small-honest realization of candidate (b)'s discriminator): read_since(offset, 1000) over rows beyond the receiver's position names the residual class — rows durable undelivered (wake/re-drain class) vs no rows (publish-visibility class); with the parked recv a wake-arrived-but-re-drain-missed arm is structurally implausible (only a read error could miss, and it surfaces Err). Candidate (b)'s literal bridge-side counter skipped, reason recorded in the task (surfacing the bridge wake count across the dyn EventReceiver boundary needs downcast/keyed-global machinery beyond the small-honest bar); the small honest piece did land: wait_wake's Lagged(n) arm now logs (eprintln, house posture matching notify.rs's bridge_loop — and forwarder.rs's 'the receiver bridge's Lagged arm logs / recovers' doc claim now true at both bridges), stream name threaded into wait_wake's signature so the log attributes the lag. Verified: pg stream module 21/21 vs harness server (full module run), tx_publishes_compose_with_the_handle 5 solo re-runs green (determinism re-check under the parked shape), two consecutive full pg harness runs green (121 lib + 25 suite + 9 schema, ~72 s each), cargo test -p alkstore-postgres green server-less (skips clean), workspace cargo test green (12 binaries), clippy -D warnings, fmt clean
2026-10-10 08:00:29 +00:00
glm-5.3-flash
9a00fb8a7e
Review gate — wave-4 fix batch passed: all seven pg-fix resolutions verified against review 002's failure mechanisms (code-read + live gates), one minor doc mismatch fixed inline (outbox wake comment's 'scheduler's fire-wake' claim — the tick fires issue no wake and schedule() can never target a reserved backing queue); no pinned posture regressed; pg suite green vs harness twice + compose determinism 20/20 re-verified; wave 5 may decompose (task review-wave-4-fixes)
2026-10-10 05:18:27 +00:00
glm-5.3-flash
49face898d
docs alignment: v1 TLS posture owned by deployment.md, QueueOpts numeric consumer-obligation notes (task pg-fix-docs-alignment, review 002 Finding 6 remainder)
...
- forwarder.rs's ListenerConnection doc corrected: NoTls is hardwired on
every connection path (pooled, listener, reconnect) — the pooled path
never rode the consumer's Config sslmode (a sslmode=require DSN fails
at connect); grep-audited no other in-crate doc repeats the claim
- PgOpts doc carries the corrected one-line TLS pointer (engine-crate-
docs posture, ADR-016 §2)
- deployment.md: new 'TLS posture (v1)' subsection (NoTls everywhere,
sslmode=require DSN fails at connect, topology-level confidentiality
is the v1 substitute, TLS a post-v1 deployment concern) and a new
'Consumer-obligation notes on engine options' section carrying the
QueueOpts trusted-as-given note with code-verified per-field symptoms
(max_attempts <= 0: never claimed, dead-lettered at the next claim
call's pre-claim sweep; negative visibility: instantly-reclaimable
claims; negative retention: every dead row at the next sweep_expired)
plus the PgOpts::max_size 0-guard counter-case; frontmatter advanced
- alkstore/src/opts.rs: QueueOpts struct doc mirrors the
consumer-obligation note (ADR-023 §2 scoping: the domain table covers
trait-surface arguments, not consumer-constructed constants)
- cross-file doc sweep over the fix batch's touched files (forwarder,
tx, scheduler, store) found no further doc-behavior mismatch
- gates: cargo build / clippy --all-targets -D warnings / fmt --check
all green (doc-only, no test touched)
2026-10-10 05:08:27 +00:00
glm-5.3-flash
40625090f6
pg queue/tx/notify dedupe + doc truth-telling: job_from_row has one owner (queue.rs, tx.rs imports it) and get_job_tx reuses live_columns()/dead_columns() instead of inlining the 18-column lists — the contract-pinned Job shape now has exactly one decode owner (ADR-012 §2); sweep_expired's doc states the moved + retention-deleted sum in the in_tx multi-statement frame (the sum the SQLite twin returns — the doc was the only liar); notify.rs's closed-store listen error routes through the shared database_error helper; bridge_capacity() is a const with its rationale doc carried. No behavior change — identical SQL strings, error shapes, and capacity. (task pg-fix-dedupe-cleanup, review 002 minor notes + Finding 6 queue bullet)
2026-10-10 05:04:09 +00:00
glm-5.3-flash
12d0497b1c
pg: scheduler runner resilience — quarantine bad rows, retry transient ticks
...
Review 002 Finding 5 (+ Finding 6's scheduler doc bullet):
- tick: a due row whose stored spec fails @every re-parse is
quarantined, not fatal — logged with the schedule name, boundary
advanced strictly past now (skip-forward at min interval), tick tx
still commits, remaining due rows proceed
- run_schedules: a pool/database tick failure retries 3x with a short
doubling backoff (250ms -> 1s cap) before the loop exits Err; the
top-of-iteration renew keeps owning the lease-loss decision
- exiting errors carry schedule-name/tick-phase context in the source
chain (ErrorContext wrapper; Database's Display is opaque)
- module docs: the Err-exit TTL-lapse posture stated; the false
per-slice soonest re-read claim corrected to the honest slice/idle
posture (60s idle floor)
Tests: behavioral quarantine pin (tampered via direct SQL UPDATE;
runner survives to clean Ok(()), other schedule fires, bad row never
fires, boundary advanced) and a server-less retry-policy pin.
Verified: cargo test -p alkstore-postgres green against the harness
server (121+10+9), build/clippy -D warnings/fmt green server-less.
2026-10-10 04:58:43 +00:00
glm-5.3-flash
ace94f0e13
pg forwarder: generation-tagged commands — a reconnect drops its predecessor's queued commands before replaying, closing the stale-UNLISTEN window (a stale UNLISTEN replayed after the snapshot's re-issued LISTENs silently cancelled a re-registered channel, its wakes lost until the next reconnect); the reconcile decision is the pure helper (stale LISTEN and UNLISTEN dropped, current-generation commands replay in queue order, dropped stale LISTENs ack the transient mid-LISTEN error) pinned by a four-combination server-less unit test (replay-proven against the neutered shape); 'queued commands replay harmlessly' doc corrected and the post-reconnect LISTEN-set invariant (server LISTEN set = registry snapshot, dead-generation commands can't undo it) stated in the module + loop docs (task pg-fix-stale-unlisten, review 002 Finding 4)
2026-10-10 04:49:35 +00:00
glm-5.3-flash
86719a39cc
pg open path: validate max_size > 0 at entry (typed Database before any round trip — 'max_size: 0' previously hung open forever at the bootstrap checkout, deadpool 0.13/0.14 neither validates nor defaults timeouts) and append the engine's '-c synchronous_commit' SET to the DSN's parse-carried options (was: setter replaced them, a silent override); pins: the zero-guard test (bounded by inner timeout, server-less-capable — fires pre-connect) and the DSN-options coexistence test (consumer SHOW statement_timeout + engine SHOW synchronous_commit, both settings) (task pg-fix-open-path, review 002 Finding 3 + options note)
2026-10-09 23:00:35 +00:00
glm-5.3-flash
9a5d1f4705
pg tx producer paths wake commit-atomically: publish_with_key_tx/enqueue_tx/outbox_enqueue_tx issue pg_notify on their mechanism-named channel (stream name / queue name / derived backing queue) inside the caller's tx — empty payload, best-effort log-and-swallow, no double-wake on the auto-commit paths; shared tx::wake_tx owner + module-doc section pinning the semantics; new tx_tests harness test pinning pre-commit silence, commit delivery on all three channels, and rollback silence; F-1 engine arm retired in review-wave-4 + implementation.md records (tx_publishes_compose_with_the_handle deterministic: 20 solo runs green; pg suite 116/116 x2 vs harness) (task pg-fix-tx-wake, review 002 Finding 2)
2026-10-09 22:38:38 +00:00
glm-5.3-flash
7ae426a01d
pg forwarder: the reconnect arm retries failed connects until success or shutdown — one failed connect no longer kills the loop permanently; every loop exit path releases the fanout sender via the shared-slot drop guard (receivers-terminal-iff-loop-gone); reconnect-config test seam + failed-connect pins: six-cycle exhausted-backoff server-less unit, pre-flip abort, guard drop, and the harness end-to-end unreachable-outage → full-recovery test (bug replay-proven against the old shape) (task pg-fix-forwarder-reconnect, review 002 Finding 1)
2026-10-09 22:31:05 +00:00
glm-5.3-flash
fb37da617d
Postgres engine integration: StoreFactory (fresh schema per open, owned idempotent CASCADE teardown, isolation/idempotence pinned), the engine's backlog column (all ten rows green against the harness server — exemplar verified, the three ADR-023 rows verified not rewritten, the five engine-scoped rows, the new pg-arm PayloadTooLarge row discharging the SQLite task's deferred adoption), test-observation accessors cfg(test)-gated with the unused PgStore::new cut, stale stub-era doc text removed, lib docs stating the finished-engine posture (task pg-engine-integration)
2026-10-09 10:22:25 +00:00
glm-5.3-flash
77619c5e93
Postgres engine: scheduler + outbox — schedule (validation triad, the pg-owned @every-only parser, upsert over the schedule table), unschedule, run_schedules (leadership via the engine's lock machinery on __alkstore_scheduler with a per-instance owner token, in-sleep lease renewals, the row-locked FOR UPDATE tick in one pool tx — fire enqueues + boundary advance + soonest read commit together — the 64-boundary catch-up cap with skip-forward, clean stop / Err(LeadershipLost) arms), outbox (validated constructor, enqueue into the derived __alkstore_outbox:{name} with the 60/5/5 stamps + max_attempts override, run_once ack/retry-curve/false over the ordinary claim machinery, no engine-issued heartbeat) (task pg-engine-scheduler-outbox)
2026-10-09 09:53:48 +00:00
glm-5.3-flash
c3591c2d44
Postgres engine: named locks — try_lock (validated entry, duration guard, opportunistic expiry-delete + insert-or-reacquire + holder read-back over the locks table), PgLockHandle (full-window renew, consuming owner-scoped release with both boolean arms), same-owner re-acquire matched to the SQLite arm, silent-lapse posture pinned, second open re-acquires after expiry (task pg-engine-locks)
2026-10-09 09:17:39 +00:00
glm-5.3-flash
3dd83791ff
Postgres engine: queues — Queue (validated constructor over the handle's QueueOpts stamps), the FOR UPDATE SKIP LOCKED claim (one statement, the pre-claim exhausted-reclaimable sweep in the atomic claim frame), JobHandle (one-shot ack/retry/fail in tx frames under the uniform validity predicate, absolute-reset heartbeat, engine-side equal-jitter backoff), dead-letter moves transactional (the #133 class excluded), worker-less ack_batch, unconditional cancel, dead-visible get_job, both-states+retention sweep, best-effort queue-channel wake, wake-driven claim loop pinned (task pg-engine-queues)
2026-10-09 08:44:59 +00:00
glm-5.3-flash
cb067bb4be
Postgres engine: streams — StreamHandle (auto-commit publishes + best-effort pg_notify wake, ASC reads with the extent guard, monotone offsets, pool-connection trim) and the durable subscribe receiver (async bridge, wake-driven re-drains, reconnect gap-heal, shutdown-only terminal close, stateless idle wake runtime for the sync save) (task pg-engine-streams)
2026-10-09 08:05:34 +00:00
glm-5.3-flash
8f5c2add5e
Postgres engine: LISTEN forwarder full behavior + notify/listen — wake contract pg arm
...
- notify: auto-commit pg_notify path, 8000-byte typed client-side check
through the tx seam's NOTIFY_PAYLOAD_LIMIT (one limit owner), closed-store
fail-closed before payload work
- listen: acked synchronous channel registration (listen starts-from-now —
a notify racing the LISTEN cannot be lost), refcounted ChannelSet
(UNLISTEN at last-subscriber drop), PgWakeReceiver bridging Wake { channel }
only, channel-scoped fanout + reserved reconnect-wake to every subscriber,
Lagged(n) surfaced-not-silent
- receiver close semantics (ADR-021 §5 pg arm): stays open across forwarder
reconnects, terminal None only at engine shutdown — Forwarder::shutdown
takes the fanout sender so receiver-held Arc lifetimes can't pin the
broadcast open
- tests: wake-arrives, 7999/8000/8002 boundary both entry points, no-replay
during connection gaps, backend-kill reconnect through the receiver stack,
drop-unregisters (behavioral probe — pg_listening_channels is per-session),
validation both paths, the two POC deadlock pitfalls re-pinned, 11 new
tests green against the harness server, gates green server-less
(task pg-engine-notify-listen)
2026-10-09 07:37:56 +00:00
glm-5.3-flash
cf5ceea70e
Postgres engine: transactional seam — begin_tx, PgTxHandle, all eleven *_tx methods with drop=rollback detached teardown, probe-pinned unknowable-state discard arms, engine-side resolution arithmetic (task pg-engine-seam-tx)
2026-10-09 06:50:53 +00:00
glm-5.3-flash
c6a7eeaa45
Postgres engine: open constructor, PgOpts, pool + listener wiring — forwarder skeleton with the POC-pinned pitfalls structurally excluded, seam error mappings, wave-3 stub surface (task pg-engine-open-opts)
2026-10-09 06:00:31 +00:00
glm-5.3-flash
2f1353bd41
Postgres engine: schema bootstrap — engine-owned schema, table family, idempotent DDL, schema-prefixed indexes (task pg-engine-schema)
2026-10-09 05:11:48 +00:00
glm-5.3-flash
34e0b9732d
Scaffold Cargo workspace: alkstore core + sqlite/postgres engine stubs (ADR-001)
2026-10-07 14:57:34 +00:00
glm-5.3-flash