---
id: pg-engine-scheduler-outbox
name: Postgres engine — scheduler (`schedule`/`unschedule`/`run_schedules`) and outbox helper
status: pending
depends_on: [pg-engine-queues, pg-engine-locks]
scope: broad
risk: high
impact: component
level: implementation
tags: [wave-4, postgres-engine]
---

## Description

Implement the scheduler collapse's pg arm (ADR-009) and the outbox
helper (ADR-014/019 §5) — the wave's last mechanism pair, riding the
queues task's claim machinery and the locks task's lock machinery:

**Scheduler** (over the schema task's schedule table):

- **`schedule(name, spec, queue, payload, opts)`** — upsert by name;
  entry-point validation on all three name-bearing arguments
  (schedule name like a queue name; the **queue argument** identically
  per ADR-021 §3 — a schedule cannot target a reserved name, so the
  outbox backing queue stays reachable only through the outbox
  surface); payload encoded at the seam (`Codec` propagation); the
  spec grammar is `@every <n><unit>` only (`s|m|h|d`) — **the pg
  engine owns its parser** (no shared substrate parser on this side;
  the grammar is small and pinned; cron strings → `InvalidSpec`).
  Rejection is pre-storage (`InvalidSpec` before any round trip).
- **`unschedule(name)`** — the unregister; `Ok(false)` when absent.
- **`run_schedules(stop: StopToken)`** — the leader loop, the
  structural twin of the SQLite engine's runner with the pg deltas:
  - **Leadership via the engine's own lock machinery** on the
    reserved name `__alkstore_scheduler` (ADR-009 §1) — `try_lock`
    with a per-runner-instance-unique owner token (the wave-3
    finding's lesson: a shared constant would make two processes each
    see "self" holding; the token is pid + subsec-nanos or
    equivalent). TTL engine-internal (10 s), renewed at the top of
    every loop iteration **and** on a cadence across the sleep (the
    wave-3 posture: without in-sleep renewals, a boundary further out
    than the TTL lapses its own lease by drift). A failed mid-sleep
    renewal is not acted on at the sleep site — the loop's
    top-of-iteration renew owns the loss decision.
  - **Acquire-time refusal is `Err(LeadershipLost)`** — the same
    matchable loss a mid-run loser returns (the wave-3 posture: no
    runner-loop exists without the lock; the respawn recipe is
    identical either way).
  - **The tick**: fire enqueues + boundary advance commit together —
    on pg this is one pool-checked-out transaction (`BEGIN` …
    `COMMIT` on the held object, not `BEGIN IMMEDIATE` — pg's
    default isolation gives the row-locked fire tx ADR-009 §4
    names; the schedule row's `SELECT … FOR UPDATE` inside the tick
    tx is the row lock that serializes rogue second tickers). The
    tick body: boundary math over the registered specs (the
    **one-firer-per-boundary** guarantee — at-least-once per elapsed
    boundary, the **64-boundary catch-up cap with skip-forward**
    past the cap, honker parity), each fire an ordinary enqueue
    stamped with the `ScheduleOpts` resolved over the target queue's
    *derived engine defaults* (no handle is open — ADR-020 §3), then
    the soonest-boundary read for the sleep deadline.
  - **Wake-driven tick advance** (the pg delta over the SQLite
    runner's pure-sleep shape): the runner parks on the sleep
    deadline, but an enqueue-wake on the scheduler's own wake channel
    (or a plain bounded-wait interrupt — implementer's choice,
    documented) may shorten the sleep so a newly-registered schedule
    is noticed promptly; correctness never rides the wake (the
    deadline always fires; the wake only shortens idle latency —
    the SQLite runner's 60 s idle-tick posture is the acceptable
    floor if the wake path proves fussy; flag the choice in Notes).
  - **Clean stop** (`StopToken::cancel`) → `Ok(())`, the leadership
    lock released cleanly; **loss** → `Err(LeadershipLost)` before
    any tick.
- **Boundary math is this engine's own implementation** (ADR-012 §2's
  one-owner rule — the substrate's `scheduler_tick`/`scheduler_soonest`
  are SQLite-side): the `@every` interval arithmetic, the boundary
  advance, the 64-cap catch-up, and the skip-forward. The contract
  suite (wave 5) pins the two engines' boundary behavior to identical
  outcomes.

**Outbox** (a helper over queues — no separate mechanism):

- **`Store::outbox(name)`** — validated constructor (empty →
  `InvalidName`, reserved-prefixed → `ReservedName`); carries no opts
  of its own (the backing queue's derived 60/5/5 set is
  engine-documented, ADR-010 §3a).
- **`enqueue(payload, opts)`** — the auto-commit path into the
  derived backing queue: derive `__alkstore_outbox:{name}` (the
  reserved prefix makes it unreachable by `enqueue_tx`/`queue()` by
  design), stamp the 60/5/5 set with the `EnqueueOpts::max_attempts`
  override, full ADR-020 resolution, ride the queues task's enqueue
  machinery.
- **`run_once(worker_id, delivery) -> bool`** — the pull op: claim
  one job from the backing queue via the ordinary claim machinery
  (the queues task's), run the delivery callable (core's `Delivery`
  trait object — ADR-019 §5's pinned shape), `ack` on `Ok`,
  `retry(err, None)` (the queue's curve; the delivery error's
  `Display` string as the retry's exhaust string — the wave-3
  posture) on `Err`, `false` on empty. **No engine-issued heartbeat
  inside delivery** (honker parity, inherited deliberately — the
  boxed `JobHandle`'s own `heartbeat` remains available to the
  closure); the boxed `JobHandle` handed to the caller.
- **`outbox_enqueue_tx`** landed with the seam task; this task's
  tests may extend its coverage (the commit-atomicity row is
  suite-owned in wave 5).

## Acceptance Criteria

- [ ] `schedule`: validation triad (name/spec/queue), `@every`-only
      grammar (cron → `InvalidSpec` pre-storage), upsert read-back;
      the pg-owned parser pinned by tests (accepts/rejects the
      grammar's exact surface)
- [ ] `run_schedules`: leadership via `__alkstore_scheduler` with a
      per-instance owner token; acquire-time refusal =
      `Err(LeadershipLost)`; tick fires + boundary advance in one
      transaction (row-locked — a rogue second ticker cannot
      double-fire); loss returns `Err(LeadershipLost)` before any
      tick; clean stop releases the lock; in-sleep lease renewals
      hold across boundaries further out than the TTL
- [ ] One-firer-per-boundary: two runners on the same db, exactly one
      fire per boundary (the band test); 64-cap catch-up +
      skip-forward with no in-window refire
- [ ] Fired jobs stamp the derived defaults resolved over
      `ScheduleOpts` (`get_job`-visible)
- [ ] `outbox`: constructor validation; backing queue unreachable by
      `enqueue_tx`/`queue()`; 60/5/5 stamps + `max_attempts`
      override; `run_once` ok/err/false dispositions with the curve
      gap asserted; in-delivery heartbeat renewal available;
      exhaustion to dead with the delivery error string
- [ ] `cargo test -p alkstore-postgres` (harness server), clippy
      `-D warnings`, fmt clean; gates green server-less

## References

- docs/architecture/decisions/009-scheduler-collapse.md (§1–§6)
- docs/architecture/decisions/014-outbox-tx-enqueue.md
- docs/architecture/decisions/019-mechanism-handle-surfaces.md §4, §5
- docs/architecture/decisions/020-enqueue-opt-semantics-and-bridges.md §3
- docs/architecture/decisions/021-tx-reads-and-value-shape-fixes.md §3
- docs/architecture/engine-postgres.md (Mapping the contract: scheduler/outbox)
- docs/architecture/queues.md (scheduler section)
- alkstore-sqlite/src/scheduler.rs, outbox.rs (the structural twin —
  same semantics, own boundary math; its Notes carry the flake-fix
  lessons: drive catch-up tests from far-out horizons so no live
  boundary races the assertion window)
- alkstore/src/stop_token.rs, outbox.rs (the traits)

## Notes

> To be filled by implementation agent

## Summary

> To be filled on completion