Repository navigation
1.0.0-rc.2
Pre-releaseA correctness round over the paths a release candidate has to get right before anyone leans on them: crash recovery, the audit trail, and the bounds on what an operator or an attribute can declare. Nothing here adds a feature to the execution model; most of it makes a guarantee the docs already claimed actually hold on every provider.
Upgrade: drop and reprovision the Acta database, on every provider. The M001 baseline was re-cut again and the stamp now names the day it was cut, baseline-20260910; the bootstrap refuses any history without this build's stamp row. There is no upgrade path between generations before 1.0. The re-cut carries a Unicode events.actor_key, three new check constraints, a re-keyed in-flight index, and expected_version on SQL Server's job_schedule_advance_batch table type.
Highlights
- A crash can no longer take a namespace's recovery down with it. The reclaim sweep ran only from the
sys.recoveryslot, and a worker could die holding it. Every worker now checks that one slot at startup and once an hour, with a single guarded statement that re-arms it only when its lease has lapsed; two workers checking at once commit one repair. The round that found this had 571 jobs queued behind a slot whose lease had expired 36 minutes earlier. - A recurring completion honours what happened to its schedules while it ran. A schedule the deployment removed stays orphaned, a schedule edited during the attempt keeps its edit, and the slot re-arms from the schedules as they stand. Each planned advance carries the schedule version it was read at.
- Restarting an unchanged build writes nothing to its schedules. Registration updates a slot or a schedule only when the declaration changed, so a fleet restart or a scale-out no longer bumps every version in the namespace.
- A host starting no longer cancels a slot another host is running. Re-registration skips a slot held under a live lease.
- SQL Server: no more table-lock escalation on
runtimes, and the heartbeat no longer deadlocks againststart_execution.FORCESEEKon four updates, and a heartbeat that locks in clustered order. The rc.1-era throughput sweep did not rise with executors because some cells compiled the escalating plan; every profile now rises monotonically, with 17 of 18 cells above that baseline. - The audit trail keeps an operator's name as typed.
actor_keyis Unicode on every provider; a job note files under the attempt that wrote it. - The job control verbs take an optional
expectedVersion. Cancel, pause, resume, restart, reschedule and reprioritize can be guarded against a row that moved; a conflict answersVersionConflictwith the current version and writes nothing. - Waits and clocks are anchored where they are read, and fractional retention and dead-worker windows round up rather than down.
Evidence
- Four chaos-certification gates on the release tree, all PASS, sealed under
docs/certification/: PostgreSQL and SQL Server at 10,000 jobs and 64 slots, SQLite reduced at 48 slots, and the three-participant, two-namespace ensemble with 830 orphaned attempts across 152 killed workers. - One idle benchmark round per provider against the rc.1-era matrix, recorded in
docs/benchmarks/rc2-benchmarks-20260910.md, with what a single round does and does not establish stated on the page. - Coverage baseline
coverage-baseline-rc2.md: 88.4% line, 70.7% branch, and the blind spot the round's own defect landed in still open.
Full notes: docs/release-notes.md.