Repository navigation
1.0.0-rc.3
Pre-releaseThe last re-cut of the baseline before 1.0.0. Two features the data model was missing land on the existing lock store, the persisted model gains the row-shape constraints it lacked, the claim seeks on PostgreSQL instead of sorting, five defects on the recovery and completion paths are fixed, and rolling deploys become a documented shape. An alerting redesign is deferred, with its design written down.
Upgrade: drop and reprovision the Acta database, on every provider. M001 is re-cut: definitions gains the concurrency and rate policy columns, jobs.exclusive_key is renamed to concurrency_key, five CHECK constraints are added and one tightened, and two SQL Server indexes gain a physical option. The baseline stamp is now a content hash of each provider's emitted M001; a database from any earlier build refuses to start with a message naming both stamps. There is no upgrade path between baseline generations before 1.0.0.
Highlights
- Concurrency limits.
[Job(ConcurrencyLimit = N)]bounds how many jobs of a definition run at once, admitted through slot rows in the lock store; the exclusive key becomes the concurrency key, one name everywhere. - Rate limits.
[Job(RateLimit = "600/m")]meters a definition through reservations: an idle meter admits one second's worth of the rate at once and then one per interval, every early job books a turn it returns to, a booked turn stays valid for a second, and an executor sleeps out a turn under a quarter second in process. The contract is the upper envelope,R*(T+W)+Badmissions in any window of T seconds. - Rolling deploys as three rules and no schema change. Registration never retires a definition the manifest omits, so the old fleet finishes its work and a rollback finds nothing missing. A worker that claims a job it has no handler for hands it back budget-neutral and excludes that definition from its own claims for the rest of its process. Retiring a definition and cancelling its queue is an operator verb:
IActaOperations.Definitions.RetireAsync,POST /definitions/{ns}/{name}/retire, and a confirmed button on the dashboard. - The claim seeks on PostgreSQL. A Ready row always carries its due instant, so the claim reads
ix_runtimes_claim_readyin index order instead of sorting; the full matrix reads 24% to 45% more Bulk throughput than rc.2 on PostgreSQL. - Durable state is bound to CHECK constraints on every provider, and two step routines they exposed are fixed.
- Five recovery and completion defects. A claim the worker has no handler for is released instead of held under a renewed lease; the completion write and the start write retry on a provider error; a committed-then-lost start is reconciled through the completion; the alerts projector cursors by
(created_at_utc, id)so an inverted event is never skipped; recovery runs through one pass, and every worker checks the recovery slot every seven minutes. - SQL Server: page-latch waits attributed per index, and the two insert convoys opted into sequential keys.
- The ensemble certification workload carries a metered shape, and every seal reports the rate contract beside the chaos figures.
Evidence
- Four chaos-certification gates on the certified commit, all PASS with real chaos, sealed under
docs/certification/: PostgreSQL and SQL Server at 10,000 jobs and 64 slots (576 and 574 orphaned attempts, 72 dead workers each), SQLite reduced at 48 slots (576 and 72), and the three-participant, two-namespace ensemble with 772 orphaned attempts across 162 killed workers. On every gate the rate meter admitted its 600/m shape at the declared rate, never past 21 in a second against a budget of 30. - A full-matrix benchmark round on all three providers against a same-hour rc.2 control from clean worktrees, recorded in
docs/benchmarks/rc3-full-matrix-20260921.md: no Direct, Bulk, drain, enqueue, or query cell regresses beyond the machine's own movement between two runs of the same tree; PostgreSQL and SQLite read ahead on most of them.
Full notes: docs/release-notes.md.