v0.6.1

Release 0.6.1 of akadze.

v0.6.0

Breaking

• A run that returns while a cancel is pending ends succeeded with its result, not cancelled (#52).

Fixes

• Limits hold when a claim waits on another claimer's lock (#46).
• The worker keeps serving through transient database errors; a failed outcome write does not rerun the job (#47).
• Task arguments arrive as their declared types (#48).
• A worker claims only tasks it has registered (#49). Upgrade akadze on all workers before adding new tasks.
• A sync task registered with timeout logs a warning (#50).
• 0.5.0 audit tail: jobs cancel exit code, migration checker, drain readiness, overlap=False docs (#51). Migration 005_jobs_meta builds a GIN index that blocks writes to akadze.jobs; on a large table create it CONCURRENTLY first (see docs).

Performance

• Claim reads the queue in windows: fair claim and claims past a key at its limit go from 90–310 ms to 3–18 ms on 100k jobs (#53).

Docs

• Documentation site with a Russian translation: https://g4st3r-d3v.github.io/akadze/ (#54, #55). Version banners updated for the release in this PR.

---

Note

Low Risk
Version and documentation-only changes with no application or worker logic in the diff.

Overview
Releases akadze 0.6.0 by bumping __version__ and Poetry version from 0.5.0 to 0.6.0.

Documentation is updated so readers no longer see “main is ahead of PyPI (0.5.0)”. The home pages (EN/RU) now say docs track main, point “since 0.6.0” markers at behavior from that release, and link to GitHub release notes for upgrades. Claim docs (EN/RU) state windowed claim ships in 0.6.0 and later instead of “first release after 0.5.0”. Cancel-on-running-job notes in the retries guides (EN/RU) drop the awkward “on PyPI” wording while keeping the 0.6.0 vs 0.5.0 behavior contrast.

^{Reviewed by Cursor Bugbot for commit 23e74aa. Bugbot is set up for automated code reviews on this repo. Configure here.}

v0.5.0

akadze 0.5.0

Product features for multi-tenant and scheduled workloads, and a much cheaper claim. Run akadze migrate after upgrading (adds an index; see below).

New

• Cron in a timezone — @app.periodic(..., timezone="Europe/Berlin"), also in add_schedule / schedule specs. DST follows classic cron: fixed-hour entries fire once when clocks go back, wildcard-hour entries keep firing; skipped wall times fire shifted by the gap. (#38)
• Backoff policy — @app.task(..., backoff=fn) / app.register / @app.periodic, fn(attempt) -> timedelta; type akadze.BackoffPolicy. (#39)
• Bulk enqueue — task.using(...).enqueue_many(items): one statement per 1000 rows, all-or-nothing, in the caller's transaction (10 000 jobs ≈ 0.07 s). (#40)
• Job meta — using(meta={...}), list_jobs(meta=...), akadze jobs list --meta k=v, JobSummary.meta (read-only), Job.meta in hooks. (#41)
• Fair claim — Worker(fair=True) / akadze worker --fair: within a priority, a free slot goes to the limit_key group with the fewest running jobs. (#42)

Changed

• Default retry backoff now has jitter in [50 %, 100 %] of 2^(attempt-1) s (cap 300 s). Jobs requeued from a lost worker get the same jittered delay instead of all starting at once. (#39)
• A negative Retry/Snooze delay means "now" instead of failing the finish; policy delays are clamped to [0, 365 days]. (#39)
• The default claim reads the head of the queue first: ~18 ms instead of ~330 ms per claim on a 100 000-job backlog. Fair claims still scan the queue. (#43)
• tzdata is a runtime dependency. (#38)

Migration

• 005_jobs_meta builds a GIN index on akadze.jobs.meta. It blocks writes to akadze.jobs while it builds; on a large table pre-build it with CREATE INDEX CONCURRENTLY IF NOT EXISTS jobs_meta_idx ON akadze.jobs USING gin (meta jsonb_path_ops).

Docs

• Public surface and versioning policy are pinned by a test of akadze.__all__. (#44)

v0.4.1

Release 0.4.1 of akadze.

v0.4.0

Alpha DX для замены Celery: расписания из конфига приложения и регистрация задач без global bind.

Новое

• app.add_schedule / app.add_schedules — cron/every для уже зарегистрированной задачи; YAML/TOML парсит приложение, не пакет
• app.register(fn, name=…) — unbound-регистрация; @app.task идёт тем же путём
• README: статус alpha (0.3.x→0.4.x) и короткий stub миграции с Celery (unique_key, limit_key, using(...).enqueue / complete_tx, Snooze)

Не ломает

Аддитивный релиз относительно 0.3.0. Новых SQL-миграций нет (001–004 как раньше).

Ops CLI (jobs / queues / limits / workers / metrics) без изменений с 0.3.0.

v0.3.0

Операционный CLI поверх Wave 3 и одно ломающее изменение в limit_key.

Ломающее относительно 0.2.0

Если у задачи задан limit_key, а строки с этим ключом в akadze.limits нет, claim больше не блокирует задачу. Воркер считает, что лимита нет. В 0.2.0 такая задача оставалась в queued навсегда.

clear_limit теперь реально снимает ограничение: удаляет строку из akadze.limits, и задачи с этим ключом снова берутся в работу.

Новой миграции схемы нет. Достаточно уже накатанных 001–004.

Новое

• akadze jobs cancel JOB_ID — ставит запрос на отмену через request_cancel
• list_limits / clear_limit и CLI akadze limits list, akadze limits clear KEY
• list_workers и CLI akadze workers list
• CLI akadze metrics (и akadze metrics render) печатает текст из render_metrics
• Akadze.include("pkg.mod") — импортирует модуль и вызывает register(app)

Оператору

akadze limits list
akadze workers list
akadze metrics

Переменная окружения та же: AKADZE_DATABASE_URL.

v0.2.0

Wave 3 ops and cluster limits on top of 0.1.x.

• list_jobs / akadze jobs list (no args/result in summaries)
• requeue for failed jobs; pause_queue / resume_queue
• Cluster concurrency via limit_key + set_limit(max_running=…)
• Start rate via set_limit(max_starts=…, window=…)
• akadze.metrics.render_metrics → Prometheus text (no HTTP)

Schema migrations: 002_queue_pauses, 003_limits, 004_start_rate (kept separate).

v0.1.1

Release 0.1.1 of akadze.

v0.1.0

Release 0.1.0 of akadze.