Absurd
TypeScript · Python · Go (experimental)Checkpointed workflows that replay persisted steps after interruption.
A practical guide to Postgres-backed jobs
Choose a queue by the work it must protect. Explore job frameworks, durable workflows, visibility queues and event queues—with the evidence kept in view.
Benchmark runs · 2026-11-11 to 2026-11-15
9 adapters, one host, PostgreSQL 18.6. Each scenario answers one question; there is no overall score.
How many 1 ms jobs per second does one worker process complete as it gets more workers?
Tuned: River 10.9k and Awa 10.8k jobs/s; PgQue 47k events/s, producer-bound.
At a rate every system should sustain, how quickly does a job go from insert to completion?
Awa, Oban and PGMQ pick up within 20 ms p99; three systems fall behind.
With workers running and no jobs, how much does each system cost the database?
16–24 xacts/s for most; PGMQ polls at 160/s; Awa holds 276 MB of Postgres memory.
When the database restarts mid-run, does the system come back and work off what accumulated?
All eight recover; Absurd and Procrastinate need a supervisor to restart the worker.
Does a busy queue slow down unrelated transactions on the same server?
Neighbour throughput unchanged everywhere; p99 at 4× from 2.4 to 18 ms.
Does the queue coexist with CDC on the same database, including when the CDC consumer stops?
Awa 19 → 0 tables without replica identity; still the most catalog churn.
When 20,000 jobs come due at the same instant, how late do they start, and do immediate jobs suffer?
PGMQ 1.9 s, Awa 12 s, Oban 21 s to start a 20,000-job herd.
When a third of jobs start failing, does useful work keep flowing?
Every system kept the good jobs flowing through a failure storm.
What does spreading the same load over 200 queues cost, and how does each system serve many queues?
All keep up; idle cost at 200 queues from 0.02 to 1.7 Postgres cores.
What happens with 64 KiB and 1 MiB arguments, and does the database shrink back afterwards?
Awa shrinks back to 21 MB after 8 GiB; others keep 7–16 GB.
At a gentle rate with 64 KiB arguments, how much WAL is written and how much space is left behind?
Only Awa returns the space (22 MB vs 1.2 GB).
See the results →What does it cost to keep jobs running for 40 seconds each?
WAL to keep 40 s jobs alive: 4.6 KB/s (PGMQ) to 28 KB/s (Awa).
How fast does a 200,000-job backlog drain, and how big is the database afterwards?
PGMQ and PgQue drain at ~14.5k jobs/s; only Awa shrinks back.
With a million jobs queued, how long does an operator wait to find one?
River finds one job in 2 ms; Awa 0.6 times out; Procrastinate stalls behind the haystack.
See the results →When an operator pauses a busy queue, does work stop everywhere, and how fast does it come back?
A pause survives a restart only where it's stored: Awa, River, PgQue.
See the results →When an entire fleet dies holding about a million claimed jobs, how long until those jobs have run?
PgQue 16 s, PGMQ 5.9 min, pg-boss 14 min, Awa 20.5 min to finish 1M abandoned jobs.
The directory
14 systems · alphabetical order
Choose up to 8 to compareCheckpointed workflows that replay persisted steps after interruption.
Typed Rust and Python jobs with segmented storage and guarded attempts.
A Rails ActiveJob backend with a dashboard and concurrency extensions.
Node.js task executors, SQL enqueue and serial named queues.
An Ecto and OTP job framework with queues, retries and telemetry.
Node.js job queues with explicit queue policies and transaction adapters.
SQL messages with a visibility timeout and application-owned consumers.
The maintained original PgQ event engine, with tick-based batches.
A shared event log with consumer cursors and batch acknowledgements.
An alpha Python worker with Celery-shaped APIs and a Postgres transport.
Python sync/async tasks with application locks and configurable recovery.
A focused Ruby queue built around advisory locks and transaction control.
A PostgreSQL job framework with a shared format across language clients.
Rails job infrastructure with SQL dispatch and explicit failure records.
Try another language or reset your filters. Documentation-only does not mean unsupported.
Language labels describe the inspected API/SDK surface, not equal maturity or feature parity across bindings. Read each profile’s scope and snapshot.
Three kinds of evidence
Each answers a different question. None replaces the others.
Primary docs and pinned source, with qualifiers and unknowns visible.
Compare contracts →02 · Benchmark resultsSaturation, latency, restarts, noisy neighbours, CDC, large arguments and a million abandoned claims, run on one host. Grouped by scenario; no single score.
See the results →03 · Teaching simulationOperate a worker, stop it, and recover its job. This is a teaching model, not a benchmark trace.
Open the failure lab →