tinyjobs

Background jobs and cron for Python, backed by SQLite, with no dependencies.

pip install tinyjobs

That's the whole setup. No broker, no result backend, no daemon. The queue is a file.

Celery is the right tool for a lot of systems, but it asks for a Redis or RabbitMQ instance before it will run a single job. This is for the projects that never needed that: a Django app that sends emails, a scraper with a nightly refresh, a CLI that offloads slow work. When you outgrow SQLite you swap the backend for a plugin and your task code doesn't change.

Requires Python 3.11 or newer.

Quick start

# tasks.py
from tinyjobs import TinyJobs

app = TinyJobs("sqlite:///jobs.db")


@app.task
async def generate_pdf(document_id: int) -> str:
    ...


@app.task(max_attempts=5, timeout=300)
def transcode(video_id: int) -> None:
    ...

Enqueue from anywhere, including plain sync code:

from tasks import generate_pdf

job = generate_pdf.enqueue(123)
print(job.id)          # 01a061bf-c08a-738d-...

enqueue gives you back a snapshot of the row it just wrote, not a live handle, so the Job you are holding keeps saying queued: the worker that runs your task is a different process writing to the database, and nothing reaches back into your object. Ask for the current state when you want it:

job.refresh()          # updates this instance in place, and returns it
job.status             # queued -> running -> success
job.attempts           # 3, if it took three tries
job.last_error         # 'ConnectionError: refused', when it failed

app.get(job.id)        # or a fresh object, if you prefer not to mutate

Attribute access never does I/O on its own. refresh() is the one place a round trip happens, which keeps for job in jobs: print(job.status) from quietly becoming a thousand queries.

If the task returns something you want back, ask for it to be kept:

@app.task(store_result=True)
async def generate_pdf(document_id: int) -> str:
    return f"/tmp/{document_id}.pdf"

job = generate_pdf.enqueue(123)
...
job.refresh().result()     # '/tmp/123.pdf'

Results are off by default because most jobs are run for their side effects, and storing return values nobody reads is write amplification on the hot path.

Keeping the database from growing forever

Terminal jobs are pruned on a schedule. Successes and cancellations are kept for a week, failures for thirty days, because failures are the ones you come back to:

app = TinyJobs("sqlite:///jobs.db", retention=RetentionPolicy(failed=90 * 86400))
app = TinyJobs("sqlite:///jobs.db", retention=KEEP_EVERYTHING)   # prune nothing

tinyjobs worker --app tasks:app --retain-success 7d --retain-failed 30d
tinyjobs worker --app tasks:app --no-purge

tinyjobs purge --app tasks:app --before 30d --dry-run

Every worker checks whether a sweep is due, and exactly one of them wins, because claiming the sweep is a single conditional UPDATE on a marker row. So running forty workers still means one sweep per interval, with no leader to elect. Queued and running jobs are never touched at any horizon, and each sweep deletes in small batches with a cap, so it never holds the write lock long enough to stall a claim.

Note that pruning does not shrink the file. SQLite reuses the freed pages, so the database plateaus rather than shrinking. Handing space back to the filesystem needs VACUUM, which wants exclusive access, so it stays a manual thing.

If you want to block until a job is finished, wait polls for you:

job = generate_pdf.enqueue(123)
print(job.wait(timeout=30).result())     # '/tmp/123.pdf'

await job.awaited(timeout=30)            # from async code

It returns the job whatever the outcome, rather than raising, so a failed job leaves you holding job.status, job.attempts and job.last_error to look at. Only the timeout raises, as WaitTimeout, which is also a TimeoutError if you would rather catch that.

The async twin is called awaited and not await for the boring reason that await is a keyword.

Waiting on a job from the request that created it means you did not need a background job. It is meant for scripts, tests and glue code.

Or from async code, which is the native path:

job = await generate_pdf.aenqueue(123)

Run a worker:

tinyjobs worker --app tasks:app --concurrency 8

tinyjobs is installed as a command, but python -m tinyjobs does the same thing and is handy when you have not activated a virtualenv:

python3.13 -m tinyjobs worker --app tasks:app --concurrency 8

Look at what happened:

tinyjobs jobs list --app tasks:app
tinyjobs jobs list --app tasks:app --status failed
tinyjobs jobs show <job-id> --app tasks:app

jobs show prints every attempt, not just the last one:

attempts
    1  failed        4ms  laptop/8821/9389541d  ConnectionError
    2  failed        3ms  laptop/8821/9389541d  ConnectionError
    3  success      12ms  laptop/8821/9389541d

There's a runnable example in examples/demo.py covering retries, timeouts, priorities, delays, deduplication and cron:

python -m examples.demo
tinyjobs worker --app examples.demo:app --concurrency 4

Cron

@app.task(cron="*/15 * * * *")
async def refresh_cache() -> None:
    ...


@app.task(cron="30 2 * * *", timezone="Europe/Rome")
def nightly_report() -> None:
    ...

There is no extra process to run. Every tinyjobs worker also runs the scheduler, and however many workers you start, each firing creates exactly one job: the schedule moves on with a single conditional UPDATE, and only the worker whose update matched writes the job, in the same transaction. Kill any of them, even with kill -9, and the rest carry on.

What it creates is an ordinary job, with the task's retries, queue and priority.

tinyjobs schedules --app tasks:app

NAME                  CRON          TIMEZONE     NEXT RUN             LAST RUN             LAST JOB
--------------------  ------------  -----------  -------------------  -------------------  ------------------------------------
tasks.nightly_report  30 2 * * *    Europe/Rome  2026-10-11 02:30:00  -                    -
tasks.refresh_cache   */15 * * * *  UTC          2026-10-10 14:15:00  2026-10-10 14:09:00  01a125b7-6b62-7135-b9a0-b7707bf39c1c

tinyjobs worker --app tasks:app --no-scheduler     # runs jobs, fires nothing
tinyjobs scheduler --app tasks:app                 # fires schedules, runs nothing

When nothing was running at the moment a schedule was due, misfire decides what happens on restart:

@app.task(cron="0 * * * *")                                   # skip, the default
@app.task(cron="0 * * * *", misfire="run_once")               # one catch-up run
@app.task(cron="0 0 * * *", misfire="run_all", overlap=True)  # every missed day, oldest first

A firing only counts as missed when it is more than a minute late (--misfire-grace). A schedule won't start a run while the previous one is still queued or running, unless overlap=True.

A malformed expression, an unknown timezone, 0 0 30 2 * (which never fires) or a cron task that needs arguments all raise at import, not at 2 a.m.

Worth knowing:

• The resolution is one minute.
• An async task that blocks the event loop also blocks its worker's scheduler, which then skips the firings it slept through. Write blocking work as def, or run a standalone tinyjobs scheduler.
• Daylight saving. A time of day (30 2 * * *) fires once, even when 02:30 happens twice or not at all. A frequency (*/10 * * * *) follows the real clock: the repeated hour runs twice, the skipped one not at all. Local timezones on Windows need pip install tzdata.
• Clocks that disagree by less than the grace are harmless. A jump forward counts as an outage; a jump back is noticed on the next tick.
• One schedule per task name. For two cadences, register the function twice under two names.
• Volume. Ten thousand schedules on the same minute all fire exactly once, the last few seconds late. Numbers in Measured.

Watching it

@app.on_job_failure
def alert(job, error, will_retry):
    if not will_retry:
        sentry_sdk.capture_exception(error)


@app.on_job_success
def timing(job, result, duration):
    statsd.timing(job.task, duration)

on_job_start(job) completes the set. Hooks run in the worker and can be def or async def. One that raises is logged and ignored, so a broken metrics client never fails a job. A def hook runs on the worker's event loop, so keep it quick.

tinyjobs stats --app tasks:app
tinyjobs check --app tasks:app --max-wait 5m      # exits 1 on any problem

QUEUE    QUEUED  RUNNING  SUCCESS  FAILED  CANCELLED  OLDEST DUE
-------  ------  -------  -------  ------  ---------  ----------
default  2       2        0        0       0          16s
mail     1       0        0        0       0          16s

running past their lease: 2

A job past its lease means the worker holding it stopped reporting, usually because it died. check fails on that, on queued jobs for a task the app doesn't register (a missing --import), on a due job waiting longer than --max-wait, and on a cron schedule more than a minute overdue, which means no worker or scheduler is running. Both are app.stats() and app.check() in code.

Testing your tasks

app = TinyJobs("memory://", eager=True)

job = generate_pdf.enqueue(123)     # runs before enqueue returns
job.status                          # success
job.result()                        # '/tmp/123.pdf'

Or keep the queue and run it when you choose, with a clock you control:

from tinyjobs import FrozenClock

clock = FrozenClock(time.time())
app = TinyJobs("memory://", clock=clock)

flaky.enqueue()
app.drain()             # runs everything due, highest priority first; returns how many ran
clock.advance(10)       # the retry's backoff
app.drain()

memory:// is the SQLite backend on a private in-memory database, so tests run the same claim, dedup and retries as production. Delays and backoff are respected in eager mode too: advance the clock and drain. A failing task is recorded on the job, not raised, as it would be in a worker.

What you get

Async and sync tasks	async def runs on the worker's event loop, def in a thread pool
Retries	exponential backoff with jitter, per-exception policy, Retry and Abandon
Cron	cron="*/15 * * * *", timezones and DST, misfire and overlap policies, in every worker
Delayed jobs	delay=60 or run_at=<datetime>
Priorities and queues	integers, highest first; workers subscribe to a subset
Timeouts	hard for async tasks, soft for threaded sync ones (see below)
Crash recovery	leases plus heartbeats, so a killed worker's jobs come back
Deduplication	key="invoice:123" collapses duplicate enqueues
Attempt history	every attempt is a row, with its traceback and timings
Multi-worker	multiple processes and machines over one file or backend
Async throughout	the backend protocol is async, with a sync facade for ordinary code
Cancellation	before a job starts; cooperative after
Retention	old terminal jobs are pruned so the database stops growing
Hooks	on_job_start, on_job_success, on_job_failure, for metrics and alerts
Health	tinyjobs stats, and tinyjobs check for probes
Testing	memory://, eager=True, app.drain() and a frozen clock

Delivery is at-least-once. A worker can finish a job and die before recording it, in which case the job runs again. Write your tasks so that running twice is the same as running once. There's no way around that without a transaction spanning both the queue and whatever your task touches, and anything claiming exactly-once is either lying to you or a great deal more complicated than this.

Two things are worth knowing before you hit them:

Timeouts on sync tasks are advisory. You can't kill a running Python thread. When a threaded task exceeds its timeout the job is marked failed and the worker moves on, but the thread keeps going until the function returns. Pass timeouts to the library you are calling, or write the task as async def where cancellation actually works.

Arguments are serialized as JSON, extended with tags for datetime, date, Decimal, UUID, Path, set, tuple, bytes and non-string dict keys. Pickle is available behind TinyJobs(..., allow_pickle=True) but it's off by default, because deserializing a pickle runs arbitrary code and the roadmap includes backends that reach over a network. Dataclasses and enums need one line of registration:

app.codec.register_dataclass(Money)
app.codec.register_enum(Priority)

How it works

jobs.db holds two tables. jobs is the queue, executions is the history of attempts.

A worker claims work with a single statement, which is what makes it safe to run as many workers as you like:

UPDATE jobs SET status = 'running', worker_id = ?, attempts = attempts + 1, lease_expires_at = ?
 WHERE id IN (SELECT id FROM jobs
               WHERE status = 'queued' AND run_at <= ? AND queue IN (?)
               ORDER BY priority DESC, run_at ASC LIMIT ?)
RETURNING *;

Claiming a job takes a lease. The worker refreshes it while the job runs, and if the worker dies the lease expires and another worker picks the job up. On SQLite older than 3.35 the same claim runs as two statements inside BEGIN IMMEDIATE.

Because it's just SQLite, you can read the queue with anything:

sqlite3 jobs.db "SELECT task, status, attempts, payload FROM jobs WHERE status = 'failed';"

sync_invoice|failed|3|{"args":[8842],"kwargs":{}}

That readability is most of the reason for preferring JSON over pickle.

Other backends

Backend choice is a URL:

app = TinyJobs("sqlite:///jobs.db")
app = TinyJobs("redis://localhost/0")

The core package doesn't import or know about optional backends. It looks the scheme up in the tinyjobs.backends entry point group, so a third-party package registers itself with:

[project.entry-points."tinyjobs.backends"]
redis = "tinyjobs_redis:RedisBackend"

Backends declare what they support (priority, delayed, dedup, ...) and asking for something unsupported raises when you enqueue rather than failing quietly later. No Redis or Postgres backend exists yet.

The protocol is async, so a backend built on a network client talks to it directly with no threads in the way. SQLite is the odd one out: sqlite3 is a blocking C library, so that backend does its own thread hop internally, on a dedicated single-worker executor. tests/test_async_backend.py implements a working in-memory backend in about 120 lines if you want a template.

Measured

Ten hours of steady load on an Apple silicon laptop: three worker processes, eight slots each, one SQLite file, a deliberate 25 jobs per second.

892,825 jobs        24.8/s sustained, zero enqueue failures
68 worker restarts  39 by SIGTERM, 29 by SIGKILL, zero unexpected deaths

Sampled once a minute for the whole run, comparing the first hour against the last:

	hour 0	hour 9
worker RSS	26.8 MB	26.6 MB
threads per worker	10	10
open file descriptors	47	47
database on disk	22 MB	23 MB
end-to-end p95	405 ms	406 ms

Nothing drifted. The p95 at hour nine is one millisecond off the p95 at hour zero, and the file descriptor count was the same 47 in all 595 samples.

A separate burst harness pushes the other end: 2,398 enqueues per second from 24 threads, 572 jobs per second end-to-end with six workers, and a crash recovery check that kills every worker while 48 jobs are mid-flight and confirms all 48 come back and finish. Some of them run twice, which is what at-least-once means; the check is that each one applied its side effect once, because the handlers are idempotent.

That run predates the built-in retention, so the harness did the pruning. Retention has since been measured on its own, same harness with the library doing the work: with it off the database climbed from 4.7 to 10.2 MB in three minutes and kept going, and with it on the same curve flattened at 6.9 MB and stayed there. Under a SIGKILL on a worker every thirty seconds it moved 0.6% across eight minutes.

Cron was measured on its own. Ten thousand schedules all due on the same minute fire in about a second from one scheduler, roughly 110 microseconds each. With four workers racing for them, 2,000 schedules a minute were fired 8,000 times in four minutes, each exactly once, p95 0.57 seconds late; 10,000 a minute, 40,000 times, each exactly once, p95 6.4 seconds late. Fifty-six schedulers racing on one schedule for twelve minutes produced twelve jobs, and fifty SIGKILLs across four workers and a standalone scheduler in ten minutes left no gap and no duplicate.

Two caveats worth stating. The tasks are synthetic, with sleep standing in for real I/O. And 25 jobs per second was chosen as a comfortable rate to hold for ten hours, not as a ceiling.

Status

v0.4.0. Working and measured, but young. What's in:

• task registry, enqueue, app.send
• SQLite backend, atomic claim, leases, reaper
• async backend protocol with a synchronous facade
• worker with async and threaded executors, timeouts, graceful shutdown on SIGTERM
• retries with backoff and jitter, Retry and Abandon
• priorities, named queues, delayed jobs, deduplication keys
• job.refresh(), job.wait() and job.result()
• retention, so the database plateaus instead of growing
• lifecycle hooks for metrics and alerting
• memory://, eager mode and app.drain() for testing
• cron schedules, fired leader-free from every worker, with timezones, misfire and overlap policies
• tinyjobs worker / scheduler / schedules / jobs / stats / check / send / purge
• the backend plugin seam

What's not in yet:

• transactional enqueue, where a job is only created if your own transaction commits
• a process executor for CPU-bound work
• any backend other than SQLite

docs/design.md explains the design and why it looks like this, including the parts not built yet.

Working on it

Nothing to install beyond an interpreter, but python3 on macOS is still 3.9, which is too old. A venv saves you from remembering:

uv venv --python 3.12 .venv     # or: python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .

Then python, tinyjobs and the example all point at the right interpreter:

python -m examples.demo
tinyjobs worker --app examples.demo:app --concurrency 4

Importing the package on anything older than 3.11 raises with the version it found and the path to it, rather than an unhelpful error about StrEnum.

Tests

python -m unittest discover -s tests -t .

284 tests, no test dependencies. The one that matters most is tests/test_claim_concurrency.py, which runs six processes and eight threads against one database and asserts every job was claimed exactly once.

License

MIT.