Redis job queue with redis-py

Implement a Redis job queue in Python with redis-py

This guide shows you how to implement a Redis-backed job queue in Python with redis-py. It includes a small local web server built with the Python standard library so you can enqueue jobs, watch a pool of workers drain them, and see the reclaimer recover jobs from a simulated worker crash.

Overview

A job queue lets your application offload background work — sending email, processing payments, image transcoding, ML inference, webhooks — from the request path. Producers enqueue jobs in milliseconds and return to the user; workers pull from the queue and process them on their own schedule.

That gives you:

  • Low-latency user-facing requests, even when downstream work is slow or bursty
  • Horizontal scale across many worker processes that share one Redis instance
  • At-least-once delivery so a worker crash doesn't lose work
  • Visibility-timeout reclaim that returns stuck jobs to the queue automatically
  • Job metadata, retry counts, and completion results in Redis hashes with TTL

In this example, each job is identified by a random hex ID and its payload, status, and result live in a Redis hash under queue:jobs:job:{id}. Pending IDs sit in a list, claimed IDs move atomically to a processing list, and completed or failed IDs land in short history lists.

How it works

The flow looks like this:

  1. The application calls queue.enqueue(payload)
  2. The helper writes the job metadata hash and LPUSHes the job ID onto the pending list
  3. A worker calls queue.claim(timeout_ms)
  4. The helper runs BRPOPLPUSH to atomically move the next pending ID into the processing list and writes a per-claim claim_token plus claimed_at_ms on the hash
  5. The worker runs the job and calls queue.complete(job, result) or queue.fail(job, error)
  6. complete removes the job from the processing list, writes the result, and LPUSHes the ID onto the completed history (with LTRIM and an EXPIRE on the hash for cleanup)
  7. fail either retries the job (back to pending) or moves it to the failed list once retries are exhausted

If a worker dies before completing a job, the job sits in the processing list with a claimed_at_ms older than the visibility timeout. A periodic call to queue.reclaim_stuck() finds those jobs and moves them back to pending so another worker can pick them up.

Every state change holds the token: a worker that has been reclaimed cannot later complete or fail a job another worker has already claimed.

The job queue helper

The RedisJobQueue class wraps the queue operations (source):

import redis
from job_queue import RedisJobQueue

r = redis.Redis(host="localhost", port=6379, decode_responses=True)
queue = RedisJobQueue(redis_client=r, visibility_ms=5000)

job_id = queue.enqueue({"kind": "email", "recipient": "[email protected]"})

# In a worker process:
job = queue.claim(timeout_ms=1000)
if job is not None:
    try:
        # ... run the job ...
        queue.complete(job, result={"sent_at": "2026-05-11T15:00:00Z"})
    except Exception as exc:
        queue.fail(job, error=str(exc))

# In a periodic sweeper:
reclaimed = queue.reclaim_stuck()

Data model

Each job's state lives in a Redis hash plus a position in one of four lists:

queue:jobs:pending          (list)   pending job IDs, oldest at the right
queue:jobs:processing       (list)   claimed but not yet completed
queue:jobs:completed        (list)   recent successes (LTRIM-capped history)
queue:jobs:failed           (list)   terminally failed jobs
queue:jobs:job:{id}         (hash)   per-job metadata
queue:jobs:events           (pubsub) completion notifications

A job's hash carries:

queue:jobs:job:9a4f...
  id              = 9a4f...
  payload         = {"kind":"email","recipient":"[email protected]"}
  status          = pending | processing | completed | failed
  attempts        = 1
  enqueued_at_ms  = 1715441000000
  claimed_at_ms   = 1715441000123
  claim_token     = b3c0d1e2...        (per-claim random token)
  completed_at_ms = 1715441000456
  result          = {"sent_at":"..."}
  last_error      = "smtp timeout"

The implementation uses:

  • LPUSH to add new job IDs to the pending list
  • BRPOPLPUSH to atomically claim a job into the processing list
  • LREM to remove a claimed job from the processing list on complete or fail
  • LTRIM to cap the completed and failed history lists
  • HSET / HGETALL for job metadata
  • EXPIRE on completed and failed hashes for automatic cleanup
  • PUBLISH on queue:jobs:events for completion signalling
  • Lua scripting (EVALSHA) for the complete, fail, and reclaim flows so each runs atomically against the processing list and metadata hash

Enqueueing jobs

enqueue() writes the metadata hash and pushes the job ID onto the pending list in one pipeline:

def enqueue(self, payload: dict) -> str:
    job_id = secrets.token_hex(8)
    now_ms = self._now_ms()
    meta = {
        "id": job_id,
        "payload": json.dumps(payload),
        "status": "pending",
        "attempts": 0,
        "enqueued_at_ms": now_ms,
        "claim_token": "",
    }
    pipe = self.redis.pipeline()
    pipe.hset(self._meta_key(job_id), mapping=meta)
    pipe.lpush(self.pending_key, job_id)
    pipe.execute()
    return job_id

The payload is stored as JSON so the queue can carry arbitrary nested structures without forcing every field into a hash.

Claiming jobs with BRPOPLPUSH

A worker blocks until a job is available, then atomically pops it from the pending list and pushes it onto the processing list. BRPOPLPUSH does both in a single Redis call:

def claim(self, timeout_ms: int = 1000) -> Optional[ClaimedJob]:
    timeout_s = max(timeout_ms / 1000.0, 0.1)
    job_id = self.redis.brpoplpush(self.pending_key, self.processing_key, timeout=timeout_s)
    if job_id is None:
        return None

    token = secrets.token_hex(8)
    now_ms = self._now_ms()
    meta_key = self._meta_key(job_id)
    pipe = self.redis.pipeline()
    pipe.hset(meta_key, mapping={
        "status": "processing",
        "claimed_at_ms": now_ms,
        "claim_token": token,
    })
    pipe.hincrby(meta_key, "attempts", 1)
    pipe.hgetall(meta_key)
    _, _, meta = pipe.execute()
    return ClaimedJob(job_id, json.loads(meta["payload"]), int(meta["attempts"]), token)

The claim_token is the worker's proof of ownership for this attempt. Every subsequent state change (complete, fail) checks it before touching the processing list, so a worker that hung past the visibility timeout cannot interfere with the new claimant.

Completing jobs

complete() runs a Lua script so the processing-list removal, the metadata write, and the history push happen atomically:

def complete(self, job: ClaimedJob, result: dict) -> bool:
    ok = self._complete(
        keys=[self.meta_prefix, self.processing_key, self.completed_key],
        args=[
            job.id,
            job.claim_token,
            "completed",
            self._now_ms(),
            json.dumps(result),
            self.completed_ttl,
            self.completed_history,
        ],
    )
    if not ok:
        return False
    self.redis.publish(self.events_channel, json.dumps({"id": job.id, "status": "completed"}))
    return True

The Lua script checks the token first and returns 0 if the worker no longer owns the job (because the reclaimer moved it back to pending). The metadata hash also gets an EXPIRE so completed jobs are cleaned up automatically.

Failing and retrying

fail() either retries the job (back to pending) or moves it to the failed list once retries are exhausted:

def fail(self, job: ClaimedJob, error: str) -> bool:
    retry = job.attempts < self.max_attempts
    result = self._fail(
        keys=[self.meta_prefix, self.processing_key, self.pending_key, self.failed_key],
        args=[
            job.id,
            job.claim_token,
            error,
            self._now_ms(),
            self.completed_ttl,
            self.completed_history,
            "1" if retry else "0",
        ],
    )
    return bool(result)

The attempt counter is incremented on every claim(), so a job that fails three times is moved to the failed list with attempts = 3 and the final last_error preserved.

Reclaiming stuck jobs

If a worker dies mid-job — the process is killed, the host loses power, the network partitions — the job sits in the processing list with a claimed_at_ms that never advances. A periodic call to reclaim_stuck() walks the processing list and moves any job past the visibility timeout back to pending:

def reclaim_stuck(self) -> list[str]:
    reclaimed = self._reclaim(
        keys=[self.pending_key, self.processing_key, self.meta_prefix],
        args=[self._now_ms(), self.visibility_ms],
    )
    return list(reclaimed)

The Lua script also handles a narrower race: a worker that crashed between BRPOPLPUSH and writing claimed_at_ms. Those jobs are reclaimed after 2 × visibility_ms using enqueued_at_ms as a fallback timer, so they aren't stranded forever.

Stats and history

stats() reports queue depth plus per-process counters:

def stats(self) -> dict:
    pipe = self.redis.pipeline()
    pipe.llen(self.pending_key)
    pipe.llen(self.processing_key)
    pipe.llen(self.completed_key)
    pipe.llen(self.failed_key)
    pending, processing, completed, failed = pipe.execute()
    return {
        "enqueued_total": self._enqueued,
        "completed_total": self._completed,
        "failed_total": self._failed,
        "reclaimed_total": self._reclaimed,
        "pending_depth": pending,
        "processing_depth": processing,
        "completed_depth": completed,
        "failed_depth": failed,
        "visibility_ms": self.visibility_ms,
    }

The completed and failed lists are capped via LTRIM so they never grow unbounded; a real deployment would also write completion events to a longer-term audit log if needed.

Prerequisites

  • Redis 6.2 or later running locally on the default port (6379). Earlier versions still work, since the helper uses commands that have existed since Redis 2.6.

  • Python 3.9 or later.

  • The redis-py client. Install it with:

    pip install "redis>=5.0"

Running the demo

Get the source files

The demo consists of three Python files. Download them from the redis-py source folder on GitHub, or grab them with curl:

mkdir job-queue-demo && cd job-queue-demo
BASE=https://raw.githubusercontent.com/redis/docs/main/content/develop/use-cases/job-queue/redis-py
curl -O $BASE/job_queue.py
curl -O $BASE/worker.py
curl -O $BASE/demo_server.py

Start the demo server

From that directory:

python3 demo_server.py

You should see:

Redis job-queue demo server listening on http://127.0.0.1:8090
Using Redis at localhost:6379
Visibility timeout: 5000 ms

Open http://127.0.0.1:8090 in a browser. You can:

  • Enqueue jobs of different kinds (email, webhook, thumbnail, invoice) in batches.
  • Start a pool of workers with configurable size, work latency, and failure / hang rates. A non-zero hang rate simulates worker crashes.
  • Click Run reclaim sweep to move any timed-out processing jobs back to pending.
  • Watch pending / processing / completed / failed lists update every 800 ms.

If your Redis server is running elsewhere, start the demo with --redis-host and --redis-port. You can also tune the visibility timeout with --visibility-ms.

The mock worker pool

The demo includes a small Worker and WorkerPool (source) that stands in for whatever real background work your application would run. Each worker:

  • Blocks on queue.claim() for new jobs.
  • Sleeps work_latency_ms to simulate doing the work.
  • Either completes successfully, fails (calling queue.fail()), or hangs — returning without completing or failing the job so the reclaimer has to recover it.

The fail_rate and hang_rate knobs let you watch the at-least-once delivery and reclaim behaviours from the UI without writing test code.

Production usage

Choose a visibility timeout that matches your worst-case job latency

The visibility timeout has to exceed the longest real job time, with margin. If it's too short, a healthy worker that's running a slow job will get its work duplicated when the reclaimer fires. If it's too long, a real crash takes longer to detect. Most production deployments use a per-queue value tuned to the 99th-percentile job latency — for example, 2 minutes for email and 30 minutes for video transcoding.

Run the reclaimer on a schedule

The demo only reclaims when you click the button. In production, run reclaim_stuck() from a periodic task (every few seconds for fast queues, every minute for slow ones), or from each worker before it blocks on claim(). Both patterns work as long as someone runs the sweep.

Use a separate Redis database or key prefix per queue

The helper takes a queue_name argument so you can run multiple independent queues against one Redis instance — for example, one queue per priority level, or one per job kind. Keep queue keys under a clearly-namespaced prefix (here, queue:jobs:*) so they're easy to inspect and easy to clear without touching application data.

Cap the completed and failed history

The demo keeps the last 50 completed and 50 failed job IDs via LTRIM. If you need longer history for audit purposes, write completion events to a separate Redis Stream (or to an external store) and keep the in-queue history short. Stream consumer groups give you the same fan-out semantics with a much richer history.

Tune max_attempts per job kind

A blanket max_attempts = 3 is a reasonable default for transient failures (network timeouts, rate limits). Jobs that talk to non-idempotent external systems — for example, posting a Stripe charge — need either application-level idempotency keys or a much lower retry count. The helper exposes max_attempts so each queue can pick its own policy.

Inspect queue state directly in Redis

Because the queue is just lists and hashes, you can inspect it with redis-cli:

# How many pending jobs?
redis-cli LLEN queue:jobs:pending

# Look at the next 5 jobs to be picked up.
redis-cli LRANGE queue:jobs:pending -5 -1

# Read a job's metadata.
redis-cli HGETALL queue:jobs:job:9a4f0d1c

# How many jobs are currently being processed?
redis-cli LLEN queue:jobs:processing

# Clear everything for this queue (be careful — this deletes work).
redis-cli --scan --pattern 'queue:jobs:*' | xargs redis-cli DEL

Learn more

This example uses the following Redis commands:

  • LPUSH to enqueue a job ID.
  • BRPOPLPUSH to atomically claim a job into the processing list.
  • LREM to remove a job from the processing list on complete or fail.
  • LRANGE and LLEN to read queue depth and list contents.
  • LTRIM to cap the completed and failed history.
  • HSET and HGETALL for job metadata.
  • HINCRBY for the attempt counter.
  • EXPIRE for automatic cleanup of completed and failed jobs.
  • PUBLISH for job-completion notifications.
  • EVALSHA for atomic complete, fail, and reclaim flows.

See the redis-py documentation for full client reference.

RATE THIS PAGE
Back to top ↑