redis-message-queue

Lightweight Python message queuing with Redis and built-in publish-side deduplication. Deduplicate publishes within a TTL window, with optional crash recovery — across any number of producers and consumers.

pip install "redis-message-queue>=8.0.0,<9.0.0"

Requires Redis server >= 6.2.

Quickstart

Redis must be running locally first: use redis-server or docker run -p 6379:6379 redis:7.

from redis import Redis
from redis_message_queue import RedisMessageQueue

client = Redis.from_url("redis://localhost:6379/0", decode_responses=True)
queue = RedisMessageQueue(
    "quickstart",
    client=client,
    deduplication=True,
    get_deduplication_key=lambda msg: msg["id"],
)
queue.publish({"id": "msg-1", "text": "hello"})
with queue.process_message() as message:
    if message is not None:
        print(f"got {message['text']}")
# Expected output: got hello

RedisMessageQueue itself is not a context manager. Use with queue.process_message() as message: for each message.

Important: In the sync API, work inside process_message() must be synchronous. If your handler is async def, returns a coroutine, or returns any other awaitable, use redis_message_queue.asyncio.RedisMessageQueue. The sync context manager does not inspect the handler's return value; an unawaited coroutine can be dropped while the message is acked. An ergonomic callback API that detects this is planned for v8.1.

Async quickstart

import asyncio
from redis.asyncio import Redis
from redis_message_queue.asyncio import RedisMessageQueue

async def main():
    client = Redis.from_url("redis://localhost:6379/0", decode_responses=True)
    queue = RedisMessageQueue(
        "quickstart",
        client=client,
        deduplication=True,
        get_deduplication_key=lambda msg: msg["id"],
    )
    await queue.publish({"id": "msg-1", "text": "hello"})
    async with queue.process_message() as message:
        print(f"got {message['text']}")
    await client.aclose()

asyncio.run(main())  # Expected output: got hello

Why redis-message-queue

The problem: You're sending messages between services or workers and need guarantees. Simple Redis LPUSH/BRPOP loses messages on crashes, doesn't deduplicate, and gives you no visibility into what succeeded or failed.

The solution: Atomic Lua scripts for publish + dedup, a processing queue for in-flight tracking (with optional crash recovery via visibility timeouts), and optional success/failure logs for observability.

Feature	Details
Deduplicated publish	Lua-scripted atomic SET NX + LPUSH prevents duplicate enqueues within a configurable TTL window (default: 1 hour), even with producer retries. Requires an explicit get_deduplication_key callable so your application defines what counts as a duplicate. Note: deduplication is publish-side only and does not prevent duplicate delivery under at-least-once visibility-timeout reclaim
Visibility-timeout redelivery	Crashed or stalled consumers' messages are reclaimed and redelivered when a visibility timeout is configured
Success & failure logs	Optional completed/failed queues for auditing and reprocessing, with configurable max length to prevent unbounded growth
Dead-letter queue	Poison messages that exceed a configurable delivery count are automatically routed to a dead-letter queue instead of being redelivered indefinitely
Graceful shutdown	Built-in interrupt handler lets consumers finish current work before stopping
Lease heartbeats	Optional background lease renewal keeps long-running handlers from being redelivered prematurely
Connection retries	Exponential backoff with jitter for Redis operations (deduplicated publish, ack, lease renewal). Publish and cleanup paths use replay markers so retryable connection drops preserve the original result within the same call. Message-claim paths use idempotent Lua claim IDs plus persisted claim metadata so retryable errors can recover the original claim safely, either in the same wait call or on the next call from the same gateway instance if the original wait had to give up before Redis became reachable again. Active waits keep their in-flight claim IDs private until they exit, so a concurrent caller on the same gateway instance cannot recover the same claim twice. Timed waits also stay bounded: once the configured wait window expires, the queue only replays persisted state for that same claim attempt and will not claim fresh work after the deadline. If a graceful interrupt arrives during claim recovery, the wait call stops instead of taking fresh work. Non-deduplicated publish is not retried — the exception propagates so the caller can decide whether to retry (accepting potential duplicates)
Async support	Drop-in async variant with identical API

All features are optional and can be enabled or disabled as needed.

Delivery semantics

Configuration	Delivery guarantee
Default (visibility_timeout_seconds=300)	At-least-once — expired messages are reclaimed and redelivered
With visibility_timeout_seconds=None	At-most-once — a consumer crash loses the in-flight message

See Crash recovery with visibility timeout for details and tradeoffs.

Important: Handler exceptions are terminal. This library is a payload queue, not a task framework: raising inside process_message() does not requeue the message. With enable_failed_queue=False, the message is removed from processing; with enable_failed_queue=True, it is moved to the failed list.

Configuration

Deduplication

# Deduplicate by order ID for a 1-hour TTL
queue = RedisMessageQueue(
    "q", client=client,
    deduplication=True,
    get_deduplication_key=lambda msg: msg["order_id"],
)

# Disable deduplication entirely
queue = RedisMessageQueue("q", client=client, deduplication=False)

Dedup key callable must return a non-empty, high-cardinality, tenant-scoped string

When deduplication=True, get_deduplication_key is required. The callable is called once per publish and must return a str that uniquely represents the deduplication scope for that message. Returning None or "" raises ConfigurationError at publish time; returning a non-str value raises TypeError.

Use stable, high-cardinality keys that include any tenant or account boundary needed by your system:

queue = RedisMessageQueue(
    "orders",
    client=client,
    deduplication=True,
    get_deduplication_key=lambda msg: f"{msg['tenant_id']}:{msg['order_id']}",
)

Avoid fallback patterns such as lambda msg: msg.get("order_id", ""). Missing fields should fail loudly instead of collapsing unrelated messages into one deduplication key.

Deduplication markers and publish retry-safety markers are Redis TTL keys. A large forward step in the Redis server expiration clock during an in-call retry window can expire those markers before the Python-side monotonic retry budget elapses, allowing a duplicate publish. This is an extreme anomaly, mainly relevant under cluster-wide NTP step corrections while a producer is retrying after an ambiguous Redis write.

Success and failure tracking

queue = RedisMessageQueue(
    "q", client=client,
    enable_completed_queue=True,   # track successful messages
    enable_failed_queue=True,      # track failed messages for reprocessing
)

Completed and failed tracking queues are capped at 1,000 entries by default when enabled. Override the caps when you need a different retention window:

queue = RedisMessageQueue(
    "q", client=client,
    enable_completed_queue=True,
    enable_failed_queue=True,
    max_completed_length=10000,    # keep only the most recent 10,000
    max_failed_length=1000,        # keep only the most recent 1,000
)

When set, LTRIM is called after each message is moved to the completed/failed queue. This is best-effort cleanup — if the trim fails, the queue is slightly longer until the next successful trim. Pass max_completed_length=None or max_failed_length=None explicitly if you want unbounded tracking queues.

Publish backpressure

By default, the pending queue is unbounded (max_pending_length=None), matching the v5 behavior. Set max_pending_length when producers can outrun consumers and Redis memory must fail closed before the broker is exhausted:

queue = RedisMessageQueue(
    "q",
    client=client,
    max_pending_length=100_000,
    pending_overload_policy="raise",  # "raise", "drop_oldest", or "block"
)

The built-in Redis path checks pending depth and enqueues in the same Lua script, so concurrent publishers cannot race above the configured cap. Overload policies:

• raise raises QueueBackpressureError and leaves the pending list unchanged.
• drop_oldest removes the oldest pending message (RPOP) before enqueueing the new message. This is silent data loss by design; deduplication markers for dropped messages are not removed, so a dropped duplicate may still be suppressed until its dedup TTL expires.
• block retries the atomic check until space opens or pending_overload_block_timeout_seconds elapses (default: 1.0), then raises QueueBackpressureError.

These limits apply only to the pending list at publish time. They do not cap messages already in processing, dead-letter queues, deduplication keys, or replay metadata. max_completed_length and max_failed_length only bound the completed/failed history lists. Size pending payload memory separately from the dedup/replay metadata described in Redis memory sizing.

When using gateway=, configure backpressure on the gateway directly, for example RedisGateway(redis_client=client, max_pending_length=100_000).

Crash recovery with visibility timeout

queue = RedisMessageQueue(
    "q",
    client=client,
    visibility_timeout_seconds=300,
    heartbeat_interval_seconds=60,
)

Important: visibility_timeout_seconds is a lease, not a handler runtime cap. rmq never interrupts a long-running handler. If the lease expires while the handler continues, another consumer can reclaim and process the same message concurrently.

This enables lease-based redelivery for messages left in processing by a crashed worker and renews the lease while a healthy long-running handler is still working. Tradeoffs:

• delivery becomes at-least-once after lease expiry
• the timeout must be longer than your normal processing time if you do not use heartbeats
• if you do use heartbeats, the heartbeat interval must be less than half of the visibility timeout
• recovery happens on consumer polling cadence rather than instantly
• heartbeats add background renewal work for active messages
• if a heartbeat fails (network error or stale lease), the heartbeat stops silently; the consumer continues processing but may find at ack time that the message was reclaimed by another consumer

Pass on_heartbeat_failure to receive a best-effort callback when the heartbeat stops because renewal failed:

queue = RedisMessageQueue(
    "q", client=client,
    visibility_timeout_seconds=300,
    heartbeat_interval_seconds=60,
    on_heartbeat_failure=lambda: log.warning("heartbeat failed; lease may be stale"),
)

The callback is advisory — it may fire briefly after a successful process_message exit when a final renewal coincided with the success path. Use it for metrics or alerting, not as a correctness signal. For the async queue (redis_message_queue.asyncio), the callback may also be async def.

Without a visibility timeout, messages already moved to processing remain there indefinitely after a consumer crash and are not redelivered, even if the crash happened before your handler started running.

Visibility deadlines use Redis server time (TIME), not Python process time. A forward step in the Redis server clock can make a live lease appear expired and allow premature redelivery while the original consumer is still processing; a backward step can delay reclaim of truly abandoned messages. Treat NTP step corrections on Redis hosts as a deployment risk. Prefer time-synchronization discipline that slews corrections rather than stepping the Redis clock.

Ordering and multi-consumer fairness

The built-in queue is a shared-pull Redis list. Successful publishes push to the left side of the pending list, and claims pop from the right side, so Redis grants claims in enqueue order in the no-failure path.

This is a claim-order guarantee only. It is not a completion-order guarantee: multiple consumers process concurrently, handlers can run for different durations, and younger messages can finish before older messages.

With visibility_timeout_seconds enabled, expired messages from processing are reclaimed before fresh pending work on the next consumer poll. A reclaimed message may be delivered after younger messages were already processed, and may be processed concurrently with a stale original handler if that handler keeps running after its lease expires.

Expired reclaims are ordered by lease deadline within one reclaim batch. CLAIM_MESSAGE_WITH_VISIBILITY_TIMEOUT_LUA_SCRIPT selects expired leases with ZRANGEBYSCORE ... LIMIT 0, 100 to bound Redis Lua execution time. When more than 100 messages expire together, the next poll can append a later reclaim batch at the claimable end of the pending list ahead of leftovers from the previous batch, so cross-batch redelivery order is not guaranteed.

max_delivery_count can skip over poison messages during a claim poll by moving over-limit messages to the dead-letter queue and returning a later pending message. Deduplication is publish-side only: duplicate publishes are not enqueued and therefore do not occupy a queue position.

Handler exceptions are not retries: the default behavior removes the message from processing, or moves it to the failed queue when enabled. Redelivery is for crash, stall, or stale-lease paths where cleanup does not complete.

Multiple consumers contend for the same queue. The next message goes to the consumer whose claim request Redis executes next. There is no round-robin, equal-share, or starvation-freedom guarantee; faster consumers can receive more than 1/N of messages.

If you need stronger ordering or fairness guarantees

• Strict queue-wide processing order — use a single consumer per queue. Multiple consumers will interleave handler completions.
• Per-key processing order — partition by key into multiple queues (queue_<hash(key) % N>), and consume each partition with a single consumer.
• Equal-share / round-robin fairness across consumers — choose a different scheduler. This queue does not guarantee that any individual consumer makes forward progress at any specific rate.
• Cross-batch ordering after reclaim — accept that reclaimed messages will reappear after newer un-reclaimed messages have been consumed. If your handler must observe original publish order, persist that order in the payload (for example, a sequence number set by the producer). For clock-related operator detail behind reclaim behavior, see production readiness R11.

Dead-letter queue

queue = RedisMessageQueue(
    "q",
    client=client,
    visibility_timeout_seconds=300,
    max_delivery_count=5,
)

When a message has been delivered more than max_delivery_count times (due to consumer crashes causing visibility-timeout reclaim), it is automatically routed to a dead-letter queue ({name}::dlq) instead of being redelivered. max_delivery_count defaults to 10 on the built-in client= path, with the DLQ name auto-derived from the queue name. This prevents poison messages from cycling indefinitely.

Notes:

• requires visibility_timeout_seconds to be set (poison messages are only a concern with VT reclaim)
• the delivery count is tracked per-message in a Redis HASH and cleaned up on successful ack or move to completed/failed
• the delivery count increments when Redis grants the claim/lease, not when your handler begins running. If a process exits after Redis claims a message, that claim still counts toward max_delivery_count
• max_delivery_count=1 means the message is delivered once; any reclaim routes it to the dead-letter queue
• set max_delivery_count=None explicitly for unlimited redelivery
• dead-lettered messages contain the raw payload only — the internal envelope (which carries a per-delivery UUID) is stripped before pushing to the DLQ, consistent with how completed/failed queues store messages. Two identical payloads dead-lettered separately are indistinguishable in the DLQ

Graceful shutdown

from redis_message_queue import RedisMessageQueue, GracefulInterruptHandler

interrupt = GracefulInterruptHandler()
queue = RedisMessageQueue("q", client=client, interrupt=interrupt)

while not interrupt.is_interrupted():
    with queue.process_message() as message:
        if message is not None:
            process(message)
# Consumer finishes current message before exiting on Ctrl+C

Note: GracefulInterruptHandler claims process-global signal handlers for its signals (default: SIGINT, SIGTERM, SIGHUP), but only when those signals are still using Python's default disposition. If another handler is already installed, or if another GracefulInterruptHandler already owns the signal, construction raises ValueError. A repeated owned signal falls back to the default behavior (for example, a second Ctrl+C raises KeyboardInterrupt). If you need multiple shutdown hooks, use a single handler and fan out in your own code.

Process-global signal ownership cannot be safely chained with task-worker CLIs such as Celery, RQ, or Dramatiq. Run sibling workers in separate processes, or install one top-level signal owner that calls queue.drain() / queue.aclose() or sets an application stop event.

If another library owns SIGTERM/SIGINT in the same process, adapt its shutdown signal to rmq with a user-owned event instead of installing rmq signal handlers:

import threading

from redis_message_queue import EventDrivenInterruptHandler, RedisMessageQueue

stop_event = threading.Event()
interrupt = EventDrivenInterruptHandler(stop_event)
queue = RedisMessageQueue("q", client=client, interrupt=interrupt)

while not interrupt.is_interrupted():
    with queue.process_message() as message:
        if message is not None:
            process(message)

# In the sibling library's shutdown hook:
stop_event.set()
queue.drain(timeout=25)

The caller MUST set stop_event before exiting. rmq observes is_interrupted() and exits cooperatively; it does not call sys.exit() or otherwise force process shutdown.

There are three distinct shutdown shapes; pick the one that matches your runtime:

Shape	Trigger	In-flight handler	Pending claim IDs
Flag-based soft drain (GracefulInterruptHandler)	First SIGINT/SIGTERM flips a flag	Runs to completion	Drained on the next claim call, not on signal arrival
Async task cancellation (asyncio.CancelledError)	Framework cancels the worker task (Uvicorn/K8s SIGTERM in many setups)	Hard abort — message stays in processing; with VT it is reclaimed at deadline expiry, without VT it is orphaned	Not drained
Explicit drain (drain() / aclose())	You call the method	Caller's responsibility to let it finish (drain does not cancel)	Drained synchronously via the gateway recovery path; new publishes are refused

Use drain() / aclose() to bridge K8s preStop / SIGTERM grace windows without relying on signal interception:

# sync — in your SIGTERM handler or preStop hook
queue.drain(timeout=25)   # refuses new publishes/claims, recovers pending claim IDs
worker_thread.join()      # wait for in-flight process_message to finish

# async — same shape
await queue.aclose(timeout=25)
await worker_task         # task observes ``_draining`` and exits its loop

drain() / aclose() set a queue-local flag so subsequent process_message() calls yield None immediately and subsequent publish() calls raise QueueDrainedError("queue is drained"). Drain also gates the publish path: if a publish is already inside the queue instance's publish path, drain waits for that publish to finish before it returns; publishes that arrive after the drained flag is set are rejected. The drained state is local to that Python queue object and is not written to Redis, so constructing a fresh RedisMessageQueue(...) over the same keys remains usable. A separate process or separate queue instance against the same Redis keys is not marked drained by this call. For multi-process graceful shutdown, each process must drain its own queue instances.

Drain does not cancel in-flight handlers — the caller must arrange handler exit through normal thread/task coordination. Returns True if all in-memory pending claim IDs were recovered within the timeout; False if the deadline fired or transient Redis errors left claim IDs pending (call again to retry). timeout=0 reports current state without attempting recovery.

Abandoned in-flight messages

Abandoned in-flight messages are recovered lazily. Async tasks cancelled without aclose(), or sync processes killed mid-handler, can leave the message and its processing/lease metadata in Redis until a later consumer claim path triggers visibility-timeout reclaim. With visibility timeouts enabled, this is the designed at-least-once recovery path: the message is delayed by the lease, not lost. With visibility_timeout_seconds=None, there is no automatic reclaim path. For low-visibility-timeout workloads, prefer an explicit drain() / aclose() during shutdown so local pending claim IDs are recovered before process exit.

drain() / aclose() timeouts are measured with Python monotonic clocks, but any lease deadlines they recover were created from Redis server time. The same Redis-clock step caveats from Crash recovery with visibility timeout apply to when abandoned work becomes reclaimable.

Heartbeat caveat (best-effort stop): when heartbeat_interval_seconds is set, the heartbeat sidecar's stop() is bounded but not strictly quiescent — a slow renewal in flight when process_message exits may still write to Redis after the caller believes shutdown is complete. The renewal is bounded by the configured visibility timeout and the lease token check on the Redis side, but plan for a small post-shutdown overlap rather than instant quiesce.

Custom gateway

from redis_message_queue import RedisGateway

# Tune retry budget, dedup TTL, or wait interval
gateway = RedisGateway(
    redis_client=client,
    retry_budget_seconds=120,          # total retry window (set 0 to disable retry)
    retry_max_delay_seconds=5.0,       # cap on per-attempt backoff
    retry_initial_delay_seconds=0.01,  # first backoff
    message_deduplication_log_ttl_seconds=3600,
    message_wait_interval_seconds=10,
    message_visibility_timeout_seconds=300,
)
queue = RedisMessageQueue("q", gateway=gateway)

The retry knobs configure an internal tenacity strategy: exponential backoff with jitter, retry on transient Redis errors only, capped at retry_budget_seconds. The budget is monotonic elapsed time from the first attempt (including attempt duration), not inter-attempt delay; it is unaffected by Python-host NTP jumps. A single attempt that takes longer than the budget results in zero retries. Setting retry_budget_seconds=0 disables retry entirely (single attempt; exceptions propagate). The library uses retry_budget_seconds to size the operation-result cache TTL automatically, so the previous footgun of an over-long retry budget out-living the cache and producing misleading "cleanup was a no-op" warnings is now structurally impossible. Note: tenacity may allow one additional attempt beyond the budget if the budget check passes at attempt start, so total monotonic elapsed time can exceed retry_budget_seconds by the duration of that final attempt.

To plug in a different retry library (backoff, asyncstdlib.retry, or your own logic) or fundamentally different semantics, subclass AbstractRedisGateway from redis_message_queue (or redis_message_queue.asyncio for the async sibling) and override the operation methods directly.

If your custom gateway uses visibility timeouts, it must expose a public message_visibility_timeout_seconds value and return ClaimedMessage from wait_for_message_and_move(). The queue now fails closed if a lease-capable gateway returns plain str/bytes, because cleanup without a lease token can ack a message that has already been reclaimed by another consumer. If a lease-capable custom gateway omits message_visibility_timeout_seconds, the queue cannot detect that lease semantics are in play and will treat the gateway as a non-lease gateway. In that misconfigured state, lease-token safety checks and heartbeat validation are bypassed.

When using a custom gateway with dead-letter queue support, configure max_delivery_count and dead_letter_queue directly on the gateway — do not pass max_delivery_count to RedisMessageQueue:

gateway = RedisGateway(
    redis_client=client,
    message_visibility_timeout_seconds=300,
    max_delivery_count=3,
    dead_letter_queue="myqueue::dead_letter",
)
queue = RedisMessageQueue("myqueue", gateway=gateway)

Use a separate gateway instance per queue when max_delivery_count is enabled. Dead-letter routing is gateway-scoped, so reusing the same gateway across different queues is rejected.

If you use Redis Sentinel, pass the Redis client returned by sentinel.master_for(name) to client= or RedisGateway(redis_client=...), not the sentinel object itself.

Connection pool sizing

Each queue with heartbeat_interval_seconds set uses up to 2 simultaneous connections: one for the main operation and one for heartbeat renewal. Size Redis client pools with max_connections >= 2 * number_of_queues + headroom.

Async API

Replace the import to use the async variant — the API is identical:

from redis_message_queue.asyncio import RedisMessageQueue

The sync and async classes intentionally share names. In modules that use both, alias the imports explicitly, for example from redis_message_queue import RedisMessageQueue as SyncRedisMessageQueue and from redis_message_queue.asyncio import RedisMessageQueue as AsyncRedisMessageQueue.

All examples work the same way. Remember to close the connection when done:

import redis.asyncio as redis

client = redis.Redis()
# ... your code
await client.aclose()

For the sync Redis client, call client.close() during application shutdown when you own the client lifecycle.

Migrating from RQ / Celery / Dramatiq / taskiq

redis-message-queue is a payload queue, not a task framework. It has no task registry, job object, result backend, scheduler, workflow canvas, callback graph, or handler-level retry policy. Producers publish a str or dict payload, and consumers decide what that payload means.

The most important semantic differences from sibling task libraries are:

• Handler exceptions are terminal. Raising inside process_message() removes the message from processing, or moves it to the failed list when enable_failed_queue=True; it does not requeue or retry the message.
• visibility_timeout_seconds is a crash/stall recovery lease, not a runtime limit. Slow handlers are not interrupted; after the lease expires another consumer can process the same payload concurrently.
• on_event is telemetry only. Callback exceptions are logged and emitted as RuntimeWarning, but they do not affect ack/nack, failed-queue movement, or any other message outcome. Do not use on_event for sagas, follow-up writes, billing callbacks, or other correctness-critical work.
• Dict payloads are JSON data, not Python call arguments. JSON does not preserve every Python type: tuples become lists, and sets or custom objects raise unless you encode them into JSON-native values first.
• Process-global signal ownership cannot be safely chained with Celery, RQ, or Dramatiq CLI workers. Prefer one top-level owner that calls queue.drain() or sets an application stop event, and run sibling workers in separate processes.

When migrating on the same Redis deployment, prefer separate Redis DBs or hard namespaces. Do not point a Celery, RQ, Dramatiq, or taskiq worker at an rmq pending key. A sibling worker can pop the rmq stored message, fail its own decoder, and leave the rmq queue without that message. Also avoid custom key_separator values that synthesize another library's key namespace, such as using ":queue:" with a queue name that overlaps RQ keys. rmq has no fixed library prefix; generated keys share the Redis DB namespace with every other Redis user.

Set strict_envelope_decoding=True if this Redis is shared with sibling task libraries (Celery, RQ, Dramatiq) to fail-fast on foreign payloads. With the default False, non-rmq values that do not start with the rmq envelope prefix remain backward-compatible raw messages and are yielded to the handler.

Production notes

Fork safety and pre-fork servers

Construct Redis clients and RedisMessageQueue instances after a process forks. This is the recommended pattern for multiprocessing, ProcessPoolExecutor, and pre-fork servers such as gunicorn with --preload.

def worker_main():
    client = redis.Redis()
    queue = RedisMessageQueue("jobs", client=client)
    ...

Avoid constructing a queue/client in a parent process and then using that same object in forked children, especially if the parent has already run any Redis command. The queue stores the user-provided Redis client and process-local claim-recovery state. Inherited Redis sockets can corrupt the Redis protocol if two processes use the same file descriptor.

Notes:

• The sync redis-py pooled client attempts to reset its connection pool after fork, but this does not apply to every client shape.
• The built-in sync gateway rejects redis.Redis(single_connection_client=True) because that mode pins one socket instead of using the pool.
• Do not share redis.asyncio.Redis or async queues across fork; create or reconnect them in the child process.
• If you use GracefulInterruptHandler, create it in the worker process after fork so signal ownership is local to that worker.
• The heartbeat sidecar is lazy and starts only while processing a leased message. Do not call fork() from inside active message handlers unless the child exits without using the inherited queue/client.

Forking after constructing GracefulInterruptHandler

If your application constructed GracefulInterruptHandler in the parent process before os.fork() (for example, via module import in a pre-fork app server), forked children cannot construct a fresh handler for the same signal because the inherited signal table still routes to the parent-process handler.

In each child process, call parent_handler.reset() before constructing a fresh handler:

def worker_main():
    # Inherited handler from parent - reset it.
    if shared.interrupt_handler is not None:
        shared.interrupt_handler.reset()

    # Now safe to construct a fresh handler for this child.
    interrupt = GracefulInterruptHandler()
    queue = RedisMessageQueue("jobs", client=redis.Redis(), interrupt=interrupt)
    ...

Alternatively, defer all construction (handler and queue) to inside worker_main() and pass --no-preload (or equivalent) to your app server. That avoids the parent-construct hazard entirely.

Redis memory sizing for deduplication and replay metadata

When deduplication is enabled, each distinct dedup key creates one Redis string for message_deduplication_log_ttl_seconds (default: 3600 seconds). The dedup key is whatever your get_deduplication_key callable returns, so choose a short, stable logical ID and size Redis for:

peak_unique_publish_rate_per_second
* message_deduplication_log_ttl_seconds
* bytes_per_dedup_key

Use 200 bytes per dedup key as a conservative starting point for short queue names, then validate with MEMORY USAGE in your Redis version. Example: 1,000 unique messages/s * 3,600s * 200 B ~= 720 MB for dedup markers alone. A 24h dedup window at the same rate is 86.4M keys, or roughly 17 GB before message payload lists, lease metadata, completed/failed queues, and allocator fragmentation.

Operation-result replay keys are normally deleted after a successful call, but may live until their TTL after ambiguous connection drops or failed cleanup deletes. With visibility timeouts, active claims also store replay metadata until ack or reclaim. Without visibility timeouts, abandoned claims leave claim_result_ids and claim_result_backrefs fields until the message is acked or manually cleaned.

max_completed_length and max_failed_length only bound the completed/failed lists. They do not bound deduplication keys or replay metadata.

Avoid sharing queue Redis DBs with unrelated high-cardinality workloads. If idempotency matters, prefer explicit capacity planning and noeviction with alerts over LRU/random eviction policies: evicting dedup/replay keys before their TTL can weaken duplicate suppression and retry result replay.

Observability

Queue instances accept an optional on_event callback for metrics, tracing, or structured logging. The sync queue expects a regular callable; the async queue expects an async callable:

from redis_message_queue import QueueEvent, RedisMessageQueue

def on_event(event: QueueEvent) -> None:
    ...

queue = RedisMessageQueue("jobs", client=client, on_event=on_event)

Events cover publish, dedup hits, claim/empty polls, reclaim, ack/nack, completed/failed cleanup, DLQ moves, heartbeat renewal, stale leases, cleanup and trim failures, and retry attempts. Callback exceptions are logged and reported with RuntimeWarning, but never propagate into queue operations. on_event is telemetry only: use it for metrics, tracing, and logging, not for sagas, follow-up writes, billing callbacks, or other correctness-critical work. Package logs remain diagnostic; use on_event rather than log parsing for metrics.

from opentelemetry import trace
from prometheus_client import Counter
from redis_message_queue import QueueEvent, RedisMessageQueue

events_total = Counter(
    "rmq_events_total",
    "redis-message-queue lifecycle events",
    ["queue", "operation", "outcome", "exception_type"],
)
SPAN_SINK_TRUSTED = False

def observe(event: QueueEvent) -> None:
    events_total.labels(
        event.queue, event.operation, event.outcome, event.exception_type or ""
    ).inc()
    if event.error is not None and SPAN_SINK_TRUSTED:
        trace.get_current_span().record_exception(event.error)

queue = RedisMessageQueue("jobs", client=client, on_event=observe)

⚠ Secrets in event.error

event.error is the actual exception object — it retains the exception message, __cause__ chain, and traceback. These can contain sensitive content: Redis credentials in connection-error messages, message payloads in handler exceptions, environment values in stack-frame locals.

When exporting to telemetry sinks (OpenTelemetry, Sentry, Datadog), prefer the redaction-friendly event.exception_type for metrics and labels. Use event.error for full structured error data ONLY if your sink is trust-equivalent to your application logs and is access-controlled.

Recommended pattern:

def on_event(event: QueueEvent) -> None:
    # Metric labels — always safe (just the exception class name)
    metric_counter.labels(
        operation=event.operation,
        outcome=event.outcome,
        exception_type=event.exception_type or "none",
    ).inc()

    # Full exception — only if your span sink is trusted
    if event.error is not None and SPAN_SINK_TRUSTED:
        span.record_exception(event.error)

Event dispatch context

Callbacks fire inline:

• Sync queue: the callback runs in the caller's thread. It sees contextvars, the OpenTelemetry current span, and structlog contextvars bound by the caller.
• Async queue: the callback is awaited in the current asyncio task. It has the same contextvars, span, and structlog visibility.
• Sync heartbeat: heartbeat events fire from a separate threading.Thread. That thread does not inherit caller contextvars or the caller's OpenTelemetry current span. Use event.message_id and event.lease_token_hash for correlation.
• Async heartbeat: heartbeat events fire from an asyncio task. The task copies the context present when the heartbeat was started, so contextvars and OpenTelemetry spans bound at handler entry are visible.

Event timing vs. Redis commit

Most events are post-commit, emitted after the Redis command or Lua script returned: publish/success, publish_dedup_hit, claim/success, claim_empty, claim_reclaim, ack, nack, completed, dlq, lease_renew, trim_failed, and stale_lease_*.

Pre-commit and mid-flight exceptions:

• failed/failure fires after the handler raises but before failed-queue cleanup completes. Use nack for cleanup-commit metrics; use failed for handler-exception attribution.
• retry_attempt/failure and retry_exhausted fire on the claim-loop retry path. The first Redis attempt may or may not have committed.
• publish/failure, claim/failure, and cleanup_failed/failure follow exceptions. Under an ambiguous lost response, Redis may have committed despite the exception. Treat them as "operation did not succeed from the caller's perspective", not "Redis did not commit".

Intentionally silent paths

The following operations have no on_event surface by design:

• B1 Cluster pcall cleanup failure: three lease-aware Lua scripts wrap a data-derived DEL in redis.pcall(...) and ignore the result. This preserves queue safety on Cluster CROSSSLOT rejection but cannot be observed through on_event. Operators watching key-TTL behavior or Redis slow logs can detect orphans.
• VT claim-store OOM compensation: if the visibility-timeout Lua script cannot store the claim result, it removes the message from processing, pushes it back to pending, and returns false. Python translates that into claim_empty/skipped, the same shape as an empty poll. This is intentional fail-safe behavior; the message is not lost.
• drain() / close() / aclose() lifecycle: explicit shutdown operations do not emit lifecycle events. Pending-claim-drain recovery work counts as claim_reclaim events when reached.
• Non-claim-loop retry attempts: tenacity retries in deduplicated publish, ack/remove, move-to-completed/failed, and lease renewal collapse into the terminal operation's failure event. There is no per-attempt event for those paths.

The public exception hierarchy is rooted at RedisMessageQueueError. Configuration value/combinations raise ConfigurationError (also a ValueError), custom gateway contract violations raise GatewayContractError (also a TypeError), and Lua redis.error_reply(...) failures raise LuaScriptError (also a redis-py ResponseError). Publish overload raises QueueBackpressureError; publish after explicit drain raises QueueDrainedError. CleanupFailedError and RetryBudgetExhaustedError are reserved categories for cleanup and retry surfaces.

Known limitations

• Timed waits use polling claim loops. To make claims recoverable after ambiguous connection drops, wait_for_message_and_move() uses idempotent Lua claim polling instead of raw blocking list-move commands. This adds a small polling cadence during timed waits.

• Redis Lua is atomic, not rollback-transactional. The built-in scripts now preflight queue key types and fail closed on WRONGTYPE before mutating queue state, but Redis does not undo earlier writes if a later script command fails for another reason (for example OOM under severe memory pressure).

• Batch reclaim limit of 100. The visibility-timeout reclaim Lua script processes at most 100 expired messages per consumer poll. Under extreme backlog this may delay recovery, but prevents any single poll from blocking Redis.

• Claim-attempt loop limit of 100 per poll. The VT claim Lua script attempts at most 100 LMOVE+delivery-count checks per invocation. Under pathological conditions (>100 consecutive poison messages in pending), a single poll returns no message even though non-poison messages exist deeper in the queue. Subsequent polls drain the poison batch 100 at a time.

• Cluster detection uses isinstance(client, RedisCluster). Wrapped or instrumented cluster clients that delegate without inheriting will bypass hash-tag validation. Custom gateways should set is_redis_cluster = True explicitly.

• Redis Cluster requires hash tags. The built-in queue uses multiple Redis keys per operation. Wrap the queue name in hash tags (for example {myqueue}) so every generated key lands in the same slot. When you pass a Redis Cluster client to the built-in queue/gateway path, incompatible names are rejected early.

• Non-ASCII payloads use ~2x storage. The default ensure_ascii=True in JSON serialization encodes non-ASCII characters as \uXXXX escape sequences. This is a deliberate compatibility choice.

• Client-side Retry can duplicate non-deduplicated publishes. If you construct your redis.Redis or redis.asyncio.Redis client with retry=Retry(...), redis-py retries ConnectionError / TimeoutError at the connection layer — below this library. Idempotent operations (deduplicated publish(), lease-scoped cleanup) are safe because their Lua scripts replay the original result. add_message() (used by publish() when deduplication=False) is a bare LPUSH by default, or a single non-idempotent Lua enqueue when max_pending_length is set: this library deliberately does not retry it, but a client-level Retry will, and if the server executed the command before the response was lost the message is enqueued twice. redis-py 6.0+ changed the default standalone Redis() / redis.asyncio.Redis() retry policy from None (no retry) to a 3-attempt ExponentialWithJitterBackoff; pass retry=None explicitly if you need strict at-most-once semantics for non-deduplicated publishes, or accept the duplication risk. More broadly, any non-idempotent enqueue path is vulnerable if the connection drops after server execution but before the client receives the response; all other built-in operations (deduplicated publish, lease-scoped ack/move, lease renewal) use replay markers and are safe under client-level Retry.

  import redis
  from redis_message_queue import RedisMessageQueue
  
  # Strict at-most-once for non-dedup messages: disable redis-py's
  # default 3-retry policy explicitly.
  client = redis.Redis(retry=None)
  queue = RedisMessageQueue("jobs", client=client)

  import redis.asyncio as redis
  from redis_message_queue.asyncio import RedisMessageQueue
  
  # Strict at-most-once for non-dedup messages: disable redis-py's
  # default 3-retry policy explicitly.
  client = redis.Redis(retry=None)
  queue = RedisMessageQueue("jobs", client=client)

• Redis Cluster default retry can stack with this library's retry budget. In redis-py 6.0+, RedisCluster() constructs a default ExponentialWithJitterBackoff retry below this library's retry_budget_seconds. If you need a single retry surface, pass retry=Retry(NoBackoff(), 0) to the cluster client or reduce retry_budget_seconds to account for the lower-level retry window.

For a full analysis, see docs/production-readiness.md.

Upgrading

v7 to v8 migration

v8.0.0 removes implicit dedup key generation. Deduplication is opt-in and deduplication=True now requires get_deduplication_key. If you were relying on v7's automatic content hash, provide the equivalent callable explicitly.

Before:

queue = RedisMessageQueue(
    "orders",
    client=client,
    deduplication=True,
)

After, prefer a stable logical ID:

queue = RedisMessageQueue(
    "orders",
    client=client,
    deduplication=True,
    get_deduplication_key=lambda msg: msg["order_id"],
)

To preserve v7 content-hash behavior mechanically:

import hashlib
import json

queue = RedisMessageQueue(
    "orders",
    client=client,
    deduplication=True,
    get_deduplication_key=lambda msg: hashlib.sha256(
        json.dumps(msg, sort_keys=True).encode()
    ).hexdigest(),
)

If you do not need publish-side deduplication, omit deduplication or set deduplication=False.

v6 to v7 migration

v7.0.0 has four breaking changes to check during upgrade.

AC-02: queue event operation/outcome types are StrEnum members. Runtime string comparisons keep working because StrEnum subclasses str, but type-strict code should replace old Literal[...] annotations and raw string constants with enum members.

Before:

from typing import Literal

QueueOperation = Literal["publish", "claim", "ack"]

def record(operation: QueueOperation) -> None:
    if operation == "publish":
        print("published")

After:

from redis_message_queue import EventOperation

def record(operation: EventOperation) -> None:
    if operation is EventOperation.PUBLISH:
        print("published")

AC-03: drained queue instances refuse new publishes. After queue.drain() / queue.close() (sync) or await queue.drain() / await queue.aclose() (async), the same queue instance rejects publish() with QueueDrainedError("queue is drained").

This state is queue-local and process-local; it is not stored in Redis. If a producer must continue publishing after a worker has drained, use a separate RedisMessageQueue(...) instance for that producer lifecycle. During shutdown, catch QueueDrainedError only at boundaries where late publishes are expected and safe to drop or reschedule.

from redis_message_queue import QueueDrainedError

try:
    queue.publish("late shutdown audit event")
except QueueDrainedError:
    # The queue instance already began draining; drop or reschedule elsewhere.
    pass

AC-09: unsafe drop_oldest combinations now fail at construction. These configurations raise ConfigurationError before the queue or gateway is created:

• pending_overload_policy="drop_oldest" with max_pending_length=None: drop_oldest requires max_pending_length to be set. Use a positive max_pending_length to define what can be dropped, or use pending_overload_policy='raise' or 'block' for unbounded queues.
• pending_overload_policy="drop_oldest" with deduplication enabled or get_deduplication_key configured: 'pending_overload_policy=drop_oldest' cannot be used with deduplication because dropped messages leave their deduplication keys in Redis, causing future publishes of the same payload to be silently suppressed. Use 'raise' or 'block' for deduplicated queues, or disable deduplication if 'drop_oldest' is required.
• pending_overload_policy="drop_oldest" with max_delivery_count set: drop_oldest is incompatible with max_delivery_count (set max_delivery_count=None or pick another policy to avoid silent loss of pending DLQ candidates). Use pending_overload_policy='raise' or 'block' when dead-letter handling is required.

AC-16: redis-py is capped below 8.0.0. The package dependency is redis>=5.0.0,<8.0.0 until redis-py 8 RESP3-default behavior is verified. Users on redis-py 7.x and earlier are unaffected. If you installed a redis-py 8.0.0 beta explicitly, downgrade with pip install "redis<8.0.0".

Configuration changes on live queues

Warning: These changes are destructive on live queues. Drain the queue completely before applying them.

• Do not change name or key_separator on a live queue. Both settings define the Redis key namespace. Existing Redis keys become invisible to the new key scheme. Drain the queue completely before changing either value.
• Do not rename dead_letter_queue on a live queue. Existing DLQ records stay in the old list, while new failures route to the new list. Inspect or drain the old DLQ manually before switching names.
• Do not toggle visibility timeout in either direction with messages in processing. Messages claimed by non-VT consumers have no lease metadata, so VT-enabled consumers cannot reclaim them. Disabling VT later orphans existing lease deadline, lease token, and delivery count metadata and removes crash recovery for those in-flight messages. Drain the processing queue first.
• Reducing max_delivery_count retroactively DLQs messages. The delivery count hash persists across restarts. Messages whose accumulated count exceeds the new limit are immediately dead-lettered on next claim.
• Changing max_delivery_count from a number to None leaves delivery metadata behind. The delivery count hash continues to exist but is no longer consulted. Use this only after draining or after planning manual cleanup of the delivery-count hash.
• Changing get_deduplication_key changes the dedup keyspace. Existing dedup records become inert for the duration of their TTL. Drain the queue or clear the old deduplication keys before switching key functions.
• Disabling deduplication has a retention-window overlap. Existing dedup records remain in Redis until their TTL expires, but new publishes bypass them. Republishes that would have been suppressed under the old setting can enqueue during that window.
• Disabling enable_failed_queue stops recording handler failures. Existing failed entries remain in Redis, but new failures are removed from processing without being appended to the failed queue. If max_delivery_count=None is also set, repeated handler failures can be dropped with no DLQ or failed-queue record; see Dead-letter queue.
• Lowering max_completed_length or max_failed_length trims existing history. The next completed or failed move calls LTRIM, so changing None to N or lowering N can immediately reduce historical entries to the new cap.
• Do not switch sync and async gateway instances mid-process while claims are active. Redis state is compatible across deploys, but each gateway instance keeps its own pending claim-recovery IDs. In-flight claim recovery state does not transfer between instances.
• Switching between gateway= and client= can retarget the DLQ. The built-in client= path derives the DLQ from the queue name. If a custom gateway used a different dead_letter_queue, switching paths has the same orphaning impact as renaming the DLQ.

v5 to v6 migration

v6.0.0 is a non-breaking-defaults release that adds new public APIs. v5 code continues to work; v6 adds opt-in features.

New APIs (opt in as needed):

• max_pending_length=N caps pending-list depth; with pending_overload_policy="raise" (default) producers see QueueBackpressureError when the cap is hit; "block" waits up to pending_overload_block_timeout_seconds; "drop_oldest" evicts silently, so use it only when data loss is acceptable.
• queue.drain(timeout=...) (sync) and await queue.aclose(timeout=...) (async) are explicit graceful-shutdown hooks. They refuse new claims and recover pending claim IDs but do not cancel in-flight handlers; join or await your worker separately.
• on_event=callback receives a QueueEvent dataclass for every publish/claim/ack/reclaim/dedup/cleanup lifecycle event. Use it for metrics, tracing, and structured logging. See examples/production/observability.py for the adapter pattern.
• See examples/production/backpressure.py and examples/production/graceful_shutdown.py for sync production patterns, with async siblings under examples/production/asyncio/.

When using a pre-fork app server (gunicorn --preload, uvicorn workers that import the app at master startup), call make_queue() from your worker startup hook - NOT at module import. See Fork safety for why.

New constructor rejections:

• Passing a redis.sentinel.Sentinel manager object now raises at construction. Use sentinel.master_for(name) instead.
• Passing redis.Redis(single_connection_client=True) to the sync built-in gateway now raises. Use a normal pooled redis.Redis client.

Custom gateway migration:

If you subclass AbstractRedisGateway and override renew_message_lease, add a keyword argument:

def renew_message_lease(
    self,
    queue,
    message,
    lease_token,
    *,
    is_interrupted=None,  # NEW in v6: heartbeat passes a stop-signal observer
) -> bool:
    ...

Async gateways need the same signature on async def. Honor is_interrupted.is_interrupted() in your retry loops to stop renewing when the queue is shutting down.

Exception handling:

• Catch QueueBackpressureError around publish if you opt into backpressure.
• A new base class RedisMessageQueueError lets you catch all library-owned exceptions in one place. Existing catches for ValueError, TypeError, redis.RedisError, etc. continue to work because the new subclasses preserve those bases.

v2 to v3 migration

v3.0.0 replaced the retry_strategy: Callable constructor parameter with retry_budget_seconds, retry_max_delay_seconds, and retry_initial_delay_seconds. Users with custom retry strategies should subclass AbstractRedisGateway instead (see Custom gateway).

Running locally

You'll need a Redis server:

docker run -it --rm -p 6379:6379 redis

Try the examples with multiple terminals:

# Two publishers
poetry run python -m examples.send_messages
poetry run python -m examples.send_messages

# Three consumers
poetry run python -m examples.receive_messages
poetry run python -m examples.receive_messages
poetry run python -m examples.receive_messages