Redis streaming with redis-py

Implement a Redis event-streaming pipeline in Python with redis-py

This guide shows you how to build a Redis-backed event-streaming pipeline in Python with redis-py. It includes a small local web server built with the Python standard library so you can produce events into a single Redis Stream, watch two independent consumer groups read it at their own pace, and recover stuck deliveries with XAUTOCLAIM after simulating a consumer crash.

Overview

A Redis Stream is an append-only log of field/value entries with auto-generated, time-ordered IDs. Producers append with XADD; consumers belong to consumer groups and read with XREADGROUP. The group as a whole tracks a single last-delivered-id cursor, and each consumer gets its own pending-entries list (PEL) of messages it has been handed but not yet acknowledged. Once a consumer has processed an entry it calls XACK to clear the entry from its PEL; entries left unacknowledged past an idle threshold can be reassigned to a healthy consumer with XAUTOCLAIM.

That gives you:

  • Ordered, durable history that many independent consumer groups can read at their own pace
  • At-least-once delivery, with per-consumer pending lists and automatic recovery of crashed consumers
  • Horizontal scaling within a group — add a consumer and Redis automatically splits the work
  • Replay of any range with XRANGE, independent of consumer-group state
  • Bounded retention through XADD MAXLEN ~ or XTRIM MINID ~, without a separate cleanup job

In this example, producers append order events (order.placed, order.paid, order.shipped, order.cancelled) to a single stream at demo:events:orders. Two consumer groups read the same stream:

  • notifications — two consumers (worker-a, worker-b) sharing the work, modelling a fan-out worker pool.
  • analytics — one consumer (worker-c) processing the full event flow on its own.

How it works

The flow looks like this:

  1. The application calls stream.produce(event_type, payload) which runs XADD with an approximate MAXLEN ~ cap. Redis assigns an auto-generated time-ordered ID.
  2. Each consumer thread loops on XREADGROUP with the special ID > (meaning "deliver entries this group has not yet delivered to anyone") and a short block timeout.
  3. After processing each entry, the consumer calls XACK so Redis can drop it from the group's pending list.
  4. If a consumer is killed (or crashes) before acking, its entries sit in the group's PEL. A periodic XAUTOCLAIM sweep reassigns idle entries to a healthy consumer.
  5. Anyone — including code outside the consumer groups — can read history with XRANGE without affecting any group's cursor.

Each consumer group has its own cursor (last-delivered-id) and its own pending list, so the two groups in this demo process the same events without coordinating with each other.

The event-stream helper

The RedisEventStream class wraps the stream operations (source):

import redis
from event_stream import RedisEventStream

r = redis.Redis(host="localhost", port=6379, decode_responses=True)
stream = RedisEventStream(
    redis_client=r,
    stream_key="demo:events:orders",
    maxlen_approx=2000,        # retention guardrail
    claim_min_idle_ms=5000,    # XAUTOCLAIM threshold
)

# Producer
stream_id = stream.produce(
    "order.placed",
    {"order_id": "o-1234", "customer": "alice", "amount": "49.50"},
)

# Consumer group + one consumer
stream.ensure_group("notifications", start_id="0-0")
entries = stream.consume("notifications", "worker-a", count=10, block_ms=500)
for entry_id, fields in entries:
    handle(fields)                              # your processing
    stream.ack("notifications", [entry_id])     # XACK

# Recover stuck PEL entries by reaping them into a healthy consumer.
# The textbook pattern: each consumer periodically calls XAUTOCLAIM
# with itself as the target and processes whatever it claimed.
# `ConsumerWorker.reap_idle_pel` wraps that flow; the low-level helper
# `stream.autoclaim(group, target_name)` is also available if you
# want to drive XAUTOCLAIM directly.
result = worker_b.reap_idle_pel()
# result == {"claimed": N, "processed": M, "deleted_ids": [...]}
# deleted_ids are PEL entries whose payload was already trimmed.
# Redis 7+ has already removed those slots from the PEL, so no XACK
# is needed — log them and route to a dead-letter store for audit.

# Replay history (independent of any group's cursor)
for entry_id, fields in stream.replay("-", "+", count=50):
    print(entry_id, fields)

Data model

Each event is a single stream entry — a flat dict of field/value strings — with an auto-generated time-ordered ID:

demo:events:orders
  1716998413541-0   type=order.placed     order_id=o-1234   customer=alice  amount=49.50  ts_ms=...
  1716998413542-0   type=order.paid       order_id=o-1234   customer=alice  amount=49.50  ts_ms=...
  1716998413542-1   type=order.shipped    order_id=o-1235   customer=bob    amount=12.00  ts_ms=...
  ...

The ID is {milliseconds}-{sequence}, monotonically increasing within the stream, so you can range-query by approximate wall-clock time without an extra index. (IDs are ordered within a stream, not across streams — two events appended to different streams at the same millisecond can produce the same ID.) The implementation uses:

Producing events

produce_batch pipelines XADD calls in a single round trip. Each call carries an approximate MAXLEN ~ cap so the stream stays bounded as it rolls forward:

def produce_batch(self, events: Iterable[tuple[str, dict]]) -> list[str]:
    pipe = self.redis.pipeline(transaction=False)
    for event_type, payload in events:
        fields = self._encode_fields(event_type, payload)
        pipe.xadd(
            self.stream_key,
            fields,
            maxlen=self.maxlen_approx,
            approximate=True,
        )
    ids = pipe.execute()
    ...
    return list(ids)

The ~ flavour of MAXLEN lets Redis trim at a macro-node boundary, which is much cheaper than exact trimming and is what you want when the cap is a retention guardrail, not a hard size constraint. With 300 events produced and MAXLEN ~ 50, you might end up with 100 entries left — Redis released the oldest whole macro-node and stopped. The next XADD will keep length stable.

If you genuinely need an exact cap (rare), drop approximate=True. The performance difference is significant on busy streams.

Reading with a consumer group

Each consumer in a group runs the same XREADGROUP loop. The special ID > means "deliver entries this group has not yet delivered to anyone":

def consume(
    self,
    group: str,
    consumer: str,
    count: int = 10,
    block_ms: int = 500,
) -> list[Entry]:
    result = self.redis.xreadgroup(
        group,
        consumer,
        {self.stream_key: ">"},
        count=count,
        block=block_ms,
    )
    return _flatten_entries(result)

block_ms makes the call efficient even when the stream is idle: the client parks on the server until either an entry arrives or the timeout expires, so consumers don't busy-loop.

Reading with an explicit ID like 0-0 instead of > does something different — it replays entries already delivered to this consumer name (its private PEL). That is the canonical recovery path when the same consumer restarts: catch up on its own pending entries first, then resume reading new ones.

Acknowledging entries

Once the consumer has processed an entry, XACK tells Redis it can drop the entry from the group's pending list:

def ack(self, group: str, ids: Iterable[str]) -> int:
    ids = list(ids)
    if not ids:
        return 0
    return int(self.redis.xack(self.stream_key, group, *ids))

This is the linchpin of at-least-once delivery: an entry that is never acked stays in the PEL until a claim moves it elsewhere. If your consumer thread crashes between processing and ack, the next claim sweep picks the entry back up. The one caveat is retention: XADD MAXLEN ~ and XTRIM can release the entry's payload even while its ID is still in the PEL. The next XAUTOCLAIM returns those IDs in its deleted list and removes them from the PEL inside the same command — the entry cannot be retried, so the caller should log it and route to a dead-letter store for audit. The example handles this explicitly in _handle_autoclaim further down.

The trade-off is the opposite of pub/sub: a slow or crashed consumer doesn't lose messages, but it does mean your downstream system must be idempotent. If you process an order twice because the first attempt died after the side effect but before the ack, the second attempt must be safe.

Multiple consumer groups, one stream

The big difference between Redis Streams and a job queue is that any number of independent consumer groups can read the same stream. The demo sets up two groups on demo:events:orders:

stream.ensure_group("notifications", start_id="0-0")
stream.ensure_group("analytics",     start_id="0-0")

Each group has its own cursor. Producing 5 events results in notifications and analytics each receiving all 5, with no coordination between them. Within notifications, the work is split across worker-a and worker-b: Redis hands each XREADGROUP call whatever entries are not yet delivered to anyone in the group, so adding a second worker doubles throughput without any rebalance logic.

The start_id="0-0" argument means "deliver everything in the stream from the beginning" — useful in a demo and for fresh groups bootstrapping from history. In production, a brand-new group reading a long-existing stream usually starts at $ ("only events after this point") and uses XRANGE explicitly if it needs history.

Recovering crashed consumers with XAUTOCLAIM

The demo's "Crash next 3" button tells a chosen consumer to drop its next three deliveries on the floor without acking them — the same effect as a worker process dying mid-message. Those entries stay in the group's PEL with their delivery counter incremented. Once they have been idle for at least claim_min_idle_ms, any healthy consumer in the group can rescue them by calling XAUTOCLAIM with itself as the target. ConsumerWorker.reap_idle_pel wraps that pattern:

def reap_idle_pel(self) -> dict:
    claimed, deleted = self.stream.autoclaim(
        self.group, self.name, page_count=100, max_pages=10,
    )
    processed = 0
    for entry_id, fields in claimed:
        try:
            self._handle_entry(entry_id, fields)
            processed += 1
        except Exception as exc:
            print(f"reap failed on {entry_id}: {exc}")
    return {
        "claimed": len(claimed),
        "deleted_ids": deleted,
        "processed": processed,
    }

The underlying stream.autoclaim helper pages through the group's PEL with XAUTOCLAIM's continuation cursor:

def autoclaim(
    self,
    group: str,
    consumer: str,
    page_count: int = 100,
    start_id: str = "0-0",
    max_pages: int = 10,
) -> tuple[list[Entry], list[str]]:
    claimed_all, deleted_all = [], []
    cursor = start_id
    for _ in range(max_pages):
        next_id, claimed, deleted = self.redis.xautoclaim(
            self.stream_key,
            group,
            consumer,
            min_idle_time=self.claim_min_idle_ms,
            start_id=cursor,
            count=page_count,
        )
        claimed_all.extend(claimed)
        deleted_all.extend(deleted or [])
        if next_id == "0-0":
            break
        cursor = next_id
    return claimed_all, deleted_all

A single XAUTOCLAIM call scans up to page_count PEL entries starting at start_id, reassigns the ones idle for at least min_idle_time to the named consumer, and returns a continuation cursor in the first slot of the reply. For a full sweep, loop until the cursor returns to 0-0 (with a max_pages safety net so one call cannot monopolise a very large PEL). The delivery counter is incremented on every claim — after a few cycles you can use it to spot a poison-pill message that crashes every consumer that touches it, and route it to a dead-letter stream so the bad entry stops cycling. (New entries keep flowing past the poison pill — XREADGROUP > still delivers fresh work — but the bad entry's repeated reclaim wastes consumer time and keeps the PEL larger than it needs to be.)

The deleted list contains PEL entry IDs whose stream payload was already trimmed by the time the claim ran (typically because MAXLEN ~ retention outran a slow consumer). XAUTOCLAIM removes those dangling slots from the PEL itself, so the caller does not need to XACK them — but the entries cannot be retried either, so log and route them to a dead-letter store for offline inspection. Redis 7.0 introduced this third return element; the example requires Redis 7.0+ for that reason.

reap_idle_pel is the right primitive for the recovery path because it claims and processes in one step: every entry the call returned is now in this consumer's PEL, so the same consumer is responsible for processing and acking it. In production each consumer thread runs reap_idle_pel periodically (every few seconds, on a timer) so a crashed peer's entries never sit invisibly. The demo exposes it as a manual button so you can trigger the reap after waiting for the idle threshold.

XCLAIM (singular, no auto) does the same thing for a specific list of entry IDs you already have in hand — useful when you want to take ownership of one known stuck entry, or when you need to move a specific consumer's PEL to a peer (the case the demo's "Remove consumer" button handles via handover_pending). XAUTOCLAIM cannot filter by source consumer, so it cannot be used for a per-consumer handover.

Replay with XRANGE

XRANGE reads a slice of history. It is completely independent of any consumer group — no cursors move, no acks happen — so it is safe to call any number of times, from any process:

def replay(
    self,
    start_id: str = "-",
    end_id: str = "+",
    count: int = 100,
) -> list[Entry]:
    return list(self.redis.xrange(
        self.stream_key, min=start_id, max=end_id, count=count,
    ))

The special IDs - and + mean "from the very beginning" and "to the very end". You can also pass real IDs (1716998413541-0) or just the millisecond part (1716998413541, which Redis interprets as "any entry with this timestamp").

Typical uses:

  • Bootstrapping a new projection — read the entire stream from - and build a derived view in another store (a search index, a SQL table, a different cache). Doing this against a consumer group would consume the entries; XRANGE lets you do it without disrupting live consumers.
  • Auditing recent activity — read the last few minutes by ID range without touching any group cursor.
  • Debugging — fetch one specific entry by its ID, or a tight range around an incident timestamp, to see exactly what producers wrote.

The consumer worker thread

ConsumerWorker wraps the XREADGROUP → process → XACK loop in a daemon thread (source):

def _run(self) -> None:
    while not self._stop_event.is_set():
        if self._paused.is_set():
            time.sleep(0.05)
            continue
        try:
            entries = self.stream.consume(
                self.group, self.name, count=10, block_ms=500,
            )
        except Exception as exc:
            print(f"[{self.group}/{self.name}] read failed: {exc}")
            time.sleep(0.5)
            continue

        for entry_id, fields in entries:
            if self.process_latency_ms:
                time.sleep(self.process_latency_ms / 1000.0)
            self._handle_entry(entry_id, fields)

_handle_entry either acks (the normal path) or, when the demo has asked the worker to "crash", drops the entry on the floor and increments a counter so the UI can show what is currently in the PEL waiting to be claimed.

Recovery of stuck PEL entries — this consumer's, after a restart, or another consumer's, after a crash — runs through a separate reap_idle_pel method rather than the read loop. That method calls XAUTOCLAIM with this consumer as the target, then processes whatever was claimed in the same flow as new entries. This is the textbook Streams pattern: each consumer is its own reaper, running XAUTOCLAIM(self) periodically (or on demand) so a crashed peer's entries never sit invisibly in the PEL. The demo's "XAUTOCLAIM to selected" button calls reap_idle_pel on the chosen consumer; in production you would run it from a timer every few seconds.

Note that the worker's main read loop deliberately does not call XREADGROUP 0 to drain its own PEL on every iteration. That would re-deliver every pending entry continuously and reset its idle counter to zero each time, which would keep crashed entries below the XAUTOCLAIM threshold forever. Using XAUTOCLAIM(self) as the recovery primitive — which only fires for entries idle longer than min_idle_time — avoids that whole class of bug.

The pause and crash levers exist only for the demo. A real consumer is just the read-process-ack loop — everything else in this class is instrumentation.

Prerequisites

  • Redis 7.0 or later. XAUTOCLAIM was added in Redis 6.2, but its reply gained a third element (the list of deleted IDs) in 7.0; the example relies on that shape.

  • Python 3.9 or later.

  • The redis-py client. Install it with:

    pip install "redis>=5.0"

If your Redis server is running elsewhere, start the demo with --redis-host and --redis-port.

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 streaming-demo && cd streaming-demo
BASE=https://raw.githubusercontent.com/redis/docs/main/content/develop/use-cases/streaming/redis-py
curl -O $BASE/event_stream.py
curl -O $BASE/consumer_worker.py
curl -O $BASE/demo_server.py

Start the demo server

From that directory:

python3 demo_server.py

You should see:

Deleting any existing data at key 'demo:events:orders' for a clean demo run (pass --no-reset to keep it).
Redis streaming demo server listening on http://127.0.0.1:8083
Using Redis at localhost:6379 with stream key 'demo:events:orders' (MAXLEN ~ 2000)
Seeded 3 consumer(s) across 2 group(s)

By default the demo wipes the configured stream key on startup so each run starts from a clean state. Pass --no-reset to keep any existing data at the key (useful when re-running against the same stream to inspect prior state), or --stream-key <name> to point the demo at a different key entirely.

Open http://127.0.0.1:8083 in a browser. You can:

  • Produce any number of events of a chosen type (or random types). Watch the stream length grow and the tail update.
  • See each consumer group: its last-delivered-id, the size of its pending list, and the consumers in it. Each consumer shows its processed count, pending count, and idle time.
  • Add or remove consumers within a group at runtime to see Redis split the work across the new shape.
  • Click Crash next 3 on a consumer to drop its next three deliveries — the same effect as a worker process dying after XREADGROUP but before XACK. Watch the Pending entries (XPENDING) panel fill up.
  • Wait until the idle time exceeds the threshold (default 5000 ms), pick a healthy target consumer, and click XAUTOCLAIM to selected — the stuck entries are reassigned and the delivery counter increments.
  • Replay (XRANGE) any range to confirm the full history is independent of consumer-group state.
  • XTRIM with an approximate MAXLEN to bound retention. Note that an approximate trim only releases whole macro-nodes — MAXLEN ~ 50 on a small stream may not delete anything; on a 300-entry stream it typically lands at around 100.
  • Click Reset demo to drop the stream and re-seed the default groups.

Production usage

Pick retention by length or by minimum ID

The demo uses MAXLEN ~ on every XADD. Two alternatives are worth considering:

  • MINID ~ <id> — keep only entries newer than an ID. If you want "the last 24 hours", compute the wall-clock cutoff and pass XTRIM MINID ~ <ms>-0. This is the right pattern when retention is time-bounded.
  • No cap on XADD plus a periodic XTRIM job — useful if your producer is hot and the per-XADD work has to stay minimal, or if retention rules are complex (a separate process can also factor in consumer-group lag).

In all three cases the trimming is approximate by default. Use exact trimming (MAXLEN n or MINID id without ~) only when you genuinely need an exact count.

Don't let consumer-group lag silently grow

XINFO GROUPS reports each group's lag (entries the group has not yet read) and pending (entries delivered but not acked). In production, alert on either of these crossing a threshold — a steadily growing pending count usually means consumers are crashing without XAUTOCLAIM running, and a growing lag means consumers can't keep up with producers.

The same applies inside a group: XINFO CONSUMERS reports per-consumer pending counts and idle times, so you can spot one slow consumer holding entries that the rest of the group is waiting on.

Make consumer logic idempotent

XAUTOCLAIM can re-deliver an entry to a different consumer after a crash. If your processing has side effects (sending email, charging a card, updating a downstream store), make sure the same entry processed twice gives the same result — use an idempotency key, an upsert with conditional check, or a once-per-id guard table. Redis Streams cannot give you exactly-once semantics on its own.

Bound the delivery counter as a poison-pill signal

XPENDING returns each entry's delivery count, incremented on every claim. If an entry has been delivered (and dropped) several times, the next consumer is unlikely to fare better. After some threshold — deliveries >= 5, say — route the entry to a dead-letter stream, ack it on the original group, and alert. New entries keep flowing past a poison pill (XREADGROUP > still delivers fresh work), but the bad entry's repeated reclaim wastes consumer time and keeps the PEL bigger than it needs to be — without a DLQ threshold it can also slowly trip retention/lag alerts.

Partition by tenant or entity for scale

A single Redis Stream is a single key, and on a Redis Cluster a single key lives on a single shard. If your throughput exceeds what one shard can handle, partition the stream — for example by tenant ID (events:orders:{tenant_a}, events:orders:{tenant_b}) — so different tenants land on different shards. Hash-tags ({tenant_a}) keep all related streams on the same shard if you need to multi-stream atomically.

Per-entity partitioning (events:order:{order_id}) is the canonical pattern when you treat each entity's stream as the event-sourcing log for that entity: every state change for one order goes on its own stream, which is also bounded in size by the entity's lifetime.

Use a separate consumer pool per group

The demo runs every consumer in one process. In production each consumer group is usually its own deployment — its own pool of pods or VMs — so a slow projection in analytics cannot pull notifications workers off their stream. Each pod runs one consumer thread per CPU core, with XAUTOCLAIM either embedded in the consumer loop (every N reads, claim idle entries to self) or run by a separate reaper.

Don't read with XREAD (no group) and then try to ack

XREAD and XREADGROUP are different mechanisms. XREAD is a tail-the-log read with no consumer-group state — entries are not added to any PEL, and you cannot XACK them. If you want at-least-once delivery and crash recovery, you must read through a consumer group.

XREAD is still useful for read-only tail clients (a UI streaming events, a debugger, a tail -f-style command-line tool). It's just not part of the at-least-once path.

Inspect the stream directly with redis-cli

When testing or troubleshooting, inspect the stream directly to confirm the consumer state is what you expect:

# Stream summary
redis-cli XLEN demo:events:orders
redis-cli XINFO STREAM demo:events:orders

# Group cursors and pending counts
redis-cli XINFO GROUPS demo:events:orders

# Consumers within a group
redis-cli XINFO CONSUMERS demo:events:orders notifications

# Pending entries with idle time and delivery count
redis-cli XPENDING demo:events:orders notifications - + 20

# Tail the stream live (no consumer-group state — like tail -f)
redis-cli XREAD BLOCK 0 STREAMS demo:events:orders '$'

# Replay a range
redis-cli XRANGE demo:events:orders - + COUNT 50

If a group's lag is growing while consumers' idle times are short, consumers are healthy but producers are outpacing them — add more consumers. If pending is growing while lag is small, consumers are receiving entries but not acking them — either they are crashing mid-message or your acking logic has a bug.

Learn more

This example uses the following Redis commands:

See the redis-py documentation for the full client reference, and the Streams overview for the deeper conceptual model — consumer groups, the PEL, claim semantics, capped streams, and the differences with Kafka partitions.

RATE THIS PAGE
Back to top ↑