redsync

Redis-based distributed synchronization primitives for Python. Async API using redis.asyncio.

Features

Core

• Blocking, no polling – Uses Redis BLPOP: the connection blocks on the server until a permit or signal is available. No busy-waiting, no lock + pub/sub overhead.
• Async-first – Built on redis.asyncio; use with async/await.
• Python 3.10+ – Modern Python support.

Semaphore

• N permits – Semaphore count from 1 to 4096 for limiting concurrency across processes.
• Configurable init – LUA (atomic, default) or OPTIMISTIC_LOCKING strategy for creating the permit pool.
• Automatic lifecycle – Built-in watchdog keeps semaphores alive while in use; Redis auto-deletes keys when all clients disconnect or crash.

Event

• Distributed Signal – Simple one-to-one signaling between processes.

TODO

• Automatic leaked permit recovery – Implement a mechanism (e.g. via heartbeats in metadata) to detect and reclaim permits that were leaked because a worker crashed while holding them.
• Other sync primitives – Add more primitives.

Installation

pip install redsync

Or with uv:

uv add redsync

Requirements: Redis server, redis>=5.0.0 (async support).

Semaphore

Usage

import asyncio
from redis.asyncio import Redis
from redsync import RedisSemaphore, RedisSemaphoreTimeoutError

async def main():
    r = Redis()
    sem = await RedisSemaphore.create(r, "my_resource", count=1)

    # acquire() raises RedisSemaphoreTimeoutError on timeout
    try:
        await sem.acquire(timeout=10)
        try:
            # do work
            pass
        finally:
            await sem.release()
    except RedisSemaphoreTimeoutError:
        pass  # handle timeout

    # or use context manager (raises on timeout)
    async with sem:
        # do work
        pass

asyncio.run(main())

N permits and attaching

Use count > 1 to allow N concurrent holders. count must be between 1 and 4096.

from redsync import SemaphoreInitStrategy

# Creator initializes the pool
sem = await RedisSemaphore.create(r, "pool", count=5, semaphore_init_strategy=SemaphoreInitStrategy.LUA)

# Other workers can attach without knowing the count
worker_sem = await RedisSemaphore.attach(r, "pool", timeout=60.0)
print(f"Total permits: {await worker_sem.get_count()}")

await worker_sem.acquire()
# ...
await worker_sem.release()

create vs attach

In a distributed environment, you have two options for connecting to a semaphore:

1. Call create() everywhere (Idempotent): If your application consists of multiple identical worker nodes running the exact same codebase, they can all safely call RedisSemaphore.create(..., count=5). The first worker to execute it will atomically initialize the semaphore, and the rest will instantly validate that their requested count matches the existing one.
2. Call create() once, and attach() elsewhere (Consumer): If your architecture has a central "manager" process that dictates concurrency limits, the manager calls create(..., count=5). The worker processes then call RedisSemaphore.attach(..., timeout=60.0). attach() does not require a count, never initializes the pool, and simply polls until the creator sets it up.

Semantic Intent: While using create() everywhere works perfectly, using attach() cleanly separates your Control Plane (the entity that decides the concurrency limits and creates the resources) from your Data Plane (the workers that just consume the resources). It simplifies worker code because workers do not need to hardcode or know the count beforehand—they just say "Give me a permit for the pool the manager set up."

Init strategies

The semaphore uses a Redis list as a permit pool. The list must be created and filled with count elements before anyone can BLPOP. Two strategies are supported:

	Lua	Optimistic Locking
Idea	Run a script that atomically ensures the list has N elements (if LLEN == 0 then RPUSH N times).	Uses a Redis transaction (WATCH + MULTI/EXEC) to atomically check if the metadata exists, and if not, creates the list and metadata.
Pros	Single atomic op; no extra key; idempotent.	No Lua; perfectly atomic; crash-proof.
Cons	Requires Lua (standard in Redis).	Transaction retry loop in Python code.

Default is SemaphoreInitStrategy.LUA. Use SemaphoreInitStrategy.OPTIMISTIC_LOCKING to avoid Lua.

Event

A distributed version of a one-to-one signal. set() pushes a signal to Redis, and wait() blocks until a signal is available.

Usage

from redsync import RedisEvent, RedisEventTimeoutError

async def worker():
    event = RedisEvent(r, "task_done")

    print("Waiting for signal...")
    await event.wait(timeout=60)
    print("Signal received, continuing work!")

async def trigger():
    event = RedisEvent(r, "task_done")
    await event.set()  # Wakes up exactly one waiter

Characteristics

• One-to-One: Each set() call wakes up exactly one wait() call.
• Persistence: If set() is called when no one is waiting, the signal is stored in Redis until a waiter arrives.
• No Polling: Uses BLPOP for efficient blocking.

Lifecycle (Watchdog)

Every RedisSemaphore instance runs a background watchdog task that periodically extends the TTL of the underlying Redis keys. This ensures:

• Keys stay alive as long as at least one client holds a reference to the semaphore.
• Automatic cleanup when all clients disconnect or crash — Redis expires the keys after lease_ttl seconds with no renewals.
• No thundering herd — each watchdog adds random jitter to its renewal interval so multiple clients don't all hit Redis at the same instant.

Parameter	Default	Description
lease_ttl	300.0 (5 min)	TTL set on Redis keys. The watchdog renews every lease_ttl / 3 seconds (±20% jitter).

# Custom lease TTL
sem = await RedisSemaphore.create(r, "my_resource", count=3, lease_ttl=120.0)

# Always close when done to stop the watchdog immediately
await sem.close()

If close() is not called (e.g. process crash), the keys will naturally expire after lease_ttl seconds.

Exceptions

• RedsyncError - Base exception
• RedisSemaphoreError, RedisEventError - Component base exceptions
• RedisSemaphoreTimeoutError, RedisEventTimeoutError – Timeout occurred
• RedisSemaphoreNotAcquiredError – release() called without acquiring
• RedisSemaphoreCountError – count not in 1–4096
• RedisSemaphoreCountMismatchError – create() count mismatch

API Reference

RedisSemaphore

class RedisSemaphore:
    @classmethod
    async def create(cls, redis_client, name: str, *, count: int = 1,
                    lease_ttl: float = 300.0,
                    semaphore_init_strategy: SemaphoreInitStrategy = SemaphoreInitStrategy.LUA,
                    key_prefix: str = "redsync:semaphore") -> RedisSemaphore

    @classmethod
    async def attach(cls, redis_client, name: str, *, timeout: float | None = 60.0,
                    lease_ttl: float = 300.0,
                    key_prefix: str = "redsync:semaphore") -> RedisSemaphore

    async def get_count(self) -> int | None

    async def acquire(self, timeout: float | None = None) -> None  # None = block until available
    async def release(self) -> None
    async def close(self) -> None  # Stop the watchdog task
    async def __aenter__(self) -> RedisSemaphore
    async def __aexit__(...) -> None

• name – Semaphore identifier (shared across processes).
• count – Number of permits (1–4096).
• lease_ttl – TTL in seconds for Redis keys; the watchdog renews every lease_ttl / 3 seconds.
• timeout – For acquire(): seconds to wait; None blocks indefinitely. Raises RedisSemaphoreTimeoutError on timeout.

RedisEvent

class RedisEvent:
    def __init__(
        self, 
        redis_client, name: str, *,
        key_prefix: str = "redsync:event"
    ) -> None
    async def set(self) -> None
    async def wait(self, timeout: float | None = None) -> None
    async def clear(self) -> None
    async def is_set(self) -> bool

Running tests

pytest
# or
uv run pytest

Set REDIS_URL if Redis is not on localhost:6379.

License

MIT License – see LICENSE.