Skip to content

Redis presence store

contrib ยท v0.1.0

Shared presence roster backed by Redis, so presence stays correct across multiple workers. Drop-in replacement for the presence kit's in-memory default.
Install copit add @chanx-kit/redis-presence-store
Import from redis_presence_store
Requires kits presence
Python packages chanx>=2.10.0,<3, redis>=5
Tags presence, redis, scaling
Source kits/redis_presence_store

Shared presence roster backed by Redis, so presence stays correct across multiple workers.

copit add @chanx-kit/redis-presence-store   # also installs: presence

Why

The presence kit defaults to InMemoryPresenceStore, which keeps the roster in the worker process. Run two workers and each one reports only the connections it happens to hold, so clients see a roster that is missing people. This store puts that state in Redis instead.

Use it

from redis.asyncio import Redis

from .ws_kits.presence import PresenceTopic
from .ws_kits.redis_presence_store import RedisPresenceStore


class SharedPresenceTopic(PresenceTopic):
    presence_store = RedisPresenceStore(Redis.from_url("redis://localhost:6379"))

Nothing else changes: it satisfies the same PresenceStore protocol.

How state is kept

One Redis hash per scope, mapping connection id to the member JSON. Membership is derived from that hash rather than tracked in a separate counter, so a worker that dies mid-connection cannot corrupt a count and leave someone permanently "online".

Each write refreshes a TTL (default one hour) on the hash, so a scope abandoned by a crashed worker expires instead of lingering forever.

RedisPresenceStore(redis, key_prefix="myapp:presence", ttl=1800)

Trade-offs

  • Every join and leave costs a round trip; members() costs one HGETALL. For very large scopes, page the roster instead of sending it whole.
  • The TTL is a backstop, not a heartbeat. If you need sub-hour accuracy for crashed workers, lower ttl and re-register connections periodically.
  • add() and discard() read then write without a transaction, so one member's tabs churning across workers at the same instant can duplicate or skip a join/leave announcement. The roster itself self-heals on the next read. If announcements must be exact, make the membership check atomic with a Lua script.