TopolearnPub/Sub is the simplest way to push real-time events, such as chat or cache invalidations, to many processes.
Pub/Sub decouples senders from receivers. Subscribers call SUBSCRIBE news (or PSUBSCRIBE news.* for glob patterns) and the connection enters a listening mode. Publishers call PUBLISH news "hello", which returns how many subscribers received it.
Delivery is fire-and-forget, at most once:
- messages are not stored; a subscriber that is offline or reconnecting misses them,
- there is no acknowledgement or replay,
- a slow subscriber whose output buffer grows past the limit is disconnected.
That's fine for ephemeral signals: chat typing indicators, live dashboards, telling every app server to drop a local cache entry, or WebSocket fan-out across servers. When you need durability, replay or work distribution among consumers, use streams.
A subscribed connection can only run subscription commands (in RESP2), so client libraries use a dedicated connection for it. Channel names are global across logical databases.
In a Redis Cluster, classic PUBLISH is broadcast to every node. Sharded Pub/Sub (SSUBSCRIBE/SPUBLISH, Redis 7.0+) maps a channel to a hash slot so messages stay on one shard, which scales much better.
Keyspace notifications (notify-keyspace-events) use Pub/Sub to announce events like key expiry, but they inherit the same fire-and-forget delivery.
Example
import threading, time
import redis
r = redis.Redis(decode_responses=True)
p = r.pubsub()
p.subscribe("news")
p.psubscribe("alerts.*")
def listen():
for msg in p.listen():
if msg["type"] in ("message", "pmessage"):
print(msg["channel"], msg["data"])
threading.Thread(target=listen, daemon=True).start()
time.sleep(0.1)
print(r.publish("news", "hello")) # 1 subscriber received it
r.publish("alerts.disk", "90% full") # matched by the pattern
r.publish("nobody-listens", "lost") # returns 0, message is gone
time.sleep(0.1)
# news hello
# alerts.disk 90% fullRemember
- SUBSCRIBE/PSUBSCRIBE receive, PUBLISH sends and returns receiver count
- At-most-once: no storage, no replay, no acks
- Use a dedicated connection for subscribing
- In a cluster prefer sharded Pub/Sub (SPUBLISH/SSUBSCRIBE)
Watch out for
- Using Pub/Sub as a job queue: jobs published while workers restart are lost.