Summary
y/hub builds its Redis connections with redis.createClient() only (src/stream.js:166 in v0.7.0), which cannot follow MOVED redirects. Managed clustered Redis offerings (AWS MemoryDB, clustered ElastiCache, most enterprise Redis) therefore cannot host y/hub even when all keys are hash-tagged into one slot — the standalone client still needs to discover which node owns the slot.
A second, subtler blocker: the Lua scripts interpolate the worker-stream and disabled-set key names into the script text with NUMBER_OF_KEYS: 1 (src/stream.js:171-276), so Redis Cluster cannot validate the keys the script touches; they must be declared via KEYS[].
What we run in production (offered as a PR)
redis.clusterMode?: boolean config; when set, the primary connection is built with createCluster() using the configured URL as root node and the normalized standalone options as per-node defaults (credentials, TLS/SNI, reconnect strategy carry over).- The blocking subscription connection is created with
primary.duplicate() so it preserves standalone vs cluster mode.
- Startup validation: cluster mode requires database 0 and a non-empty hash tag in
redis.prefix (all y/hub keys must share a slot because _runSub reads every subscribed room in one multi-key XREAD).
- Every Lua-accessed key declared in
KEYS[] (addMessage, trimMessages, disableCompaction, enableCompaction).
- Keyless / pattern commands routed to the shard that owns the namespace:
SCAN-based key listing routed via a hash-tagged key, and TIME routed the same way so minted timestamps share the stream-id clock domain.
Standalone mode remains the default with byte-identical behavior. We also have a test harness that runs the suite against a real three-primary Redis Cluster with a deliberately wrong seed node (to exercise MOVED-following), plus a standalone regression smoke — happy to contribute both.
Summary
y/hub builds its Redis connections with
redis.createClient()only (src/stream.js:166 in v0.7.0), which cannot followMOVEDredirects. Managed clustered Redis offerings (AWS MemoryDB, clustered ElastiCache, most enterprise Redis) therefore cannot host y/hub even when all keys are hash-tagged into one slot — the standalone client still needs to discover which node owns the slot.A second, subtler blocker: the Lua scripts interpolate the worker-stream and disabled-set key names into the script text with
NUMBER_OF_KEYS: 1(src/stream.js:171-276), so Redis Cluster cannot validate the keys the script touches; they must be declared viaKEYS[].What we run in production (offered as a PR)
redis.clusterMode?: booleanconfig; when set, the primary connection is built withcreateCluster()using the configured URL as root node and the normalized standalone options as per-node defaults (credentials, TLS/SNI, reconnect strategy carry over).primary.duplicate()so it preserves standalone vs cluster mode.redis.prefix(all y/hub keys must share a slot because_runSubreads every subscribed room in one multi-keyXREAD).KEYS[](addMessage,trimMessages,disableCompaction,enableCompaction).SCAN-based key listing routed via a hash-tagged key, andTIMErouted the same way so minted timestamps share the stream-id clock domain.Standalone mode remains the default with byte-identical behavior. We also have a test harness that runs the suite against a real three-primary Redis Cluster with a deliberately wrong seed node (to exercise MOVED-following), plus a standalone regression smoke — happy to contribute both.