Operations
Deployment¶
Stay on one process until you cannot. Splitting gateway from workers costs you Redis, another failure mode, and more pools.
Settings¶
| Name | Purpose |
|---|---|
Redis noeviction |
When the member cache must be complete. |
| Separate Redis DBs | Isolate event transport from large caches. |
| Loopback metrics | Behind an authenticated reverse proxy if you expose them at all. |
| Explicit shard / cluster counts | Predictable rollouts. |
| Graceful shutdown | Cache and token-scoped files cleared. |
| SIGTERM | Through the client/supervisor lifecycle. Leave grace for tasks. |
Roles¶
| Name | Purpose |
|---|---|
role="monolith" |
Default. One client owns shards, callbacks, optional in-memory cache. No Redis required for that. |
gateway |
Owns shard sockets. May publish events for workers. |
worker |
Consumes events. Does not own gateway sessions. |
redis |
Bus coordination. |
cache_redis |
Cache storage. May be the same service; separate DBs isolate failure. |
| Same namespace | Gateway and worker must agree on IPC/Redis and application code. |
| Shard ownership | Each shard exactly once. |
Redis ACLs, TLS, persistence, eviction, memory headroom, and failover are yours.
Capacity¶
| Name | Purpose |
|---|---|
| Redis pools | One per Python process. Size for the fleet. |
| Cache bounds | Entry counts, not bytes. Measure RAM before raising them. |
| Watch | Dropped writes, queue depth, event lag, worker partitions, REST 429s, reconnects. |
| Forced kill | Last event may replay. Cache may be incomplete. |
| Metrics ports | Consecutive ports for every local cluster. Scrape each process. |
| Host CPU/RSS | Same machine if several processes share a host. Not separate budgets. |