Realtime Architecture Notes

Boring connection layer

WebSocket servers should do one thing: hold connections and forward frames. Business logic belongs upstream in consumers that write to pub/sub channels.

Write path vs read path

Client → API → Kafka → Consumer → Redis PUBLISH → WS pod → Client

Kafka gives you replay and backpressure. Redis gives you fan-out speed. Don't skip the durable log — you'll need it for reconnect snapshots.

Partitioning strategy

• Channel key = roomId or userId depending on isolation needs.
• Edge pods subscribe to a hash slot subset to limit Redis fan-in.

Deploy without dropping messages

1. Mark pod as draining (stop new connections).
2. Send GOAWAY with reconnect hint to existing clients.
3. Wait for connection count → 0 or timeout.
4. Kill pod.

Metrics that matter

• Publish lag (Kafka → Redis)
• Per-pod connection count
• Reconnect rate after deploy

If reconnect spikes, your drain window is too aggressive.