Realtime chat with WebSockets, rooms, and presence

Resume bullet (when finished)

“Built a multi-room realtime chat backend in FastAPI + Redis pub/sub, supporting presence, typing indicators, message history, and 500 concurrent WebSocket connections on a single 1-CPU container.”

Locked tech stack

No "choose your language" — analysis paralysis kills completion. Follow the stack to the letter on your first build.

Milestones (6 · ~23h)

1. M1~3h

   WebSocket echo + auth

   `/ws/{room}` accepts WebSocket, validates JWT in query string, echoes messages.

   CHECK BEFORE MOVING ON:

   • Why JWT in query, not header?
   • What's wrong with persistent unauth'd connections?
   $ git commit -m "feat(ws): echo endpoint with JWT auth"
2. M2~3h

   Multi-room broadcast

   Messages to one client fan out to everyone in the same room via in-memory dict.

   CHECK BEFORE MOVING ON:

   • What's the fan-out cost as N rooms × M users grows?
   • Where does in-memory fall apart?
   $ git commit -m "feat(ws): multi-room broadcast (in-memory)"
3. M3~5h

   Redis pub/sub for multi-node

   Replace in-memory fan-out with Redis pub/sub so two API pods can share rooms.

   CHECK BEFORE MOVING ON:

   • What does pub/sub guarantee — and not?
   • Why pub/sub here and not Redis Streams?
   $ git commit -m "feat(ws): Redis pub/sub fan-out"
4. M4~4h

   Persistent history + paging

   Messages saved to Postgres. `GET /rooms/{id}/messages?before=<cursor>` returns the last 50.

   CHECK BEFORE MOVING ON:

   • Why cursor pagination, not offset?
   • What's the right index on the messages table?
   $ git commit -m "feat: persistent history with cursor paging"
5. M5~4h

   Presence + typing

   Online users heartbeat every 10s, expire after 30s. Typing events use a separate channel with TTL.

   CHECK BEFORE MOVING ON:

   • Why heartbeats vs server-side detection?
   • What can go wrong if typing events are persisted?
   $ git commit -m "feat: presence + typing indicators"
6. M6~4h

   Load test + observability

   500 concurrent connections sustained on 1-CPU. Prometheus tracks open conns, broadcast lag, dropped messages.

   CHECK BEFORE MOVING ON:

   • Why is broadcast lag the right SLO?
   • What does p99 of zero typically mean?
   $ git commit -m "ops: load test + Prometheus metrics"

60-second demo storyboard

What you say in the recruiter screen when they ask "tell me about your latest project." Practice it out loud.

1. 0-5s: 'Realtime chat in FastAPI — 500 concurrent connections, multi-room, presence.'
2. 5-25s: open 3 browser tabs, demo typing, presence, and a message round-trip.
3. 25-45s: kill one API pod live, show the other pod still receiving messages (Redis pub/sub).
4. 45-60s: load test 500-conn graph in Grafana.

STAR talking points for behavioral round

Production references — how grown-up systems do this

Self-review rubric (before you claim done)