Presence and online indicators

Business constraint

Product wants a green dot when a user is "online". Support, chat, and gaming all treat this as a soft signal — but sales dashboards treat missing dots as outages. You must bound false positives and keep write load cheap.

Naive v0

Store last_seen in a primary DB row per user. Every heartbeat UPDATEs the row. UI polls every 5s. Works on a laptop demo with 20 users.

Failure drill

At 200k concurrent users with 15s heartbeats you issue ~13k writes/s to the primary. Disconnects leave dots green for minutes. Two tabs fight. A mobile radio blip looks like logout. Support tickets spike when the DB is slow — presence becomes a canary for unrelated load.

Evolution path

Iteration 1

Move heartbeats to an in-memory TTL map (or Redis) keyed by user id. Expire after ttl = heartbeat_interval * miss_budget. UI reads from the TTL store, not the OLTP primary.

Iteration 2

Separate session presence from device presence. Aggregate devices → user. Soft TTL: serve last-known briefly while a refresh is in flight so UI does not flicker on a single missed beat.

Iteration 3

Shard the presence map by user-id hash. Cap write fan-out with coalescing: multiple heartbeats in a window collapse to one write. Measure false-online rate under simulated disconnect storms.

Implementation cut

Ship a process-local map with monotonic expiry first (lab presence). Document that multi-node presence needs sticky routing or a shared TTL store. Never let presence writes share a transaction pool with checkout.

Numbers

Target: heartbeat write path <1% of DB IOPS. False-online under 30s disconnect <5% of sessions. p99 presence read <5ms from TTL store.

Pattern tags

• soft-ttl
• hot-key-tactics

Self-check

1. Why is polling the primary a bad default for presence?
2. What does miss_budget buy you?
3. How do multi-device sessions change the model?
4. What metric catches false-online?
5. When would sticky sessions help presence reads?

Walkthrough

Recordings will appear here when published. Until then, work the failure drill and lab locally — pause after each iteration and write the metric that moved.

Lab package: see /systems/labs/presence/.