Connection pooling vs read replicas for a 50k-member forum

Marcus Vance2026-08-31T04:00:00Z

We used to raise max_connections whenever p99 login latency spiked. After about 4_000 concurrent sessions that stopped working — the wait moved to row locks on topic_subscriptions.

What we measured

sqlx pool wait (target under 8ms).
PostgreSQL wait_event for Lock vs IO.
Valkey session GET p99, which stayed flat.

Feed SELECTs now hit a replica. Votes still go to primary. If you have run a forum past 50k DAU, how did you split read-your-writes from the home feed?

5 replies

System Administrator2026-08-31T05:00:00Z

We hit the same wall on connection. Replica lag was the surprise, not CPU.

Permalink

Sarah Chen2026-08-31T06:08:00Z

Staff view: please do not delete the thread if this is a duplicate. Link the canonical Issue instead.

Permalink

Leo Mensah2026-08-31T08:24:00Z

For campus we used entity_type=course and left the template as Question. Search Console looks sane.

Permalink

Maya Lin2026-08-31T07:16:00Z

Posted a Guide recap of this thread so we are not asking people to mine comments.

Permalink

System Administrator2026-08-31T09:32:00Z

Thanks — leaving this up for the next team that searches the same string.

Permalink

Weekly winners

Connection pooling vs read replicas for a 50k-member forum

Marcus Vance·2 months ago· 127 5

We used to raise max_connections whenever p99 login latency spiked. After about 4_000 concurrent sessions that stopped working — the wait moved to row locks on topic_subscriptions.

What we measured

sqlx pool wait (target under 8ms).
PostgreSQL wait_event for Lock vs IO.
Valkey session GET p99, which stayed flat.

Feed SELECTs now hit a replica. Votes still go to primary. If you have run a forum past 50k DAU, how did you split read-your-writes from the home feed?

#rust #postgresql #performance