Jipamba started as a pilot in one county. When growth hit 40x in six months, the original architecture started showing strain: long-tail Postgres queries, a single-region deployment, and ad-hoc analytics on the production read replica.
We moved the hot path onto materialized views, sharded the activity feed by tenant, and split analytics into a separate warehouse fed by logical replication. P95 latency dropped from 2.4s to 180ms.
The lesson: re-architect before you need to, not after the on-call rotation burns out.