By Net Innovix · Published on 2026-06-02 · Engineering
The caching, query and connection-pooling changes that took FinFlow p95 latency from 940ms to 300ms in six weeks. No rewrite required.
When FinFlow came to us their core ledger API was buckling under peak-hour traffic. Rather than propose a risky full rewrite, we ran a two-week profiling sprint to find the real bottlenecks: N+1 queries in the transaction history endpoint, a missing read-replica strategy, and connection pool exhaustion under burst load.
We introduced a tiered caching layer (Redis for hot balances, materialized views for reporting queries), moved reporting reads to dedicated replicas, and rightsized the connection pool with PgBouncer. Every change shipped behind a feature flag and was validated against production shadow traffic before rollout.
The result: p95 latency dropped from 940ms to 300ms, database CPU utilization fell by 40%, and the team avoided a six-month rewrite that would have frozen feature development. The bigger lesson — most "we need a rewrite" conversations are actually "we need better observability" conversations.