We had a discussion to replace Postgres with Cassandra because “we needed to scale”. 1. Ran a quick workload check + Peak traffic: ~250 reads/sec, ~40 writes/sec + Data size: 60-80 GB + Biggest pain: p95 latency spikes during a couple heavy queries 2. Did a simple cost + complexity comparison + Postgres (RDS) upgrade + read replica: cheaper and boring + Cassandra: more nodes, more ops, more tuning, more weird failure modes + Also you lose strong constraints and ad-hoc querying becomes pain 3. Measured the real bottleneck + Missing index on a high-cardinality filter + One endpoint doing N+1 queries + No connection pooling, app opened too many connections under load + Long transactions causing lock waits 4. Fixed the basics + Added the right composite index + Rewrote the N+1 into one query + Added PgBouncer and sane pool limits + Reduced transaction scope, moved slow work out of the request path 5. Result + p95 dropped from ~800ms to ~120ms + CPU stayed flat, no more connection storms + Zero new infra, zero new operational burden Sometimes the boring solution is the right solution. Most scaling “needs” are just missing indexes, bad query shapes, and no guardrails.