Question

Your order service scaled the stateless app tier from 20 to 60 instances to handle holiday load, but write latency got worse, not better. The single primary Postgres is at 95% CPU with high WAL write rate and replication lag climbing on the replicas; read replicas are nearly idle. Connection count on the primary is near max. Adding more app instances now makes it strictly worse. There's no slow query — individual writes are fast, there are just too many. How do you triage, and what's the real ceiling here?