Database and query performance, measured

Architectural invariants

1. No query gets tuned until a trace proves it is slow.
2. Every index is a write tax; it needs a read path to justify it.
3. Caches have budgets: TTL defaults and size limits per key prefix.
4. Payload size is a cost line, so serialization is an architecture decision.

Proof in production

• Cut partner API response times 60% by removing expensive joins, partitioning indexes and pruning payloads.
• Fixed endpoint timeouts using SQL analysis, database performance tools and OpenTelemetry tracing.
• Reduced production cache memory 44% (peak utilization 94% → 56%) and built a shared client library with cluster-aware pooling and sane TTLs.
• Moved cache serialization from Java to Protobuf 3 with zero downtime, saving over $1M a year.
• Co-founded a SQL performance review committee; built storage growth prediction that was 70% more accurate.