Unbounded Database Buffering Creates Latency Spikes in High-Throughput Services

A microservice handling 200,000 events per second experienced severe latency issues when using unbounded in-memory buffering for PostgreSQL write-ahead logs. The system accumulated records in a 1 GB buffer before flushing, causing 99th-percentile latency to jump from 2 ms to over 200 ms. Three primary problems emerged: memory pressure triggering writeback pauses, extended fsync latency due to large buffer sizes, and prolonged recovery times during crashes. The solution implemented was capping batch sizes at 64 KB to maintain sub-millisecond fsync latency while preventing unbounded memory growth.