Day 3 of #DistributedDilemmas: Caching accelerates inconsistency

Cache invalidation storms often outweigh hit rates. The fastest cache is sometimes no cache at all.

Exhibit A: DoorDash experienced this issue in their systems. During peak hours and high load, cache misses due to expiration or server restarts led to multiple parallel duplicate reads, causing spikes in network traffic and latency. This behavior, known as cache stampeding, overwhelmed their backend systems.

Exhibit B: Percona’s blog discusses a similar phenomenon where expensive-to-generate cache items with high hit rates can cause server overload when multiple clients simultaneously attempt to recreate the item after a cache miss. This parallel processing can significantly increase response times, making the system slower than if there was no cache at all.

Day 3 on X ↗