Cache Made Consistent – Cache Invalidation Might No Longer Be a Hard Thing in Computer Science
summary
Meta runs Memcache and TAO serving over one quadrillion queries a day and reports improving TAO cache consistency from six nines to over ten nines. The team builds a Polaris monitoring service that measures client observable invariants and a consistency tracing library that logs only the short postwrite window where races can create stale entries. A recent bug that left cache forever stale is diagnosed and fixed in under 30 minutes, and the effort pushes toward near 100% consistency.
AI-written summary · about
chapters (8)
- 0:00Cache consistency challenges
- 2:32Cache validation importance
- 4:58Cache invalidation hard problem
- 8:13Cache consistency at scale
- 11:41Debugging cache inconsistencies
- 15:37Single cache service perspective
- 18:49Real-world cache inconsistency bug
- 21:41Summary and future work
related talks
21mRAMP-TAO: Layering Atomic Transactions on Meta's Online TAO Data Store
01h 00mBeyond Eventual Consistency
15mAccelerate Application Response Times with True Cache
54mDistributed Caching in Real-Time Payment Systems
29mOracle TimesTen: extreme performance for latency-critical applications
49mCache à la carte: a framework for in-memory caching
11mConsistent Caching in FoundationDB
52mData Consistency at Scale in Meta