Cache stampede
A cache stampede occurs when overlapping requests for one unavailable cached value independently start redundant work to reconstruct it.
Triggers: Expiration, eviction or invalidation can expose a hot key to concurrent misses. A high aggregate hit ratio does not describe the concentration of backend work around that event. [1]
Coordination: A key-bound refill token can authorize cache publication and reject a stale fill after invalidation. Regulating token issuance also suppresses repeated backend fetches. Serving an older value is a separate application freshness decision. [2]
Relations
Citations
- Vattani, Chierichetti and Lowenstein. Optimal Probabilistic Cache Stampede Prevention. PVLDB 8(8), 886-897 (2015), sections 2-5.
- Nishtala et al. Scaling Memcache at Facebook. NSDI 2013, pp. 385-398, section 3.2.1.
Sources & caveats
- Primary-source review, not a reproduction of the systems' throughput or workload measurements.