Embedding cache invalidation

On-call incidentMedium30 minNetflix and Meta style

Your task

Your personalization service fetches precomputed user + item embeddings from an embedding store (cache in front of a database) before scoring. At 18:00 p99 latency on the embedding-fetch step jumps from 15ms to 600ms and the backing DB shows a read-QPS spike. Dashboards: the embedding cache hit-rate fell from 98% to 60% at 18:00; a job that nightly regenerates and bulk-reuploads all item embeddings ran early today at 18:00 and rewrote every item key (changing values), and the cache evicted/invalidated en masse. No app deploy. Triage and mitigate.

It covers

  • Embedding cache
  • Hot key
  • Cache miss
  • Latency