Break Out of Redis
summary
A company decided to stop using Redis as a critical datastore after repeated production incidents with Redis Sentinel — full resyncs, slow failovers and Sentinel failing around 30GB/node at 2–10k RPS causing stop‑the‑world pauses. They tested migrating to Scylla: Scylla handled ~21k RPS with p50 ≈1ms and p99 ≈2–4ms, and transactions about 11ms at 2k RPS, but range scans caused tombstone GC issues that required a schema change; the solution works in tests but has not yet been deployed to production.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:10Agenda and Redis experience
- 2:15Redis as primary data source
- 5:05Redis Sentinel problems
- 9:25Capacity planning for cache
- 12:21Redis alternatives: ScyllaDB
- 24:04Consistency in ScyllaDB
- 31:48Transactions in ScyllaDB
- 32:09ScyllaDB testing
- 37:48Why not Redis-compatible DBs
- 41:00Selling new solutions
- 56:29Full Redis replacement
- 1:01:28Q&A
related talks
41mPerformance & Fault Tolerance: Building a Modern Database
59mUsing Redis beyond caching
44mScyllaDB where Cassandra meets Redis!
35mHow to Save on Scaling by Moving from Cassandra to Scylla DB
34mScaling Redis writes with cluster mode
25mCache? Where we're going, we don't need cache!
51mX-Raying Your Redis App: SDK Tuning with a Performance Engineer’s Eye
33mIntroduction to Redis