EVCache: Lowering Costs for a Low-Latency Cache with RocksDB
summary
EVCache moves terabytes of replicated RAM cache onto SSD-backed RocksDB to cut cost while keeping latency low: Netflix runs hundreds of TB and trillions of requests per day and reports 70% cost reduction after deploying the Moneta/Manetta stack. Architecture uses a Go proxy (Ren) plus a pneumonic layer that maps memcached ops to RocksDB, uses FIFO SST strategy with full-file bloom filters (no block cache, no compression), shards multiple RocksDB DBs per node, and delivers 95th <1ms in production for key workloads.
AI-written summary · about
chapters (33)
- 0:00Intro
- 0:18SANDY WEXLER
- 1:46What do caches touch?
- 2:42Home Page Request
- 3:43Ephemeral olatile Cache
- 4:37Why Optimize for AWS
- 6:32Architecture
- 8:26Reading (get)
- 9:24Use Case Lookaside Cache
- 10:14Use Case Transient Data Store
- 10:59Use Case: Primary Store
- 12:32Use Case Versioned Primary Store
- 13:39Use Case High Volume && High Availability
- 15:03Pipeline of Personalization
- 17:36Cross-Region Replication Region A
- 18:53Additional Features Key Iteration
- 19:11Cache Warming
- 22:31Old Server
- 23:51New Server
- 24:32Rend
- 27:18Mnemonic
- 27:53Why RocksDB?
- 29:21How we use RocksDB
- 34:19Region-Locality Optimizations
- 37:21FIFO Limitations
- 39:13AWS Instance Type
- 39:52Moneta Per Benchmark (High-Vol Online Requests)
- 41:22Moneta Per Benchmark (cont)
- 45:11Moneta Performance in Production (Batch Systems)
- 46:22Moneta Performance in Prod (High Vol-Online Req)
- 47:44Challenges/Concerns
- 48:35Current/Future Work
- 50:06Open Source
related talks
29mAccelerating Performance: Mastering RocksDB for High Speed Data Processing
57mRocksDB
15mImproving Performance of Flash Based Key-Value Stores Using Storage Class Memory...
24mPerformance tuning in BlueStore & RocksDB
52mRocksDB Performance and Reliability Practics
14mYour favourite second database
19mA zone-aware cache system for distributed databases
37mMoving on from RocksDB to something FASTER