A zone-aware cache system for distributed databases
summary
Built a zone-aware local cache called zoom store that partitions SSD RBAs into zones and zone-groups, writes data and metadata separately, and uses record-level indexing so one IO returns a full value. It runs on raw block devices (zoned SSDs planned), removes RocksDB-style compaction so latency spikes disappear, enlarges effective cache capacity by about 30%, boosts read QPS (≈50% in tests) and accepts controlled data sacrifice to avoid running out of space.
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 3:10Requirements and challenges
- 8:09Zone-aware cache design
- 13:55Benchmarks and results
- 16:23Conclusion and Q&A
related talks
16mZNS: Avoiding the Block Interface Tax for Flash-based SSDs
52mHow ZNS SSD's Are Changing the Landscape
30mData @Scale Seattle- Muthu Annamalai
51mSix New Important RocksDB Features And Planned Works
12mSpanDB: A Fast, Cost-Effective LSM-tree Based KV Store on Hybrid Storage
57mRocksDB
24mPerformance tuning in BlueStore & RocksDB
54mEVCache: Lowering Costs for a Low-Latency Cache with RocksDB