Characterizing, Modeling, and Benchmarking RocksDB Key-Value Workloads at Facebook
summary
Open-sourced trace collector captures RocksDB GET/PUT/DELETE/MERGE/ITERATOR traces from three Facebook workloads (UDB, ZippyDB, UP2X) and analyzers reveal that key sizes are narrow while values are long‑tailed, UDB shows a strong diurnal pattern, UP2X uses MERGE for >90% of ops, and under 5% of keys receive most accesses (<5% hot data). YCSB produces many more reads and low block-cache hits because it scatters hot keys; a key-range benchmark groups hot keys and matches storage I/O and cache behavior much better.
AI-written summary · about
chapters (30)
- 0:00Introduction
- 0:30Research Topic
- 0:42Research Area
- 1:03Challenges
- 2:15Methodology
- 2:37Benchmarking
- 6:03Production Use Cases
- 8:01User Profile Use Case
- 9:11Correct Conversation
- 9:42Column Families
- 9:59UDP Column Families
- 10:35KeySize Variation
- 11:25KeySize Distribution
- 12:18CDLCDF
- 12:55Intensive Needs
- 14:02Access Count
- 14:54Use Case
- 15:53Heat Map
- 18:13Storage Isles
- 18:54Data Capture
- 19:33Results
- 20:32Heatmap
- 21:39Key Ranges
- 22:07Final Results
- 22:48Conclusion
- 23:13Use Cases
- 23:33Weiss SP
- 23:51Future Improvements
- 24:32Ads
- 24:57Replay
related talks
52mRocksDB Performance and Reliability Practics
57mRocksDB
47mLessons Learned from Our Main Database Migrations at Facebook
11mEvolution of Development Priorities in Key-value Stores Serving Large-scale Applications
39mLessons Learned from Our Main Database Migrations
51mMyRocks deep dive and production deployment at Facebook
30mData @Scale Seattle- Muthu Annamalai
24mPerformance tuning in BlueStore & RocksDB