How We Make TiKV- a Distributed Storage, More Cost-Effective On the Cloud
summary
TiKV hibernates idle Raft regions to cut CPU and heartbeat traffic, and migrating nodes to ARM yields about 20% lower power cost for similar performance. Replacing RocksDB’s Raft WAL with a custom raft-engine and separating user data from Raft logs reduces disk I/O roughly 10–40% and cuts write amplification. Cross-AZ bandwidth drops by forcing local replica copies so the cluster copies data only once across AZs. Future work targets Lambda compaction and EBS snapshots for backups.
AI-written summary · about
chapters (10)
- 0:00Start
- 6:42Deployment
- 8:25Nothing is free on cloud
- 9:15What can we do to reduce the cost
- 10:08Resource analysis
- 12:55Using ARM
- 13:51Life of a write
- 19:51Reducing cross-AZ write traffic
- 21:42Summary
- 26:25TIDB Architecture
related talks
11mAdapting TiKV for Cloud Storage
40mTiKV Best Practices
40mTiKV Best Practices
18mHow We Reduced Write Latency in TiKV?
24mUsing a Distributed Key-Value Store
29mDistributed SQL How TiDB Modernized Data Management
29mDistributed SQL: How TiDB Modernized Data Management
18mTiFlash: Make TiKV 10x Faster and HTAP-able