How We Reduced Write Latency in TiKV?
Talkat KubeCon + CloudNativeCon North America 2021 · on Friday, 2021-10-15· at 16:30· Concourse Hall 150 ABC + Online·
summary
Raftstore turns out to be the biggest write-latency bottleneck, so the team extracts disk work into a dedicated IO thread and implements asynchronous ready (not Linux AIO) plus a synchronous-IO coordination path. Benchmarks show async IO raises QPS by >30% and cuts 99th-percentile store-duration to roughly one-third, 99.9th to about half. Applying unpersisted entries (seen in ~1/6 of tests) further trims 99th latency by ~40%. Command-batching and parallel apply are next optimizations.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:36What is TiKV
- 1:17Why we need low average latency
- 1:52Why we need low tail latency
- 2:47Latency of the right phase
- 3:44Roughstore
- 5:16Star Series
- 6:20Source Series
- 6:38Rough Model
- 7:31Sequence Diagram
- 10:53Flowchart
- 11:45Message Duration
- 14:01Results
- 14:59Still duration
- 16:52Benchmark
- 17:16Future Work
- 17:38Outro
related talks
11mAdapting TiKV for Cloud Storage
29mHow We Make TiKV- a Distributed Storage, More Cost-Effective On the Cloud
24mUsing a Distributed Key-Value Store
40mTiKV Best Practices
40mTiKV Best Practices
18mTiFlash: Make TiKV 10x Faster and HTAP-able
29mDistributed SQL How TiDB Modernized Data Management
29mDistributed SQL: How TiDB Modernized Data Management