tikv/tikv

Reduce prometheus metrics on performance critical path

オープン

#2,581 opened on 2017/12/13

 (1 件のコメント) (0 件のリアクション) (0 人の担当者)Rust (2,189 件のフォーク)batch import
component/performancecomponent/storagehelp wantedtype/enhancement

Repository metrics

Stars
 (15,922 個のスター)
PR merge metrics
 (平均マージ 28d 9h) (30d で 68 merged PRs)

説明

Currently, the bottleneck of RawKV read is the scheduler thread, it runs 100% CPU while the raftstore runs 60%, the network, and disks are not full either.

Based on the flame graph[0] of the scheduler, we can easily find out that Prometheus consumes about 10% CPU time. 10% is not big but definitely not small, there is more room for the read performance!

read-scheduler

[0] https://gist.github.com/overvenus/eb3033344df4fd4a2cf707708a2832b4

コントリビューターガイド