Prewrite and Commit take a long time

[ocbsearcher]

Performance Questions

What version of TiDB are you using?

• tidb 5.2.1

What's the observed and your expected performance respectively?

• We want to improve the writing performance of tidb. According to the information of tidb dashboard and tidb grafana, prewrite and commit stage is very slow.

Cluster's topology architecture:

1. 4 tidb & 3pd (aws c5.2xlarge 8 cores & 16G)
2. 6 tikv (aws m5.2xlarge 8 cores & 32G) (disk: gp3 & 10000 iops & 800MB/s)

Modified Cluster Params: (https://docs.pingcap.com/tidb/stable/tidb-configuration-file#txn-local-latches)

1. txn-local-latches enabled true
2. txn-local-latches capacity 4196000

Application scenario: Import data, write more and read less. Some businesses use optimistic transactions (auto commit), and some conflicting scenarios use pessimistic transactions.

Under this configuration, the writing performance of tidb is very slow. Can you give me some tuning methods? Thanks!

---

TiDB Grafana

scheduler prewrite

scheduler commit

tikv trouble shooting hot write

tikv trouble shooting hot read

(During this time period, there are basically no read requests, and almost all requests are for writing data. )

tikv trouble shotting write too slow

tikv node exporter

tikv summary

tidb summary

TiDB Dashboard

cluster qps

commit mean latency

prewrite is very slow

key visulizer