/type performance
/type regression
/sig execution
/severity critical
/label affects-8.0

/remove-label may-affects-7.5
/remove-label may-affects-7.1
/remove-label may-affects-6.5
/remove-label may-affects-6.1
/remove-label may-affects-5.4

In the new CPU profile, we can see that there is additional CPU overhead in the index usage reporter section. This overhead was introduced by commit 7087f70.
We will conduct tests on this commit to confirm whether the regression is indeed introduced by this commit.

Change the sig label to sig/sql-infra

I tried to run sysbench random_points on #50773 and #50643. The change is not significant:

Actually, the variance is too big 🤦 (The upper 3 are #50773, and the next 3 are #50643)

/remove-severity critical
/severity major

/remove-label may-affects-7.5
/remove-label may-affects-7.1
/remove-label may-affects-6.5
/remove-label may-affects-6.1
/remove-label may-affects-5.4

Tracing back to this PR will lead to regression in the following three benchmarks: BenchmarkBatchPointGet, BenchmarkPreparedPointGet, and BenchmarkPointGet.

https://github.com/pingcap/tidb/blob/master/Makefile#L299
make bench-daily TO=/path/to/file.json

To re-produce this issue, I wrote a simpler implementation of select_random_points in go, with client-side connection balancer, and it has a much stabler performance: https://gist.github.com/YangKeao/c02ada99d2602d5e7fb5fb94f0581015

For the commit with index usage, the average QPS (after enough warm-up, though I don't know why warm-up does help) is 58393, and the commit without index usage is 59029. The degradation is around 1.07%.

(Also, I don't know why this go-version is a little bit slower than the lua version, maybe the go client is not as fast as the driver used by sysbench).

I have had some ideas to boost the performance of index usage a little. See #53195

This issue is related to another issue, #51852. All three workloads involve a large number of coprocessor operations. The performance can be fixed through the PR #53429 and #52887. For detailed details, please refer to the PR analysis.