High read index duration

[flowbehappy]

By hacking the code (doLearnerRead), I am able to eliminate the readIndex operation, to evaluate the cost of it. I run a simple SQL by different threads.

Benchmark SQL: select min(actual_disp_no) from A;
Single thread duration: 230ms
No updates during the benchmark.

1. The more threads, the higher SQL execution duration.
2. The more threads, the higher read index duration. The time spent on the cost of read index can be over 30%.
3. TiFlash cannot fully utilize the CPU resources. only use up to 40% of 40 cores.

[solotzg]

You can change such item to 81920 in tikv and proxy: https://docs.pingcap.com/tidb/stable/tikv-configuration-file#raft-client-queue-size.

[solotzg]

There are a few ways to implement stale-read in tiflash.

pull mode:

• start a loop with interval N_0, N_1, .. N_MAX seconds.
• in each loop task, try to execute learner-read process and compute safe tso for all regions.
• tidb should set start-ts to latest-tso - N sec for special stale-read query.
• tiflash can check whether tso is safe to skip leader-read for queries.

push mode:

• implement stale-read like follower-read process in tikv.
• enable stale-read option in sql.