executor: Reduce the memory usage of InsertExec

[b41sh]

What problem does this PR solve?

InsertExec consumes too much memory 14355

What is changed and how it works?

allocate less rows and batch exec.
we can see memory usage of insert 500,000 rows reduce from 2.3GB to 1.7GB

Check List

Tests

• Unit test
• Integration test
• Manual test (add detailed scripts or steps below)

[shenli]

@b41sh Thanks for your amazing work!

[jackysp]

Could you show the difference of go pprof heap? @b41sh

[b41sh]

alloc_space of insertRows drop from 11.45MB to 0.77MB
@jackysp PTAL

[jackysp]

Thank, @b41sh ! But it doesn't seem to be a great improvement. Can you analyze the reason? Because this modification may bring some negative effects, such as making the insert ignore / on duplicate update statement call the BatchGet interface more times. We needs to better balance pros and cons.

[b41sh]

@jackysp @XuHuaiyu @qw4990 thanks for your guidance.
As the main use of memory is in e.Lists, and rows reset does not immediately reclaim memory. This PR will not bring improvement, so I close it.

[XuHuaiyu]

• It's my mistake here. I thought the most of the memory space is consumed during the execution phase, but actually:
  • 79.56%: parse phase and planBuilder phases
  • 8.34%: transaction buffer, and mutation during 2pc

• I bench the heap of branch master and this commit, this commit does not bring an actual reduction of the inuse_space.

branch master

(pprof) list execute
Total: 771.81MB
ROUTINE ======================== github.com/pingcap/tidb/session.(*session).execute in /home/tidb/xhy/GOPATH/src/github.com/pingcap/tidb/session/session.go
         0   659.09MB (flat, cum) 85.39% of Total

this commit

(pprof) list execute
Total: 938.31MB
ROUTINE ======================== github.com/pingcap/tidb/session.(*session).execute in /home/tidb/xhy/GOPATH/src/github.com/pingcap/tidb/session/session.go
         0   825.72MB (flat, cum) 88.00% of Total