*: print the statement causing write conflicts

[sticnarf]

What problem does this PR solve?

Fixes #14593

What is changed and how it works?

MySQL [test]> select * from t;
+----+------+
| pk | v    |
+----+------+
|  1 |    1 |
|  2 |    2 |
+----+------+
2 rows in set (0.002 sec)

Client 1	Client2
begin optimistic;
update t set v=2 where pk=1;
update t set v=3 where pk=2;
	update t set v=0 where pk=1;
commit;

The result will be:

MySQL [test]> commit;
ERROR 9007 (HY000): Write conflict, txnStartTS=414420145598627841, conflictStartTS=414420147931709441, conflictCommitTS=414420147931709442, key={tableID=5262, handle=1} primary={tableID=5262, handle=1}, conflict cause: update t set v=2 where pk=1 [try again later]

In order to know which statement causes the conflict, we need to store the statement index together with the kv pair. So MemBuffer is modified to support storing *extra bytes associated with keys. Here, the extra bytes are used to store the statement index. We might also enlarge the space and store something else.

Because ticlient does not have access to the statement history, we need to return the conflicting statement index to session. However, errors returned from ticlient don't carry data (only messages), so the conflicting statement index is just stored in tikvTxn and is fetched by the caller.

To add the statement text to error message, I have no other idea but use a dirty way. parser/terror is changed to expose the inner arguments. So I can change the messages in ErrWriteConflict.

Check List

Tests

• Unit test

Code changes

• Has exported function/method change
• Has exported variable/fields change
• Has interface methods change

Side effects

• This PR depends on history no matter whether retry is enabled, so the changes in session: do not keep history when the transaction retry is disabled #11192 is basically reverted. I'm not sure if the memory usage really counts.

[sticnarf]

/bench +sysbench

[sre-bot]

Benchmark Report

Run Sysbench Performance Test on VMs

@@                               Benchmark Diff                               @@
================================================================================
--- tidb: 7ba002e83642c2cfed90de0d6ddbf98376a201f6
+++ tidb: b322dca1dc60d87ea81e95ef7df49215160f774d
tikv: 0f08b6beafe17825329d4a1c00a22db6b6699e6c
pd: 4686d4622ad3ebc7fc267b5c64437129cee356fd
================================================================================
oltp_insert:
    * QPS: 7273.31 ± 0.93% (std=42.47) delta: 0.14% (p=0.517)
    * Latency p50: 17.59 ± 0.92% (std=0.10) delta: -0.16%
    * Latency p99: 30.55 ± 4.56% (std=0.92) delta: 2.78%
            
oltp_point_select:
    * QPS: 42620.12 ± 0.45% (std=134.47) delta: 0.98% (p=0.055)
    * Latency p50: 3.00 ± 0.50% (std=0.01) delta: -0.61%
    * Latency p99: 10.09 ± 0.00% (std=0.00) delta: -1.18%
            
oltp_read_write:
    * QPS: 14508.33 ± 0.25% (std=25.82) delta: -0.25% (p=0.148)
    * Latency p50: 176.86 ± 0.20% (std=0.25) delta: 0.28%
    * Latency p99: 327.60 ± 5.03% (std=9.81) delta: 0.48%
            
oltp_update_index:
    * QPS: 4233.11 ± 0.15% (std=5.24) delta: 0.39% (p=0.411)
    * Latency p50: 30.24 ± 0.16% (std=0.04) delta: -0.38%
    * Latency p99: 52.57 ± 1.20% (std=0.45) delta: -3.31%
            
oltp_update_non_index:
    * QPS: 4702.72 ± 1.03% (std=31.78) delta: -0.30% (p=0.505)
    * Latency p50: 27.21 ± 1.05% (std=0.19) delta: 0.29%
    * Latency p99: 42.25 ± 2.71% (std=1.14) delta: 1.55%
            

[sticnarf]

/run-integration-tests

[jackysp]

removed?

[sticnarf]

This feature requires history to be always saved. As mentioned in the PR description, #11192 is basically reverted.

[jackysp]

After discussions with @sticnarf , we feel that the implementation cost of this pr is a bit high, and the benefits brought by it are not too large for the master and release-3.0, because the new environment is now pessimistic by default. It makes sense for 2.1, but there are a lot of changes and it is not suitable to backport to release 2.1. @coocood and @siddontang PTAL.

[sticnarf]

Close this PR as the benefit does not match its cost.