executor: fix value swapping and multi-updates in UPDATE statement

[dyzsr]

What problem does this PR solve?

Issue Number: close #11911, close #19137, amendment to #20603 (#20594)

Problem Summary: In UPDATE statement:

• Cannot swap values in either single table or multiple tables. (Can not swap values by a multi-tables update #19137)
• Got wrong results for multi-updates on the same table. (Wrong result when update ... set ... on same table #11911)

What is changed and how it works?

If we have a table:

create table t(a int primary key, b int, c int);
insert into t values (1, 2, 3);

For #19137 (changes in UpdateExec.composeNewRow):

• Evaluate expressions using old row data rather than new values.

update t set a=b, b=a; # b ← 2, a ← 1

For #11911 :

• Modify updateRowKeys to allow the same row be updated once for each table alias (instead of once for each table).

# Originally
update t m, t n set m.c=5, m.b=10, n.b=20; # only updates from m will be written into t;

# Now
update t m, t n set m.c=5, m.b=10, n.b=20; # both updates from m and n will be written into t;

• Use mergedRowData to merge modifications to a row from each table alias.

update t m, t n set m.c=5, m.b=10, n.b=20;
# Before merge
# | as name |        m        |        n        |
# | --------|-----------------|-----------------|
# | column  |  a  |  b  |  c  |  a  |  b  |  c  |
# | old row |  1  |  2  |  3  |  1  |  2  |  3  |
# | new row |  1  |  10 |  5  |  1  |  20 |  3  |
# | flag    |     |  *  |  *  |     |  *  |     |
# 
# Previous updates to the same row of the same table will be merged into current row.
# For example, updates from m will be merged into n,
# | as name |        m        |        n        |
# | --------|-----------------|-----------------|
# | column  |  a  |  b  |  c  |  a  |  b  |  c  |
# | old row |  1  |  2  |  3  |  1  |  10 |  5  |
# | new row |  1  |  10 |  5  |  1  |  20 |  5  |
# | flag    |     |  *  |  *  |     |  *  |     |

• Change the computations order in UpdateExec.updateRows:
  1. Prepare handles... (UpdateExec.prepare)
  2. Compose new row for non-generated columns (composeFunc)
  3. Merge modifications for non-generated columns (UpdateExec.merge)
  4. Compose new row for generated columns using newly evaluated values (composeGeneratedColumns)
  5. Merge modifications for generated columns (UpdateExec.merge)
  6. Write back to table (UpdateExec.exec)

For #20603 :

• Reject the update to a table's primary keys if these exists actual multi-updates to that table (In checkUpdateLists). For example:

update t m, t n set m.a=10, n.b=20; # reject
update t m, t n set m.a=10;         # ok

Related changes

• Need to cherry-pick to the release branch 4.0

Check List

Tests

• Unit test
• Integration test

Side effects

• Performance regression
  • Consumes more CPU
  • Consumes more MEM
• Breaking backward compatibility

Release note

• Support value swapping in UPDATE, e.g. update t set a=b, b=a.
  • This may break the compatibility if you have queries like update t set b=a, c=b, as the assignments will use values from the original row instead of the newly evaluated row.
• Bug fixes for multi-update on the same table, e.g. update t m, t n set m.a=1, n.b=2.
  • Previously, updates from exactly one alias of the same table (m.a or n.b) can be written to storage.
  • Now this has been corrected, both updates can take effects.

[dyzsr]

PTAL @SunRunAway

[lzmhhh123]

Is the result consistent with MySQL?

[dyzsr]

No. It could be more consistent with PG and SQLite as discussed in #19137.

[lzmhhh123]

LGTM

[lzmhhh123]

/cc @lysu

[lysu]

/bench

[XuHuaiyu]

It seems not to be a compatibility-breaker, the behavior is correct for now.

[lysu]

LGTM

[dyzsr]

.

[dyzsr]

/run-all-tests

[dyzsr]

/run-common-test tidb-test=pr/1111
/run-integration-common-test tidb-test=pr/1111

[dyzsr]

/run-unit-test

[dyzsr]

PTAL @lysu @cfzjywxk

[lysu]

LGTM

[lzmhhh123]

/merge

[ti-srebot]

Your auto merge job has been accepted, waiting for:

• 21633
• 20924
• 21179

[ti-srebot]

/run-all-tests