expression: make sleep function response to the kill statement

[tiancaiamao]

What problem does this PR solve?

Fixes #10955

Execute select sleep() in one session, and kill tidb query xxx in another session,
the sleep query is not interrupted.

mysql> select sleep(10);
+-----------+
| sleep(10) |
+-----------+
|         0 |
+-----------+
1 row in set (10.00 sec)

What is changed and how it works?

Check the Killed variable once in a while during the sleep function.

Check List

Tests

• Unit test

• Need to cherry-pick to the release branch

[codecov]

Codecov Report

Merging #10959 into master will decrease coverage by 0.0038%.
The diff coverage is 100%.

@@               Coverage Diff                @@
##             master     #10959        +/-   ##
================================================
- Coverage   80.8878%   80.8839%   -0.0039%     
================================================
  Files           418        418                
  Lines         89184      89187         +3     
================================================
- Hits          72139      72138         -1     
- Misses        11804      11806         +2     
- Partials       5241       5243         +2

[codecov]

Codecov Report

Merging #10959 into master will not change coverage.
The diff coverage is n/a.

@@             Coverage Diff             @@
##             master     #10959   +/-   ##
===========================================
  Coverage   81.0646%   81.0646%           
===========================================
  Files           418        418           
  Lines         89473      89473           
===========================================
  Hits          72531      72531           
  Misses        11716      11716           
  Partials       5226       5226

[tiancaiamao]

/run-all-tests

[tiancaiamao]

/run-all-tests

[tiancaiamao]

/run-unit-test

[tiancaiamao]

PTAL @jackysp @XuHuaiyu

[jackysp]

LGTM

[lysu]

LGTM, period check maybe look good for current code - -

[lysu]

==

[lysu]

@tiancaiamao I think I found a bug 🤔

In MySQL:

select /*+ MAX_EXECUTION_TIME(1000) */ sleep(4), (select sleep(5));
+----------+-------------------+
| sleep(4) | (select sleep(5)) |
+----------+-------------------+
|        1 |                 1 |
+----------+-------------------+
1 row in set (1.00 sec)

but in TiDB:

mysql> select /*+ MAX_EXECUTION_TIME(1000) */ sleep(4), (select sleep(5));
+----------+-------------------+
| sleep(4) | (select sleep(5)) |
+----------+-------------------+
|        1 |                 0 |
+----------+-------------------+
1 row in set (6.09 sec)

it seems we only can kill one operator, but if later operators have time-wasted task, the query cannot return immediately

[tiancaiamao]

@lysu Good job, you find a bug! But this is not true:

it seems we only can kill one operator, but if later operators have time-wasted task, the query cannot return immediately

The real problem is that MAX_EXECUTION_TIME hint can only kill query in execute phase, while in your given query, sleep happens plan build phase.

To verify that, you can try this, this takes 5s:

mysql> explain select /*+ MAX_EXECUTION_TIME(1000) */ 1,  (select sleep(5));
+--------------------+-------+------+---------------+
| id                 | count | task | operator info |
+--------------------+-------+------+---------------+
| Projection_9       | 1.00  | root | 1, 0          |
| └─TableDual_10     | 1.00  | root | rows:1        |
+--------------------+-------+------+---------------+
2 rows in set (5.00 sec)

As a contrast, this query returns immediately:

mysql> explain select /*+ MAX_EXECUTION_TIME(1000) */ 1,  sleep(5);
+-------------------+-------+------+---------------+
| id                | count | task | operator info |
+-------------------+-------+------+---------------+
| Projection_3      | 1.00  | root | 1, sleep(5)   |
| └─TableDual_4     | 1.00  | root | rows:1        |
+-------------------+-------+------+---------------+
2 rows in set (0.01 sec)

When there is a sleep in (select sleep(5)), we'll execute the subquery, this happens in the plan building phase and can not be canceled by MAX_EXECUTION_TIME

[tiancaiamao]

I think we can file an issue, and merge this PR first.

[lysu]

LGTM

[tiancaiamao]

PTAL @lysu