TiDB incorrectly converts update statement to point get

[hi-rustin]

Bug Report

Please answer these questions before submitting your issue. Thanks!

1. Minimal reproduce step (Required)

CREATE TABLE golang1 (
`fcbpdt` CHAR (8) COLLATE utf8_general_ci NOT NULL,
`fcbpsq` VARCHAR (20) COLLATE utf8_general_ci NOT NULL,
`procst` char (4) COLLATE utf8_general_ci DEFAULT NULL,
`cipstx` VARCHAR (105) COLLATE utf8_general_ci DEFAULT NULL,
`cipsst` CHAR (4) COLLATE utf8_general_ci DEFAULT NULL,
`dyngtg` VARCHAR(4) COLLATE utf8_general_ci DEFAULT NULL,
`blncdt` VARCHAR (8) COLLATE utf8_general_ci DEFAULT NULL, PRIMARY KEY ( fcbpdt, fcbpsq ));

insert into golang1 values('20230925','12023092502158016','abc','','','','');

create table golang2 (
`sysgrp` varchar(20) NOT NULL,
`procst` varchar(8) NOT NULL,
`levlid` int(11) NOT NULL,PRIMARY key (procst));

insert into golang2 VALUES('COMMON','ACSC',90);
insert into golang2 VALUES('COMMON','abc',8);
insert into golang2 VALUES('COMMON','CH02',6);

UPDATE golang1 a
SET procst =(
CASE WHEN
( SELECT levlid FROM golang2 b WHERE b.sysgrp = 'COMMON' AND b.procst = 'ACSC' )
>
( SELECT levlid FROM golang2 c WHERE c.sysgrp = 'COMMON' AND c.procst = a.procst )
THEN 'ACSC' ELSE a.procst END
) ,
cipstx = 'CI010000',
cipsst = 'ACSC',
dyngtg = 'EAYT',
blncdt= '20230925'
WHERE
fcbpdt = '20230925'
AND fcbpsq = '12023092502158016';

2. What did you expect to see? (Required)

Update success.

3. What did you see instead (Required)

Panic: ERROR 1105 (HY000): runtime error: index out of range [0] with length 0

4. What is your TiDB version? (Required)

mysql> select tidb_version();
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| tidb_version()                                                                                                                                                                                                                                                                               |
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
| Release Version: v6.5.1
Edition: Community
Git Commit Hash: 4084b077d615f9dc0a41cf2e30bc6e1a02332df2
Git Branch: heads/refs/tags/v6.5.1
UTC Build Time: 2023-03-07 16:04:14
GoVersion: go1.19.5
Race Enabled: false
TiKV Min Version: 6.2.0-alpha
Check Table Before Drop: false
Store: tikv |
+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)

[hi-rustin]

It works on v6.3.0:

mysql> explain UPDATE golang1 a
    -> SET procst =(
    -> CASE WHEN
    -> ( SELECT levlid FROM golang2 b WHERE b.sysgrp = 'COMMON' AND b.procst = 'ACSC' )
    -> >
    -> ( SELECT levlid FROM golang2 c WHERE c.sysgrp = 'COMMON' AND c.procst = a.procst )
    -> THEN 'ACSC' ELSE a.procst END
    -> ) ,
    -> cipstx = 'CI010000',
    -> cipsst = 'ACSC',
    -> dyngtg = 'EAYT',
    -> blncdt= '20230925'
    -> WHERE
    -> fcbpdt = '20230925'
    -> AND fcbpsq = '12023092502158016';
+------------------------------------+---------+-----------+----------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------+
| id                                 | estRows | task      | access object                                | operator info                                                                                                                                                |
+------------------------------------+---------+-----------+----------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------+
| Update_17                          | N/A     | root      |                                              | N/A                                                                                                                                                          |
| └─IndexJoin_23                     | 1.25    | root      |                                              | left outer join, inner:IndexLookUp_22, outer key:test.golang1.procst, inner key:test.golang2.procst, equal cond:eq(test.golang1.procst, test.golang2.procst) |
|   ├─Point_Get_36(Build)            | 1.00    | root      | table:golang1, index:PRIMARY(fcbpdt, fcbpsq) |                                                                                                                                                              |
|   └─IndexLookUp_22(Probe)          | 1.00    | root      |                                              |                                                                                                                                                              |
|     ├─IndexRangeScan_19(Build)     | 1.00    | cop[tikv] | table:c, index:PRIMARY(procst)               | range: decided by [eq(test.golang2.procst, test.golang1.procst)], keep order:false, stats:pseudo                                                             |
|     └─Selection_21(Probe)          | 1.00    | cop[tikv] |                                              | eq(test.golang2.sysgrp, "COMMON")                                                                                                                            |
|       └─TableRowIDScan_20          | 1.00    | cop[tikv] | table:c                                      | keep order:false, stats:pseudo                                                                                                                               |
+------------------------------------+---------+-----------+----------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------+
7 rows in set (0.01 sec)

But on v6.5.1:

mysql> explain UPDATE golang1 a
    -> SET procst =(
    -> CASE WHEN
    -> ( SELECT levlid FROM golang2 b WHERE b.sysgrp = 'COMMON' AND b.procst = 'ACSC' )
    -> >
    -> ( SELECT levlid FROM golang2 c WHERE c.sysgrp = 'COMMON' AND c.procst = a.procst )
    -> THEN 'ACSC' ELSE a.procst END
    -> ) ,
    -> cipstx = 'CI010000',
    -> cipsst = 'ACSC',
    -> dyngtg = 'EAYT',
    -> blncdt= '20230925'
    -> WHERE
    -> fcbpdt = '20230925'
    -> AND fcbpsq = '12023092502158016';
+-------------------+---------+------+--------------------------------------------------------+---------------+
| id                | estRows | task | access object                                          | operator info |
+-------------------+---------+------+--------------------------------------------------------+---------------+
| Update_20         | N/A     | root |                                                        | N/A           |
| └─Point_Get_1     | 1.00    | root | table:golang1, clustered index:PRIMARY(fcbpdt, fcbpsq) |               |
+-------------------+---------+------+--------------------------------------------------------+---------------+
2 rows in set (0.00 sec)

[hi-rustin]

The main difference is the v6.5.1 TiDB converted the update statement to a point-get operation.

I first checked which commit introduced this bug, and I found after this commit c6e1982 the plan changed.

Then I dug into the code, and I found that we have different results when we call this code on different versions.

tidb/planner/core/point_get_plan.go

Line 1026 in 350e1c2

tblName, tblAlias := getSingleTableNameAndAlias(selStmt.From)

1. We should always return nil, because we can not convert this SQL to a simple point get.
2. For v6.3.0, we also incorrectly tried to convert it but because we didn't use the clustered index, so it just returned nil.
   We used the wrong schema to resolve indexes.
   

   tidb/planner/core/point_get_plan.go

   Line 1675 in 350e1c2

   newAssign.Expr, err = expr.ResolveIndices(plan.Schema())

3. For v6.5.1, we incorrectly converted it to a point get.
   We also used the wrong schema to resolve indexes.
   

   tidb/planner/core/point_get_plan.go

   Line 1675 in 350e1c2

   newAssign.Expr, err = expr.ResolveIndices(plan.Schema())

So both v6.3.0 and v6.5.1 all have bugs. We should return from here:

tidb/planner/core/point_get_plan.go

Line 1026 in 350e1c2

tblName, tblAlias := getSingleTableNameAndAlias(selStmt.From)

Then we don't need to try to convert to a point get.

[JasonWu0506]

/type regression

[jackysp]

It happened in v5.4.1, I think it may also affect release-6.1.

[hi-rustin]

I tested it on v6.1. It worked on v6.1.

mysql> UPDATE golang1 a
    -> SET procst =(
    -> CASE WHEN
    -> ( SELECT levlid FROM golang2 b WHERE b.sysgrp = 'COMMON' AND b.procst = 'ACSC' )
    -> >
    -> ( SELECT levlid FROM golang2 c WHERE c.sysgrp = 'COMMON' AND c.procst = a.procst )
    -> THEN 'ACSC' ELSE a.procst END
    -> ) ,
    -> cipstx = 'CI010000',
    -> cipsst = 'ACSC',
    -> dyngtg = 'EAYT',
    -> blncdt= '20230925'
    -> WHERE
    -> fcbpdt = '20230925'
    -> AND fcbpsq = '12023092502158016';
Query OK, 1 row affected (0.01 sec)
Rows matched: 1  Changed: 1  Warnings: 0

I think the bug was only triggered after we enabled the index of the cluster. To fully resolve this issue, we need to cherry-pick it. Because in version 6.1, I believe other reasons have obscured this problem.

[kennedy8312]

Regression Analysis
PR caused this regression: #38447

[hi-rustin]

Regression Analysis PR caused this regression: #38447

I don't believe that is the root cause. The problem has existed for a long time. This only triggered the bug but did not introduce it.

[kennedy8312]

Regression Analysis PR caused this regression: #38447

I don't believe that is the root cause. The problem has existed for a long time. This only triggered the bug but did not introduce it.

I agree with you for the long existing root cause and the configuration change exposed the issue. However, the change broke an existing function per user point of view, which meant a functional regression.