Different partition locations for the same key when SQL mode is differnt

[lcwangchao]

Bug Report

Please answer these questions before submitting your issue. Thanks!

1. Minimal reproduce step (Required)

> set @@sql_mode='allow_invalid_dates';
> create table t(a datetime primary key) partition by range columns(a) (
    partition p0 values less than ('2020-02-31 00:00:00'),
    partition p1 values less than (MAXVALUE)
)
> insert into t values('2020-03-01 00:00:00')
> set @@sql_mode='';
> insert into t values('2020-03-01 00:00:00')

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

The second insert should report an error because a is an primary key.

3. What did you see instead (Required)

The second insert succeeds. And we can see both of them can be found:

TiDB root@127.0.0.1:test> select * from t;
+---------------------+
| a                   |
+---------------------+
| 2020-03-01 00:00:00 |
| 2020-03-01 00:00:00 |
+---------------------+
2 rows in set
Time: 0.010s
TiDB root@127.0.0.1:test> select * from t partition(p0);
+---------------------+
| a                   |
+---------------------+
| 2020-03-01 00:00:00 |
+---------------------+
1 row in set
Time: 0.005s
TiDB root@127.0.0.1:test> select * from t partition(p1);
+---------------------+
| a                   |
+---------------------+
| 2020-03-01 00:00:00 |
+---------------------+
1 row in set
Time: 0.007s

The located in different partitions

4. What is your TiDB version? (Required)

Master. But release version may also have this issue.

[lcwangchao]

The root cause of the issue seems a little complex. First of all, the expressions in the partition table will be parsed as an expression.Expression object which will be stored as an global cache in memory:

tidb/pkg/table/tables/partition.go

Line 115 in be334af

partitionExpr, err := newPartitionExpr(tblInfo, pi.Type, pi.Expr, pi.Columns, pi.Definitions)

tidb/pkg/table/tables/partition.go

Lines 241 to 245 in be334af

	func newPartitionExpr(tblInfo *model.TableInfo, tp model.PartitionType, expr string, partCols []model.CIStr, defs []model.PartitionDefinition) (*PartitionExpr, error) {
	// a partitioned table cannot rely on session context/sql modes, so use a default one!
	ctx := mock.NewContext()
	dbName := model.NewCIStr(ctx.GetSessionVars().CurrentDB)
	columns, names, err := expression.ColumnInfos2ColumnsAndNames(ctx, dbName, tblInfo.Name, tblInfo.Cols(), tblInfo)

In this case, the upper bound of partition p0 is lt(t.a, cast(2021-02-31 00:00:00, datetime)). In most cases, the second argument of lt will be folded to a datetime constant and the cast will dispear. However, in this case, the fold operation will fail because 2021-02-31 00:00:00 is not a valid datetime when ALLOW_INVALID_DATES is not set in context (this context is a mock context and is not related with the current environment). Then the parsed expression will keep cast in it:

tidb/pkg/expression/constant_fold.go

Lines 232 to 246 in be334af

	value, err := x.Eval(ctx.GetEvalCtx(), chunk.Row{})
	retType := x.RetType.Clone()
	if !hasNullArg {
	// set right not null flag for constant value
	switch value.Kind() {
	case types.KindNull:
	retType.DelFlag(mysql.NotNullFlag)
	default:
	retType.AddFlag(mysql.NotNullFlag)
	}
	}
	if err != nil {
	logutil.BgLogger().Debug("fold expression to constant", zap.String("expression", x.ExplainInfo(ctx.GetEvalCtx())), zap.Error(err))
	return expr, isDeferredConst
	}

After the expression finished building, we started to insert 2020-03-01 00:00:00 into this table. However, the result of lt(t.a, cast(2021-02-31 00:00:00, datetime)) depends on the SQL mode:

• If SQL mode is "ALLOW_INVALID_DATES", the above expression will return false.
• If SQL mode is empty "", the above expression will return NULL because cast(2021-02-31 00:00:00, datetime) returns NULL.

When the compare returns NULL, it will be recognized as true actually:

tidb/pkg/table/tables/partition.go

Lines 1362 to 1366 in 761b592

	if isNull {
	// If the column value used to determine the partition is NULL, the row is inserted into the lowest partition.
	// See https://dev.mysql.com/doc/mysql-partitioning-excerpt/5.7/en/partitioning-handling-nulls.html
	return true // Always less than any other value (NULL cannot be in the partition definition VALUE LESS THAN).
	}

That makes the same row located to the different partitions.

[lcwangchao]

I found this bug when I surveyed the test failure for PR #54101. The quick fix for this issue is to allow invalid dates in context to make the fold operation pass. But I'm not sure whether we have a better way to fix it...