Table gone after concurrent EXCHANGE PARTITION

[mjonss]

Bug Report

Please answer these questions before submitting your issue. Thanks!

1. Minimal reproduce step (Required)

tidb1> create table t (a int);
tidb1> create table tp (a int) partition by hash (a) partitions 3;
tidb1> create table t2 (a int);

tidb2> begin;
tidb2> insert into tp values (1);

tidb1> alter table tp exchange partition p0 with table t;

tidb3> alter table tp exchange partition p0 with table t2; -- fails error 8038, information schema changed

tidb2> rollback;

tidb1> select * from t2; -- Fails: ERROR 1146 (42S02): Table 'test.t2' doesn't exist

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

Failure with out the table removed

3. What did you see instead (Required)

Failure and table t2 no longer found.

4. What is your TiDB version? (Required)

tidb_version(): Release Version: v7.4.0-alpha
Edition: Community
Git Commit Hash: d590ccb14e2e36610f7a1313ed75904088b7d84e
Git Branch: heads/refs/tags/v7.4.0-alpha
UTC Build Time: 2023-08-07 14:24:53
GoVersion: go1.20.7
Race Enabled: false
Check Table Before Drop: false
Store: tikv

[mjonss]

The issue seems to be that the partitionID is set in the ddl client session (ddl_api.go) and then is used as the new tableID when applying the diff. This works for the first alter, but for the second alter, the given partitionID (same as the first one) is then used in applyCreateTable. So the second exchange partition will drop the the 't2' table, but when it will recreate it with the new tableID, it has already been swapped with 't', so it gets the first table instead of the partition/partitioned table, since it has already exchanged with the partition. So later when the second exchange partition tries to validate the rows it fails, since it cannot find any 't2' table (and information_schema.tables shows two 't' tables with the same table_id...).

[mjonss]

Since it is only gone from the table cache, a restart will make the table appear again...