ALTER TABLE ... ENGINE=MRG_MyISAM is not binlogged as DDL

An ALTER TABLE to MERGE or MRG_MyISAM is binlogged incorrectly. It is not marked as DDL, and therefore is scheduled incorrectly in parallel replication. The MDEV-21107 is one possible symptom of this.

--source include/have_binlog_format_mixed.inc

CREATE TABLE t (i1 int, i2 int, pk int) ;

CREATE TABLE t3 LIKE t ;

ALTER TABLE t3 ENGINE = MERGE UNION (t1,t2);

The GTID for the ALTER in the binlog is missing the "ddl" mark:

#251020 13:02:56 server id 1  end_log_pos 668 CRC32 0xa81c998f  GTID 0-1-3

#251020 13:02:56 server id 1  end_log_pos 787 CRC32 0x68bf0315  Query   thread_id=5     exec_time=0     error_code=0    xid=26

ALTER TABLE t3 ENGINE = MERGE UNION (t1,t2)

The ddl mark is decided by the DID_DDL bit of THD_TRANS::m_unsafe_rollback_flags.
For the ALTER, this gets set in mysql_execute_command():

  if (stmt_causes_implicit_commit(thd, CF_IMPLICIT_COMMIT_BEGIN)) {

...

    thd->transaction->stmt.mark_trans_did_ddl();

However, for the ALTER to MERGE, the flag gets cleared again before binlogging, causing the ddl flag to be lost on the GTID event. This happens in THD_TRANS::reset():

  void reset() {

...

    m_unsafe_rollback_flags= 0;

This gets called through trans_commit_stmt() and trans_commit_implicit() from mysql_alter_table():

  /*

    We do not copy data for MERGE tables. Only the children have data.

    MERGE tables have HA_NO_COPY_ON_ALTER set.

  */

  if (!(new_table->file->ha_table_flags() & HA_NO_COPY_ON_ALTER))

...

  else {

...

    if (trans_commit_stmt(thd) || trans_commit_implicit(thd))

This seems to explain why the bug happens, the HA_NO_COPY_ON_ALTER code branch is dropping the DID_DDL flag and then the GTID event will be binlogged later without the ddl mark.

But I am not sure what is the right way to fix it. Should mysql_alter_table() save and restore the flag across the resetting of the THD_TRANS? Should mysql_alter_table() maybe mark the DID_DDL unconditionally at the end? Should the Gtid_log_event constructor be changed to detect DDL in another way? Could there be other consequences elsewhere in the code for the DID_DDL flag to be lost?

I filed a separate bug for this with a clear description of what the problem is, it is the root cause of MDEV-21107, but that bug has a long discussion that mixes several different issues.