Smart ALTER TABLE dropping non-existing tmp table (Re: CREATE INDEX is not committed?)

[Marko Mäkelä]

On Thu, Jun 21, 2007 at 12:42:17PM +0300, Marko Mäkelä wrote:
> Monty,
> 
> On Thu, Jun 21, 2007 at 12:53:51AM +0300, Michael Widenius wrote:
> > Marko> Another thing: mysql_alter_table() is trying to drop a temporary
> table
> > Marko> it did not create:
> > 
> > Marko> err1:
> > Marko>   if (new_table)
> > Marko>   {
> > Marko>     /* close_temporary_table() frees the new_table pointer. */
> > Marko>     close_temporary_table(thd, new_table, 1, 1);
> > Marko>   }
> > Marko>   else
> > Marko>     VOID(quick_rm_table(new_db_type, new_db, tmp_name, FN_IS_TMP));
> > 
> > Marko> I believe that the else branch should be removed altogether, because
> > Marko> new_table==NULL should mean that handler::add_index() was called and
> > Marko> no temporary table was created.
> > 
> > The above branch is also taken if we created a temporary table but was
> > not able to open it.  In this case we have to remove it.
> 
> Okay, I didn't analyze this thoroughly, but our modified InnoDB (which
> implements ha_innobase::add_index()) is complaining that MySQL is deleting
> a non-existing table.  I didn't check if MySQL asked InnoDB to create
> that table, but I strongly doubt it.  I will look at this after my vacation,
> sometime in August.

The problem appears to be that a .frm file is created for the temporary
table, but the table is not created in the storage engine, because
create_info->frm_only == true.  However, here in mysql_alter_table()
MySQL will attempt to remove the table also from the storage engine:

err1:
  if (new_table)
  {
    /* close_temporary_table() frees the new_table pointer. */
    close_temporary_table(thd, new_table, 1, 1);
  }
  else
    VOID(quick_rm_table(new_db_type, new_db, tmp_name, FN_IS_TMP));

Proposed fix: add the parameter bool frm_only to quick_rm_table.

	Marko