MDEV-38218 : Galera test failure on galera_bf_abort_flush_for_export #552
Conversation
hemantdangi-gc
left a comment
hemantdangi-gc
left a comment
There was a problem hiding this comment.
Problem is that FLUSH TABLES FOR EXPORT is a local operation (i.e
it is not replicated by Galera) but it takes MDL-lock. This
MDL-lock then can conflict with INSERT from other node causing
INSERT to be BF aborted. This depends on timing, if we
have enough time to find that INSERT is waiting MDL-lock
we do UNLOCK TABLES fast enough and avoid BF abort. If not
there will be BF-abort.
Test case is fixed so that no query about number of BF aborts
is counted as it is not stable.
The above two are optimization but not related to error. The error is for query 'SET SESSION wsrep_sync_wait = 0' and not for INSERT :
mysqltest: At line 19: query 'SET SESSION wsrep_sync_wait = 0' failed: ER_QUERY_INTERRUPTED (1317): Query execution was interrupted
The additional error priniting and warning will give more information for issue, but don't think thses changes resolves issue.
Problem was in wsrep_handle_mdl_conflict function that was comparing thd->lex->sql_command variable for granted MDL-lock. In this test case there is first (1) FLUSH TABLES ... FOR EXPORT that will take MDL-lock. However, thd->lex->sql_command is not stored to taken tiket. In test some cases next stament executed is INSERT from other node and as thd->lex->sql_command in (1) is SQLCOM_FLUSH INSERT is forced to wait. In some cases next stament executed is SET, in this case thd->lex_sql_command is changed. Now INSERT is executed and as granted thd sql_command is not anymore SQLCOM_FLUSH, SET (and its previous command FLUSH) is bf aborted. This is possible because SET does not cause implicit commit or take new MDL-locks (note that test had SET autocommit=OFF earlier). Fix is to store thd->lex->sql_command to MDL-ticket and in MDL conflict handling use this stored variable when making decision is request or granted aborted or do we wait granted ticket to be released.
janlindstrom
force-pushed
the
10.11-MDEV-22124
branch
from
f1cdd9b to
8a8f690
Compare
Problem is that FLUSH TABLES FOR EXPORT is a local operation (i.e it is not replicated by Galera) but it takes MDL-lock. This MDL-lock then can conflict with INSERT from other node causing INSERT to be BF aborted. This depends on timing, if we have enough time to find that INSERT is waiting MDL-lock we do UNLOCK TABLES fast enough and avoid BF abort. If not there will be BF-abort.
Test case is fixed so that no query about number of BF aborts is counted as it is not stable. Furthermore, improved error printing and added warning when query is interrupted and there is error in wsrep layer.