Description:
The positions for DDL statements are updated after processing the event on a slave. If a crash happens after processing the event and before updating the position in info repositories, the ddl statement will be re-executed again causing the sql_thread to error.
I know DDL statements are not transactional, but is there any way to fix this. One needs to manually skip the event, but this is a problem with multi threaded slave where it is difficult to see which worker thread failed.
How to repeat:
mtr test reproducing the error
-- source include/master-slave.inc
-- source include/have_innodb.inc
-- let $old_debug = `select @@global.debug;`
connection master;
create table t1 (a int) ENGINE=Innodb;
insert into t1 values(1);
sync_slave_with_master;
-- eval SET GLOBAL debug = '+d,crash_before_update_pos_rli'
-- exec echo "restart" > $MYSQLTEST_VARDIR/tmp/mysqld.2.expect
connection master;
drop table t1;
connection slave;
-- source include/wait_until_disconnected.inc
-- let $rpl_server_number = 2
-- source include/rpl_reconnect.inc
-- eval SET GLOBAL debug = `$old_debug`
-- source include/start_slave.inc
connection master;
-- source include/rpl_end.inc
This requires a debug_execute_if in sql/rpl_slave.cc
diff --git a/sql/rpl_slave.cc b/sql/rpl_slave.cc
index 99e6a25..3d7eaec 100644
--- a/sql/rpl_slave.cc
+++ b/sql/rpl_slave.cc
@@ -3601,6 +3601,7 @@ apply_event_and_update_pos(Log_event** ptr_ev, THD* thd, Relay_log_info* rli)
ev->get_type_str(), explain[reason]));
#endif
+ DBUG_EXECUTE_IF("crash_before_update_pos_rli", DBUG_SUICIDE(););
error= ev->update_pos(rli);
Suggested fix:
Don't know.
Description: The positions for DDL statements are updated after processing the event on a slave. If a crash happens after processing the event and before updating the position in info repositories, the ddl statement will be re-executed again causing the sql_thread to error. I know DDL statements are not transactional, but is there any way to fix this. One needs to manually skip the event, but this is a problem with multi threaded slave where it is difficult to see which worker thread failed. How to repeat: mtr test reproducing the error -- source include/master-slave.inc -- source include/have_innodb.inc -- let $old_debug = `select @@global.debug;` connection master; create table t1 (a int) ENGINE=Innodb; insert into t1 values(1); sync_slave_with_master; -- eval SET GLOBAL debug = '+d,crash_before_update_pos_rli' -- exec echo "restart" > $MYSQLTEST_VARDIR/tmp/mysqld.2.expect connection master; drop table t1; connection slave; -- source include/wait_until_disconnected.inc -- let $rpl_server_number = 2 -- source include/rpl_reconnect.inc -- eval SET GLOBAL debug = `$old_debug` -- source include/start_slave.inc connection master; -- source include/rpl_end.inc This requires a debug_execute_if in sql/rpl_slave.cc diff --git a/sql/rpl_slave.cc b/sql/rpl_slave.cc index 99e6a25..3d7eaec 100644 --- a/sql/rpl_slave.cc +++ b/sql/rpl_slave.cc @@ -3601,6 +3601,7 @@ apply_event_and_update_pos(Log_event** ptr_ev, THD* thd, Relay_log_info* rli) ev->get_type_str(), explain[reason])); #endif + DBUG_EXECUTE_IF("crash_before_update_pos_rli", DBUG_SUICIDE();); error= ev->update_pos(rli); Suggested fix: Don't know.