mariadb/mysql-test/suite/rpl/r/rpl_parallel.result

893 lines
26 KiB
Text
Raw Normal View History

include/rpl_init.inc [topology=1->2]
SET @old_parallel_threads=@@GLOBAL.slave_parallel_threads;
SET GLOBAL slave_parallel_threads=10;
ERROR HY000: This operation cannot be performed as you have a running slave ''; run STOP SLAVE '' first
include/stop_slave.inc
SET GLOBAL slave_parallel_threads=10;
CHANGE MASTER TO master_use_gtid=slave_pos;
include/start_slave.inc
*** Test long-running query in domain 1 can run in parallel with short queries in domain 0 ***
ALTER TABLE mysql.gtid_slave_pos ENGINE=InnoDB;
CREATE TABLE t1 (a int PRIMARY KEY) ENGINE=MyISAM;
CREATE TABLE t2 (a int PRIMARY KEY) ENGINE=InnoDB;
INSERT INTO t1 VALUES (1);
INSERT INTO t2 VALUES (1);
LOCK TABLE t1 WRITE;
SET gtid_domain_id=1;
INSERT INTO t1 VALUES (2);
SET gtid_domain_id=0;
INSERT INTO t2 VALUES (2);
INSERT INTO t2 VALUES (3);
BEGIN;
INSERT INTO t2 VALUES (4);
INSERT INTO t2 VALUES (5);
COMMIT;
INSERT INTO t2 VALUES (6);
SELECT * FROM t2 ORDER by a;
a
1
2
3
4
5
6
SELECT * FROM t1;
a
1
UNLOCK TABLES;
SELECT * FROM t1 ORDER BY a;
a
1
2
*** Test two transactions in different domains committed in opposite order on slave but in a single group commit. ***
include/stop_slave.inc
SET sql_log_bin=0;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
RETURN x;
END
||
SET sql_log_bin=1;
SET @old_format= @@SESSION.binlog_format;
SET binlog_format='statement';
SET gtid_domain_id=1;
INSERT INTO t2 VALUES (foo(10,
'commit_before_enqueue SIGNAL ready1 WAIT_FOR cont1',
'commit_after_release_LOCK_prepare_ordered SIGNAL ready2'));
FLUSH LOGS;
SET sql_log_bin=0;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
IF d1 != '' THEN
SET debug_sync = d1;
END IF;
IF d2 != '' THEN
SET debug_sync = d2;
END IF;
RETURN x;
END
||
SET sql_log_bin=1;
SET @old_format=@@GLOBAL.binlog_format;
SET GLOBAL binlog_format=statement;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
SET debug_sync='now WAIT_FOR ready1';
SET gtid_domain_id=2;
INSERT INTO t2 VALUES (foo(11,
'commit_before_enqueue SIGNAL ready3 WAIT_FOR cont3',
'commit_after_release_LOCK_prepare_ordered SIGNAL ready4 WAIT_FOR cont4'));
SET gtid_domain_id=0;
SELECT * FROM t2 WHERE a >= 10 ORDER BY a;
a
10
11
SET debug_sync='now WAIT_FOR ready3';
SET debug_sync='now SIGNAL cont3';
SET debug_sync='now WAIT_FOR ready4';
SET debug_sync='now SIGNAL cont1';
SET debug_sync='now WAIT_FOR ready2';
SET debug_sync='now SIGNAL cont4';
SELECT * FROM t2 WHERE a >= 10 ORDER BY a;
a
10
11
include/show_binlog_events.inc
Log_name Pos Event_type Server_id End_log_pos Info
slave-bin.000002 # Binlog_checkpoint # # slave-bin.000002
slave-bin.000002 # Gtid # # BEGIN GTID #-#-# cid=#
slave-bin.000002 # Query # # use `test`; INSERT INTO t2 VALUES (foo(11,
'commit_before_enqueue SIGNAL ready3 WAIT_FOR cont3',
'commit_after_release_LOCK_prepare_ordered SIGNAL ready4 WAIT_FOR cont4'))
slave-bin.000002 # Xid # # COMMIT /* XID */
slave-bin.000002 # Gtid # # BEGIN GTID #-#-# cid=#
slave-bin.000002 # Query # # use `test`; INSERT INTO t2 VALUES (foo(10,
'commit_before_enqueue SIGNAL ready1 WAIT_FOR cont1',
'commit_after_release_LOCK_prepare_ordered SIGNAL ready2'))
slave-bin.000002 # Xid # # COMMIT /* XID */
FLUSH LOGS;
include/stop_slave.inc
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET debug_sync='RESET';
include/start_slave.inc
*** Test that group-committed transactions on the master can replicate in parallel on the slave. ***
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
FLUSH LOGS;
CREATE TABLE t3 (a INT PRIMARY KEY, b INT) ENGINE=InnoDB;
INSERT INTO t3 VALUES (1,1), (3,3), (5,5), (7,7);
BEGIN;
INSERT INTO t3 VALUES (2,102);
BEGIN;
INSERT INTO t3 VALUES (4,104);
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
SET binlog_format=statement;
INSERT INTO t3 VALUES (2, foo(12,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued1 WAIT_FOR slave_cont1',
''));
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
SET binlog_format=statement;
INSERT INTO t3 VALUES (4, foo(14,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued2',
''));
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued3';
SET binlog_format=statement;
INSERT INTO t3 VALUES (6, foo(16,
'group_commit_waiting_for_prior SIGNAL slave_queued3',
''));
SET debug_sync='now WAIT_FOR master_queued3';
SET debug_sync='now SIGNAL master_cont1';
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SELECT * FROM t3 ORDER BY a;
a b
1 1
2 12
3 3
4 14
5 5
6 16
7 7
include/show_binlog_events.inc
Log_name Pos Event_type Server_id End_log_pos Info
master-bin.000002 # Binlog_checkpoint # # master-bin.000002
master-bin.000002 # Gtid # # GTID #-#-#
master-bin.000002 # Query # # use `test`; CREATE TABLE t3 (a INT PRIMARY KEY, b INT) ENGINE=InnoDB
master-bin.000002 # Gtid # # BEGIN GTID #-#-#
master-bin.000002 # Query # # use `test`; INSERT INTO t3 VALUES (1,1), (3,3), (5,5), (7,7)
master-bin.000002 # Xid # # COMMIT /* XID */
master-bin.000002 # Gtid # # BEGIN GTID #-#-# cid=#
master-bin.000002 # Query # # use `test`; INSERT INTO t3 VALUES (2, foo(12,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued1 WAIT_FOR slave_cont1',
''))
master-bin.000002 # Xid # # COMMIT /* XID */
master-bin.000002 # Gtid # # BEGIN GTID #-#-# cid=#
master-bin.000002 # Query # # use `test`; INSERT INTO t3 VALUES (4, foo(14,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued2',
''))
master-bin.000002 # Xid # # COMMIT /* XID */
master-bin.000002 # Gtid # # BEGIN GTID #-#-# cid=#
master-bin.000002 # Query # # use `test`; INSERT INTO t3 VALUES (6, foo(16,
'group_commit_waiting_for_prior SIGNAL slave_queued3',
''))
master-bin.000002 # Xid # # COMMIT /* XID */
SET debug_sync='now WAIT_FOR slave_queued3';
ROLLBACK;
SET debug_sync='now WAIT_FOR slave_queued1';
ROLLBACK;
SET debug_sync='now WAIT_FOR slave_queued2';
SET debug_sync='now SIGNAL slave_cont1';
SELECT * FROM t3 ORDER BY a;
a b
1 1
2 12
3 3
4 14
5 5
6 16
7 7
include/show_binlog_events.inc
Log_name Pos Event_type Server_id End_log_pos Info
slave-bin.000003 # Binlog_checkpoint # # slave-bin.000003
slave-bin.000003 # Gtid # # GTID #-#-#
slave-bin.000003 # Query # # use `test`; CREATE TABLE t3 (a INT PRIMARY KEY, b INT) ENGINE=InnoDB
slave-bin.000003 # Gtid # # BEGIN GTID #-#-#
slave-bin.000003 # Query # # use `test`; INSERT INTO t3 VALUES (1,1), (3,3), (5,5), (7,7)
slave-bin.000003 # Xid # # COMMIT /* XID */
slave-bin.000003 # Gtid # # BEGIN GTID #-#-# cid=#
slave-bin.000003 # Query # # use `test`; INSERT INTO t3 VALUES (2, foo(12,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued1 WAIT_FOR slave_cont1',
''))
slave-bin.000003 # Xid # # COMMIT /* XID */
slave-bin.000003 # Gtid # # BEGIN GTID #-#-# cid=#
slave-bin.000003 # Query # # use `test`; INSERT INTO t3 VALUES (4, foo(14,
'commit_after_release_LOCK_prepare_ordered SIGNAL slave_queued2',
''))
slave-bin.000003 # Xid # # COMMIT /* XID */
slave-bin.000003 # Gtid # # BEGIN GTID #-#-# cid=#
slave-bin.000003 # Query # # use `test`; INSERT INTO t3 VALUES (6, foo(16,
'group_commit_waiting_for_prior SIGNAL slave_queued3',
''))
slave-bin.000003 # Xid # # COMMIT /* XID */
*** Test STOP SLAVE in parallel mode ***
include/stop_slave.inc
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET binlog_direct_non_transactional_updates=0;
SET sql_log_bin=0;
CALL mtr.add_suppression("Statement is unsafe because it accesses a non-transactional table after accessing a transactional table within the same transaction");
SET sql_log_bin=1;
BEGIN;
INSERT INTO t2 VALUES (20);
INSERT INTO t1 VALUES (20);
INSERT INTO t2 VALUES (21);
INSERT INTO t3 VALUES (20, 20);
COMMIT;
INSERT INTO t3 VALUES(21, 21);
INSERT INTO t3 VALUES(22, 22);
SET binlog_format=@old_format;
BEGIN;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
INSERT INTO t2 VALUES (21);
START SLAVE;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET @old_dbug= @@GLOBAL.debug_dbug;
SET GLOBAL debug_dbug="+d,rpl_parallel_wait_for_done_trigger";
STOP SLAVE;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='now WAIT_FOR wait_for_done_waiting';
ROLLBACK;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET GLOBAL debug_dbug=@old_dbug;
SET debug_sync='RESET';
include/wait_for_slave_to_stop.inc
SELECT * FROM t1 WHERE a >= 20 ORDER BY a;
a
20
SELECT * FROM t2 WHERE a >= 20 ORDER BY a;
a
20
21
SELECT * FROM t3 WHERE a >= 20 ORDER BY a;
a b
20 20
include/start_slave.inc
SELECT * FROM t1 WHERE a >= 20 ORDER BY a;
a
20
SELECT * FROM t2 WHERE a >= 20 ORDER BY a;
a
20
21
SELECT * FROM t3 WHERE a >= 20 ORDER BY a;
a b
20 20
21 21
22 22
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
*** Test killing slave threads at various wait points ***
*** 1. Test killing transaction waiting in commit for previous transaction to commit ***
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
SET binlog_format=statement;
INSERT INTO t3 VALUES (31, foo(31,
'commit_before_prepare_ordered WAIT_FOR t2_waiting',
'commit_after_prepare_ordered SIGNAL t1_ready WAIT_FOR t1_cont'));
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
SET binlog_format=statement;
BEGIN;
INSERT INTO t3 VALUES (32, foo(32,
'ha_write_row_end SIGNAL t2_query WAIT_FOR t2_cont',
''));
INSERT INTO t3 VALUES (33, foo(33,
'group_commit_waiting_for_prior SIGNAL t2_waiting',
'group_commit_waiting_for_prior_killed SIGNAL t2_killed'));
COMMIT;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued3';
SET binlog_format=statement;
INSERT INTO t3 VALUES (34, foo(34,
'',
''));
SET debug_sync='now WAIT_FOR master_queued3';
SET debug_sync='now SIGNAL master_cont1';
SELECT * FROM t3 WHERE a >= 30 ORDER BY a;
a b
31 31
32 32
33 33
34 34
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET sql_log_bin=0;
CALL mtr.add_suppression("Query execution was interrupted");
CALL mtr.add_suppression("Commit failed due to failure of an earlier commit on which this one depends");
CALL mtr.add_suppression("Slave: Connection was killed");
SET sql_log_bin=1;
SET debug_sync='now WAIT_FOR t2_query';
SET debug_sync='now SIGNAL t2_cont';
SET debug_sync='now WAIT_FOR t1_ready';
KILL THD_ID;
SET debug_sync='now WAIT_FOR t2_killed';
SET debug_sync='now SIGNAL t1_cont';
MDEV-5262, MDEV-5914, MDEV-5941, MDEV-6020: Deadlocks during parallel replication causing replication to fail. Remove the temporary fix for MDEV-5914, which used READ COMMITTED for parallel replication worker threads. Replace it with a better, more selective solution. The issue is with certain edge cases of InnoDB gap locks, for example between INSERT and ranged DELETE. It is possible for the gap lock set by the DELETE to block the INSERT, if the DELETE runs first, while the record lock set by INSERT does not block the DELETE, if the INSERT runs first. This can cause a conflict between the two in parallel replication on the slave even though they ran without conflicts on the master. With this patch, InnoDB will ask the server layer about the two involved transactions before blocking on a gap lock. If the server layer tells InnoDB that the transactions are already fixed wrt. commit order, as they are in parallel replication, InnoDB will ignore the gap lock and allow the two transactions to proceed in parallel, avoiding the conflict. Improve the fix for MDEV-6020. When InnoDB itself detects a deadlock, it now asks the server layer for any preferences about which transaction to roll back. In case of parallel replication with two transactions T1 and T2 fixed to commit T1 before T2, the server layer will ask InnoDB to roll back T2 as the deadlock victim, not T1. This helps in some cases to avoid excessive deadlock rollback, as T2 will in any case need to wait for T1 to complete before it can itself commit. Also some misc. fixes found during development and testing: - Remove thd_rpl_is_parallel(), it is not used or needed. - Use KILL_CONNECTION instead of KILL_QUERY when a parallel replication worker thread is killed to resolve a deadlock with fixed commit ordering. There are some cases, eg. in sql/sql_parse.cc, where a KILL_QUERY can be ignored if the query otherwise completed successfully, and this could cause the deadlock kill to be lost, so that the deadlock was not correctly resolved. - Fix random test failure due to missing wait_for_binlog_checkpoint.inc. - Make sure that deadlock or other temporary errors during parallel replication are not printed to the the error log; there were some places around the replication code with extra error logging. These conditions can occur occasionally and are handled automatically without breaking replication, so they should not pollute the error log. - Fix handling of rgi->gtid_sub_id. We need to be able to access this also at the end of a transaction, to be able to detect and resolve deadlocks due to commit ordering. But this value was also used as a flag to mark whether record_gtid() had been called, by being set to zero, losing the value. Now, introduce a separate flag rgi->gtid_pending, so rgi->gtid_sub_id remains valid for the entire duration of the transaction. - Fix one place where the code to handle ignored errors called reset_killed() unconditionally, even if no error was caught that should be ignored. This could cause loss of a deadlock kill signal, breaking deadlock detection and resolution. - Fix a couple of missing mysql_reset_thd_for_next_command(). This could cause a prior error condition to remain for the next event executed, causing assertions about errors already being set and possibly giving incorrect error handling for following event executions. - Fix code that cleared thd->rgi_slave in the parallel replication worker threads after each event execution; this caused the deadlock detection and handling code to not be able to correctly process the associated transactions as belonging to replication worker threads. - Remove useless error code in slave_background_kill_request(). - Fix bug where wfc->wakeup_error was not cleared at wait_for_commit::unregister_wait_for_prior_commit(). This could cause the error condition to wrongly propagate to a later wait_for_prior_commit(), causing spurious ER_PRIOR_COMMIT_FAILED errors. - Do not put the binlog background thread into the processlist. It causes too many result differences in mtr, but also it probably is not useful for users to pollute the process list with a system thread that does not really perform any user-visible tasks...
2014-06-10 10:13:15 +02:00
include/wait_for_slave_sql_error.inc [errno=1317,1927,1964]
STOP SLAVE IO_THREAD;
SELECT * FROM t3 WHERE a >= 30 ORDER BY a;
a b
31 31
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
RETURN x;
END
||
SET sql_log_bin=1;
INSERT INTO t3 VALUES (39,0);
include/start_slave.inc
SELECT * FROM t3 WHERE a >= 30 ORDER BY a;
a b
31 31
32 32
33 33
34 34
39 0
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
IF d1 != '' THEN
SET debug_sync = d1;
END IF;
IF d2 != '' THEN
SET debug_sync = d2;
END IF;
RETURN x;
END
||
SET sql_log_bin=1;
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
*** 2. Same as (1), but without restarting IO thread after kill of SQL threads ***
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
SET binlog_format=statement;
INSERT INTO t3 VALUES (41, foo(41,
'commit_before_prepare_ordered WAIT_FOR t2_waiting',
'commit_after_prepare_ordered SIGNAL t1_ready WAIT_FOR t1_cont'));
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
SET binlog_format=statement;
BEGIN;
INSERT INTO t3 VALUES (42, foo(42,
'ha_write_row_end SIGNAL t2_query WAIT_FOR t2_cont',
''));
INSERT INTO t3 VALUES (43, foo(43,
'group_commit_waiting_for_prior SIGNAL t2_waiting',
'group_commit_waiting_for_prior_killed SIGNAL t2_killed'));
COMMIT;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued3';
SET binlog_format=statement;
INSERT INTO t3 VALUES (44, foo(44,
'',
''));
SET debug_sync='now WAIT_FOR master_queued3';
SET debug_sync='now SIGNAL master_cont1';
SELECT * FROM t3 WHERE a >= 40 ORDER BY a;
a b
41 41
42 42
43 43
44 44
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET debug_sync='now WAIT_FOR t2_query';
SET debug_sync='now SIGNAL t2_cont';
SET debug_sync='now WAIT_FOR t1_ready';
KILL THD_ID;
SET debug_sync='now WAIT_FOR t2_killed';
SET debug_sync='now SIGNAL t1_cont';
MDEV-5262, MDEV-5914, MDEV-5941, MDEV-6020: Deadlocks during parallel replication causing replication to fail. Remove the temporary fix for MDEV-5914, which used READ COMMITTED for parallel replication worker threads. Replace it with a better, more selective solution. The issue is with certain edge cases of InnoDB gap locks, for example between INSERT and ranged DELETE. It is possible for the gap lock set by the DELETE to block the INSERT, if the DELETE runs first, while the record lock set by INSERT does not block the DELETE, if the INSERT runs first. This can cause a conflict between the two in parallel replication on the slave even though they ran without conflicts on the master. With this patch, InnoDB will ask the server layer about the two involved transactions before blocking on a gap lock. If the server layer tells InnoDB that the transactions are already fixed wrt. commit order, as they are in parallel replication, InnoDB will ignore the gap lock and allow the two transactions to proceed in parallel, avoiding the conflict. Improve the fix for MDEV-6020. When InnoDB itself detects a deadlock, it now asks the server layer for any preferences about which transaction to roll back. In case of parallel replication with two transactions T1 and T2 fixed to commit T1 before T2, the server layer will ask InnoDB to roll back T2 as the deadlock victim, not T1. This helps in some cases to avoid excessive deadlock rollback, as T2 will in any case need to wait for T1 to complete before it can itself commit. Also some misc. fixes found during development and testing: - Remove thd_rpl_is_parallel(), it is not used or needed. - Use KILL_CONNECTION instead of KILL_QUERY when a parallel replication worker thread is killed to resolve a deadlock with fixed commit ordering. There are some cases, eg. in sql/sql_parse.cc, where a KILL_QUERY can be ignored if the query otherwise completed successfully, and this could cause the deadlock kill to be lost, so that the deadlock was not correctly resolved. - Fix random test failure due to missing wait_for_binlog_checkpoint.inc. - Make sure that deadlock or other temporary errors during parallel replication are not printed to the the error log; there were some places around the replication code with extra error logging. These conditions can occur occasionally and are handled automatically without breaking replication, so they should not pollute the error log. - Fix handling of rgi->gtid_sub_id. We need to be able to access this also at the end of a transaction, to be able to detect and resolve deadlocks due to commit ordering. But this value was also used as a flag to mark whether record_gtid() had been called, by being set to zero, losing the value. Now, introduce a separate flag rgi->gtid_pending, so rgi->gtid_sub_id remains valid for the entire duration of the transaction. - Fix one place where the code to handle ignored errors called reset_killed() unconditionally, even if no error was caught that should be ignored. This could cause loss of a deadlock kill signal, breaking deadlock detection and resolution. - Fix a couple of missing mysql_reset_thd_for_next_command(). This could cause a prior error condition to remain for the next event executed, causing assertions about errors already being set and possibly giving incorrect error handling for following event executions. - Fix code that cleared thd->rgi_slave in the parallel replication worker threads after each event execution; this caused the deadlock detection and handling code to not be able to correctly process the associated transactions as belonging to replication worker threads. - Remove useless error code in slave_background_kill_request(). - Fix bug where wfc->wakeup_error was not cleared at wait_for_commit::unregister_wait_for_prior_commit(). This could cause the error condition to wrongly propagate to a later wait_for_prior_commit(), causing spurious ER_PRIOR_COMMIT_FAILED errors. - Do not put the binlog background thread into the processlist. It causes too many result differences in mtr, but also it probably is not useful for users to pollute the process list with a system thread that does not really perform any user-visible tasks...
2014-06-10 10:13:15 +02:00
include/wait_for_slave_sql_error.inc [errno=1317,1927,1964]
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
RETURN x;
END
||
SET sql_log_bin=1;
INSERT INTO t3 VALUES (49,0);
START SLAVE SQL_THREAD;
SELECT * FROM t3 WHERE a >= 40 ORDER BY a;
a b
41 41
42 42
43 43
44 44
49 0
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
IF d1 != '' THEN
SET debug_sync = d1;
END IF;
IF d2 != '' THEN
SET debug_sync = d2;
END IF;
RETURN x;
END
||
SET sql_log_bin=1;
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
*** 3. Same as (2), but not using gtid mode ***
include/stop_slave.inc
CHANGE MASTER TO master_use_gtid=no;
include/start_slave.inc
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
SET binlog_format=statement;
INSERT INTO t3 VALUES (51, foo(51,
'commit_before_prepare_ordered WAIT_FOR t2_waiting',
'commit_after_prepare_ordered SIGNAL t1_ready WAIT_FOR t1_cont'));
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
SET binlog_format=statement;
BEGIN;
INSERT INTO t3 VALUES (52, foo(52,
'ha_write_row_end SIGNAL t2_query WAIT_FOR t2_cont',
''));
INSERT INTO t3 VALUES (53, foo(53,
'group_commit_waiting_for_prior SIGNAL t2_waiting',
'group_commit_waiting_for_prior_killed SIGNAL t2_killed'));
COMMIT;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued3';
SET binlog_format=statement;
INSERT INTO t3 VALUES (54, foo(54,
'',
''));
SET debug_sync='now WAIT_FOR master_queued3';
SET debug_sync='now SIGNAL master_cont1';
SELECT * FROM t3 WHERE a >= 50 ORDER BY a;
a b
51 51
52 52
53 53
54 54
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET debug_sync='now WAIT_FOR t2_query';
SET debug_sync='now SIGNAL t2_cont';
SET debug_sync='now WAIT_FOR t1_ready';
KILL THD_ID;
SET debug_sync='now WAIT_FOR t2_killed';
SET debug_sync='now SIGNAL t1_cont';
MDEV-5262, MDEV-5914, MDEV-5941, MDEV-6020: Deadlocks during parallel replication causing replication to fail. Remove the temporary fix for MDEV-5914, which used READ COMMITTED for parallel replication worker threads. Replace it with a better, more selective solution. The issue is with certain edge cases of InnoDB gap locks, for example between INSERT and ranged DELETE. It is possible for the gap lock set by the DELETE to block the INSERT, if the DELETE runs first, while the record lock set by INSERT does not block the DELETE, if the INSERT runs first. This can cause a conflict between the two in parallel replication on the slave even though they ran without conflicts on the master. With this patch, InnoDB will ask the server layer about the two involved transactions before blocking on a gap lock. If the server layer tells InnoDB that the transactions are already fixed wrt. commit order, as they are in parallel replication, InnoDB will ignore the gap lock and allow the two transactions to proceed in parallel, avoiding the conflict. Improve the fix for MDEV-6020. When InnoDB itself detects a deadlock, it now asks the server layer for any preferences about which transaction to roll back. In case of parallel replication with two transactions T1 and T2 fixed to commit T1 before T2, the server layer will ask InnoDB to roll back T2 as the deadlock victim, not T1. This helps in some cases to avoid excessive deadlock rollback, as T2 will in any case need to wait for T1 to complete before it can itself commit. Also some misc. fixes found during development and testing: - Remove thd_rpl_is_parallel(), it is not used or needed. - Use KILL_CONNECTION instead of KILL_QUERY when a parallel replication worker thread is killed to resolve a deadlock with fixed commit ordering. There are some cases, eg. in sql/sql_parse.cc, where a KILL_QUERY can be ignored if the query otherwise completed successfully, and this could cause the deadlock kill to be lost, so that the deadlock was not correctly resolved. - Fix random test failure due to missing wait_for_binlog_checkpoint.inc. - Make sure that deadlock or other temporary errors during parallel replication are not printed to the the error log; there were some places around the replication code with extra error logging. These conditions can occur occasionally and are handled automatically without breaking replication, so they should not pollute the error log. - Fix handling of rgi->gtid_sub_id. We need to be able to access this also at the end of a transaction, to be able to detect and resolve deadlocks due to commit ordering. But this value was also used as a flag to mark whether record_gtid() had been called, by being set to zero, losing the value. Now, introduce a separate flag rgi->gtid_pending, so rgi->gtid_sub_id remains valid for the entire duration of the transaction. - Fix one place where the code to handle ignored errors called reset_killed() unconditionally, even if no error was caught that should be ignored. This could cause loss of a deadlock kill signal, breaking deadlock detection and resolution. - Fix a couple of missing mysql_reset_thd_for_next_command(). This could cause a prior error condition to remain for the next event executed, causing assertions about errors already being set and possibly giving incorrect error handling for following event executions. - Fix code that cleared thd->rgi_slave in the parallel replication worker threads after each event execution; this caused the deadlock detection and handling code to not be able to correctly process the associated transactions as belonging to replication worker threads. - Remove useless error code in slave_background_kill_request(). - Fix bug where wfc->wakeup_error was not cleared at wait_for_commit::unregister_wait_for_prior_commit(). This could cause the error condition to wrongly propagate to a later wait_for_prior_commit(), causing spurious ER_PRIOR_COMMIT_FAILED errors. - Do not put the binlog background thread into the processlist. It causes too many result differences in mtr, but also it probably is not useful for users to pollute the process list with a system thread that does not really perform any user-visible tasks...
2014-06-10 10:13:15 +02:00
include/wait_for_slave_sql_error.inc [errno=1317,1927,1964]
SELECT * FROM t3 WHERE a >= 50 ORDER BY a;
a b
51 51
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
RETURN x;
END
||
SET sql_log_bin=1;
INSERT INTO t3 VALUES (59,0);
START SLAVE SQL_THREAD;
SELECT * FROM t3 WHERE a >= 50 ORDER BY a;
a b
51 51
52 52
53 53
54 54
59 0
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
IF d1 != '' THEN
SET debug_sync = d1;
END IF;
IF d2 != '' THEN
SET debug_sync = d2;
END IF;
RETURN x;
END
||
SET sql_log_bin=1;
include/stop_slave.inc
CHANGE MASTER TO master_use_gtid=slave_pos;
include/start_slave.inc
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET GLOBAL slave_parallel_threads=4;
include/start_slave.inc
*** 4. Test killing thread that is waiting to start transaction until previous transaction commits ***
SET binlog_format=statement;
SET gtid_domain_id=2;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
BEGIN;
INSERT INTO t3 VALUES (70, foo(70,
'rpl_parallel_start_waiting_for_prior SIGNAL t4_waiting', ''));
INSERT INTO t3 VALUES (60, foo(60,
'ha_write_row_end SIGNAL d2_query WAIT_FOR d2_cont2',
'rpl_parallel_end_of_group SIGNAL d2_done WAIT_FOR d2_cont'));
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
COMMIT;
SET gtid_domain_id=0;
SET debug_sync='now WAIT_FOR d2_query';
SET gtid_domain_id=1;
BEGIN;
INSERT INTO t3 VALUES (61, foo(61,
'rpl_parallel_start_waiting_for_prior SIGNAL t3_waiting',
'rpl_parallel_start_waiting_for_prior_killed SIGNAL t3_killed'));
INSERT INTO t3 VALUES (62, foo(62,
'ha_write_row_end SIGNAL d1_query WAIT_FOR d1_cont2',
'rpl_parallel_end_of_group SIGNAL d1_done WAIT_FOR d1_cont'));
COMMIT;
SET gtid_domain_id=0;
SET debug_sync='now WAIT_FOR d1_query';
SET gtid_domain_id=0;
INSERT INTO t3 VALUES (63, foo(63,
'ha_write_row_end SIGNAL d0_query WAIT_FOR d0_cont2',
'rpl_parallel_end_of_group SIGNAL d0_done WAIT_FOR d0_cont'));
SET debug_sync='now WAIT_FOR d0_query';
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET gtid_domain_id=3;
BEGIN;
INSERT INTO t3 VALUES (68, foo(68,
'rpl_parallel_start_waiting_for_prior SIGNAL t2_waiting', ''));
INSERT INTO t3 VALUES (69, foo(69,
'ha_write_row_end SIGNAL d3_query WAIT_FOR d3_cont2',
'rpl_parallel_end_of_group SIGNAL d3_done WAIT_FOR d3_cont'));
COMMIT;
SET gtid_domain_id=0;
SET debug_sync='now WAIT_FOR d3_query';
SET debug_sync='now SIGNAL d2_cont2';
SET debug_sync='now WAIT_FOR d2_done';
SET debug_sync='now SIGNAL d1_cont2';
SET debug_sync='now WAIT_FOR d1_done';
SET debug_sync='now SIGNAL d0_cont2';
SET debug_sync='now WAIT_FOR d0_done';
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='now SIGNAL d3_cont2';
SET debug_sync='now WAIT_FOR d3_done';
SET binlog_format=statement;
INSERT INTO t3 VALUES (64, foo(64,
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
'rpl_parallel_before_mark_start_commit SIGNAL t1_waiting WAIT_FOR t1_cont', ''));
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2 WAIT_FOR master_cont2';
INSERT INTO t3 VALUES (65, foo(65, '', ''));
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued3';
INSERT INTO t3 VALUES (66, foo(66, '', ''));
SET debug_sync='now WAIT_FOR master_queued3';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued4';
INSERT INTO t3 VALUES (67, foo(67, '', ''));
SET debug_sync='now WAIT_FOR master_queued4';
SET debug_sync='now SIGNAL master_cont2';
SELECT * FROM t3 WHERE a >= 60 ORDER BY a;
a b
60 60
61 61
62 62
63 63
64 64
65 65
66 66
67 67
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
68 68
69 69
70 70
SET debug_sync='RESET';
SET debug_sync='now SIGNAL d0_cont';
SET debug_sync='now WAIT_FOR t1_waiting';
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='now SIGNAL d3_cont';
SET debug_sync='now WAIT_FOR t2_waiting';
SET debug_sync='now SIGNAL d1_cont';
SET debug_sync='now WAIT_FOR t3_waiting';
SET debug_sync='now SIGNAL d2_cont';
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='now WAIT_FOR t4_waiting';
KILL THD_ID;
SET debug_sync='now WAIT_FOR t3_killed';
SET debug_sync='now SIGNAL t1_cont';
2014-02-10 15:12:17 +01:00
include/wait_for_slave_sql_error.inc [errno=1317,1927,1964]
STOP SLAVE IO_THREAD;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SELECT * FROM t3 WHERE a >= 60 AND a != 65 ORDER BY a;
a b
60 60
61 61
62 62
63 63
64 64
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
68 68
69 69
70 70
SET debug_sync='RESET';
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
RETURN x;
END
||
SET sql_log_bin=1;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
UPDATE t3 SET b=b+1 WHERE a=60;
include/start_slave.inc
SELECT * FROM t3 WHERE a >= 60 ORDER BY a;
a b
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
60 61
61 61
62 62
63 63
64 64
65 65
66 66
67 67
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
68 68
69 69
70 70
SET sql_log_bin=0;
DROP FUNCTION foo;
CREATE FUNCTION foo(x INT, d1 VARCHAR(500), d2 VARCHAR(500))
RETURNS INT DETERMINISTIC
BEGIN
IF d1 != '' THEN
SET debug_sync = d1;
END IF;
IF d2 != '' THEN
SET debug_sync = d2;
END IF;
RETURN x;
END
||
SET sql_log_bin=1;
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
*** 5. Test killing thread that is waiting for queue of max length to shorten ***
SET @old_max_queued= @@GLOBAL.slave_parallel_max_queued;
SET GLOBAL slave_parallel_max_queued=9000;
SET binlog_format=statement;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
INSERT INTO t3 VALUES (80, foo(0,
'ha_write_row_end SIGNAL query_waiting WAIT_FOR query_cont', ''));
SET debug_sync='now WAIT_FOR query_waiting';
SET @old_dbug= @@GLOBAL.debug_dbug;
SET GLOBAL debug_dbug="+d,rpl_parallel_wait_queue_max";
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SELECT * FROM t3 WHERE a >= 80 ORDER BY a;
a b
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
80 0
81 10000
SET debug_sync='now WAIT_FOR wait_queue_ready';
KILL THD_ID;
SET debug_sync='now WAIT_FOR wait_queue_killed';
SET debug_sync='now SIGNAL query_cont';
2014-02-10 15:12:17 +01:00
include/wait_for_slave_sql_error.inc [errno=1317,1927,1964]
STOP SLAVE IO_THREAD;
SET GLOBAL debug_dbug=@old_dbug;
SET GLOBAL slave_parallel_max_queued= @old_max_queued;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
INSERT INTO t3 VALUES (82,0);
SET binlog_format=@old_format;
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SET debug_sync='RESET';
include/start_slave.inc
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
SELECT * FROM t3 WHERE a >= 80 ORDER BY a;
a b
MDEV-5657: Parallel replication. Clean up and improve the parallel implementation code, mainly related to scheduling of work to threads and handling of stop and errors. Fix a lot of bugs in various corner cases that could lead to crashes or corruption. Fix that a single replication domain could easily grab all worker threads and stall all other domains; now a configuration variable --slave-domain-parallel-threads allows to limit the number of workers. Allow next event group to start as soon as previous group begins the commit phase (as opposed to when it ends it); this allows multiple event groups on the slave to participate in group commit, even when no other opportunities for parallelism are available. Various fixes: - Fix some races in the rpl.rpl_parallel test case. - Fix an old incorrect assertion in Log_event iocache read. - Fix repeated malloc/free of wait_for_commit and rpl_group_info objects. - Simplify wait_for_commit wakeup logic. - Fix one case in queue_for_group_commit() where killing one thread would fail to correctly signal the error to the next, causing loss of the transaction after slave restart. - Fix leaking of pthreads (and their allocated stack) due to missing PTHREAD_CREATE_DETACHED attribute. - Fix how one batch of group-committed transactions wait for the previous batch before starting to execute themselves. The old code had a very complex scheduling where the first transaction was handled differently, with subtle bugs in corner cases. Now each event group is always scheduled for a new worker (in a round-robin fashion amongst available workers). Keep a count of how many transactions have started to commit, and wait for that counter to reach the appropriate value. - Fix slave stop to wait for all workers to actually complete processing; before, the wait was for update of last_committed_sub_id, which happens a bit earlier, and could leave worker threads potentially accessing bits of the replication state that is no longer valid after slave stop. - Fix a couple of places where the test suite would kill a thread waiting inside enter_cond() in connection with debug_sync; debug_sync + kill can crash in rare cases due to a race with mysys_var_current_mutex in this case. - Fix some corner cases where we had enter_cond() but no exit_cond(). - Fix that we could get failure in wait_for_prior_commit() but forget to flag the error with my_error(). - Fix slave stop (both for normal stop and stop due to error). Now, at stop we pick a specific safe point (in terms of event groups executed) and make sure that all event groups before that point are executed to completion, and that no event group after start executing; this ensures a safe place to restart replication, even for non-transactional stuff/DDL. In error stop, make sure that all prior event groups are allowed to execute to completion, and that any later event groups that have started are rolled back, if possible. The old code could leave eg. T1 and T3 committed but T2 not, or it could even leave half a transaction not rolled back in some random worker, which would cause big problems when that worker was later reused after slave restart. - Fix the accounting of amount of events queued for one worker. Before, the amount was reduced immediately as soon as the events were dequeued (which happens all at once); this allowed twice the amount of events to be queued in memory for each single worker, which is not what users would expect. - Fix that an error set during execution of one event was sometimes not cleared before executing the next, causing problems with the error reporting. - Fix incorrect handling of thd->killed in worker threads.
2014-02-26 15:02:09 +01:00
80 0
81 10000
82 0
include/stop_slave.inc
SET GLOBAL binlog_format=@old_format;
SET GLOBAL slave_parallel_threads=0;
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
*** MDEV-5788 Incorrect free of rgi->deferred_events in parallel replication ***
include/stop_slave.inc
SET GLOBAL replicate_ignore_table="test.t3";
SET GLOBAL slave_parallel_threads=2;
include/start_slave.inc
INSERT INTO t3 VALUES (100, rand());
INSERT INTO t3 VALUES (101, rand());
INSERT INTO t3 VALUES (102, rand());
INSERT INTO t3 VALUES (103, rand());
INSERT INTO t3 VALUES (104, rand());
INSERT INTO t3 VALUES (105, rand());
include/stop_slave.inc
SET GLOBAL replicate_ignore_table="";
include/start_slave.inc
INSERT INTO t3 VALUES (106, rand());
INSERT INTO t3 VALUES (107, rand());
SELECT * FROM t3 WHERE a >= 100 ORDER BY a;
a b
106 #
107 #
*** MDEV-5921: In parallel replication, an error is not correctly signalled to the next transaction ***
include/stop_slave.inc
SET GLOBAL slave_parallel_threads=10;
include/start_slave.inc
INSERT INTO t3 VALUES (110, 1);
SELECT * FROM t3 WHERE a >= 110 ORDER BY a;
a b
110 1
SET sql_log_bin=0;
INSERT INTO t3 VALUES (111, 666);
SET sql_log_bin=1;
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
INSERT INTO t3 VALUES (111, 2);
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
INSERT INTO t3 VALUES (112, 3);
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='now SIGNAL master_cont1';
SET debug_sync='RESET';
include/wait_for_slave_sql_error.inc [errno=1062]
include/wait_for_slave_sql_to_stop.inc
SELECT * FROM t3 WHERE a >= 110 ORDER BY a;
a b
110 1
111 666
SET sql_log_bin=0;
DELETE FROM t3 WHERE a=111 AND b=666;
SET sql_log_bin=1;
START SLAVE SQL_THREAD;
SELECT * FROM t3 WHERE a >= 110 ORDER BY a;
a b
110 1
111 2
112 3
***MDEV-5914: Parallel replication deadlock due to InnoDB lock conflicts ***
include/stop_slave.inc
CREATE TABLE t4 (a INT PRIMARY KEY, b INT, KEY b_idx(b)) ENGINE=InnoDB;
INSERT INTO t4 VALUES (1,NULL), (2,2), (3,NULL), (4,4), (5, NULL), (6, 6);
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
UPDATE t4 SET b=NULL WHERE a=6;
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
DELETE FROM t4 WHERE b <= 3;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='now SIGNAL master_cont1';
SET debug_sync='RESET';
include/start_slave.inc
include/stop_slave.inc
SELECT * FROM t4 ORDER BY a;
a b
1 NULL
3 NULL
4 4
5 NULL
6 NULL
DELETE FROM t4;
INSERT INTO t4 VALUES (1,NULL), (2,2), (3,NULL), (4,4), (5, NULL), (6, 6);
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
INSERT INTO t4 VALUES (7, NULL);
SET debug_sync='now WAIT_FOR master_queued1';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
DELETE FROM t4 WHERE b <= 3;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='now SIGNAL master_cont1';
SET debug_sync='RESET';
include/start_slave.inc
include/stop_slave.inc
SELECT * FROM t4 ORDER BY a;
a b
1 NULL
3 NULL
4 4
5 NULL
6 6
7 NULL
DELETE FROM t4;
INSERT INTO t4 VALUES (1,NULL), (2,2), (3,NULL), (4,4), (5, NULL), (6, 6);
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued1 WAIT_FOR master_cont1';
UPDATE t4 SET b=NULL WHERE a=6;
SET debug_sync='now WAIT_FOR master_queued1';
SET @old_format= @@SESSION.binlog_format;
SET binlog_format='statement';
SET debug_sync='commit_after_release_LOCK_prepare_ordered SIGNAL master_queued2';
DELETE FROM t4 WHERE b <= 1;
SET debug_sync='now WAIT_FOR master_queued2';
SET debug_sync='now SIGNAL master_cont1';
SET @old_format=@@GLOBAL.binlog_format;
SET debug_sync='RESET';
SET @old_dbug= @@GLOBAL.debug_dbug;
SET GLOBAL debug_dbug="+d,disable_thd_need_ordering_with";
include/start_slave.inc
SET GLOBAL debug_dbug=@old_dbug;
SELECT * FROM t4 ORDER BY a;
a b
1 NULL
2 2
3 NULL
4 4
5 NULL
6 NULL
SET @last_gtid= 'GTID';
SELECT IF(@@gtid_slave_pos LIKE CONCAT('%',@last_gtid,'%'), "GTID found ok",
CONCAT("GTID ", @last_gtid, " not found in gtid_slave_pos=", @@gtid_slave_pos))
AS result;
result
GTID found ok
SELECT "ROW FOUND" AS `Is the row found?`
FROM mysql.gtid_slave_pos
WHERE CONCAT(domain_id, "-", server_id, "-", seq_no) = @last_gtid;
Is the row found?
ROW FOUND
*** MDEV-5938: Exec_master_log_pos not updated at log rotate in parallel replication ***
include/stop_slave.inc
SET GLOBAL slave_parallel_threads=1;
SET DEBUG_SYNC= 'RESET';
include/start_slave.inc
CREATE TABLE t5 (a INT PRIMARY KEY, b INT);
INSERT INTO t5 VALUES (1,1);
INSERT INTO t5 VALUES (2,2), (3,8);
INSERT INTO t5 VALUES (4,16);
test_check
OK
test_check
OK
FLUSH LOGS;
test_check
OK
test_check
OK
*** MDEV_6435: Incorrect error handling when query binlogged partially on master with "killed" error ***
CREATE TABLE t6 (a INT) ENGINE=MyISAM;
CREATE TRIGGER tr AFTER INSERT ON t6 FOR EACH ROW SET @a = 1;
SET @old_format= @@binlog_format;
SET binlog_format= statement;
SET debug_sync='sp_head_execute_before_loop SIGNAL ready WAIT_FOR cont';
INSERT INTO t6 VALUES (1), (2), (3);
SET debug_sync='now WAIT_FOR ready';
KILL QUERY CONID;
SET debug_sync='now SIGNAL cont';
ERROR 70100: Query execution was interrupted
SET binlog_format= @old_format;
SET debug_sync='RESET';
SET debug_sync='RESET';
include/wait_for_slave_sql_error.inc [errno=1317]
STOP SLAVE IO_THREAD;
SET GLOBAL gtid_slave_pos= 'AFTER_ERROR_GTID_POS';
include/start_slave.inc
INSERT INTO t6 VALUES (4);
SELECT * FROM t6 ORDER BY a;
a
1
4
SELECT * FROM t6 ORDER BY a;
a
4
include/stop_slave.inc
SET GLOBAL slave_parallel_threads=@old_parallel_threads;
include/start_slave.inc
2014-02-05 20:53:54 +02:00
SET DEBUG_SYNC= 'RESET';
DROP function foo;
DROP TABLE t1,t2,t3,t4,t5,t6;
2014-02-05 20:53:54 +02:00
SET DEBUG_SYNC= 'RESET';
include/rpl_end.inc