Hi Andrei! Here is the patch to move the cleanup of old GTID rows into the slave background thread. It is mostly identical to the patch I wrote a month ago, but now I got test cases added/updated and I ran it through buildbot. Want to review it? The patch should be complete now. The patch is also available in git here: https://github.com/knielsen/server/tree/gtid_delete_in_background_thread This patch could go into 10.4. With the patch, a request is sent to the background thread to delete old rows every @@gtid_cleanup_batch_size replicated transactions. Some considerations on the default of this value: - It should be reasonably large so that the background thread only needs to wake up occasionally, and a number of row deletions can share the cost of the part that is not inherently per-row. - A large enough value can group a number of row deletions in a single transaction which is probably beneficial. There is also a fixed overhead per existing gtid_domain_id to scan the GTID state, so the value should not be much smaller than the number of domains in use. - It should not be too large either, since that could cause the mysql.gtid_slave_pos table to grow unnecessarily. Ideally the table should fit in one InnoDB data page. I settled on a default of 64. Probably a wide range will work fine in >99% of cases, eg. 10 or 100 should be mostly the same. 64 should be large enough that the per-batch overhead becomes small relative to the per-row, even if hundreds of domains are used. And small enough that the table is likely to be able to fit in a single InnoDB 16k page, even considering that deletion is asynchroneous and more rows can accumulate temporarily due to parallel replication etc. I think this is a nice improvement. Moving the deletion logic away from the main replication SQL/worker threads simplifies the logic, eg. for when a transaction fails after a deletion has already been done. And it also has the potential to be more efficient; the contention on global mutex LOCK_slave_state is reduced, processing is moved out of the main replication threads, and batching row deletions could potentially be cheaper than doing them one-by-one. - Kristian. Kristian Nielsen writes:

revision-id: 8bfb140d5dc247c183787b8a0a1799cf375845bd (mariadb-10.3.10-25-g8bfb140d5dc) parent(s): 74387028a06c557f36a0fd1bbde347f1551c8fb7 author: Kristian Nielsen committer: Kristian Nielsen timestamp: 2018-11-25 19:38:33 +0100 message:

Move deletion of old GTID rows to slave background thread

This patch changes how old rows in mysql.gtid_slave_pos* tables are deleted. Instead of doing it as part of every replicated transaction in record_gtid(), it is done periodically (every @@gtid_cleanup_batch_size transaction) in the slave background thread.

This removes the deletion step from the replication process in SQL or worker threads, which could speed up replication with many small transactions. It also decreases contention on the global mutex LOCK_slave_state. And it simplifies the logic, eg. when a replicated transaction fails after having deleted old rows.

With this patch, the deletion of old GTID rows happens asynchroneously and slightly non-deterministic. Thus the number of old rows in mysql.gtid_slave_pos can temporarily exceed @@gtid_cleanup_batch_size. But all old rows will be deleted eventually after sufficiently many new GTIDs have been replicated.

--- mysql-test/main/mysqld--help.result | 10 + mysql-test/suite/rpl/r/rpl_gtid_mdev4484.result | 40 +- mysql-test/suite/rpl/r/rpl_gtid_stop_start.result | 8 +- .../suite/rpl/r/rpl_parallel_optimistic.result | 14 +- mysql-test/suite/rpl/t/rpl_gtid_mdev4484.test | 68 +++- .../suite/rpl/t/rpl_parallel_optimistic.test | 42 ++- .../sys_vars/r/sysvars_server_notembedded.result | 14 + sql/log_event.cc | 6 +- sql/mysqld.cc | 1 + sql/mysqld.h | 1 + sql/rpl_gtid.cc | 413 +++++++++++++-------- sql/rpl_gtid.h | 12 +- sql/rpl_rli.cc | 87 +---- sql/rpl_rli.h | 11 - sql/slave.cc | 35 +- sql/slave.h | 1 + sql/sys_vars.cc | 13 + .../mysql-test/rocksdb_rpl/r/mdev12179.result | 18 + .../mysql-test/rocksdb_rpl/t/mdev12179.test | 85 +++++ .../mysql-test/tokudb_rpl/r/mdev12179.result | 18 + .../tokudb/mysql-test/tokudb_rpl/t/mdev12179.test | 85 +++++ 21 files changed, 675 insertions(+), 307 deletions(-)

diff --git a/mysql-test/main/mysqld--help.result b/mysql-test/main/mysqld--help.result index 5a7153f32d3..4f801ec5275 100644 --- a/mysql-test/main/mysqld--help.result +++ b/mysql-test/main/mysqld--help.result @@ -294,6 +294,15 @@ The following specify which files/extra groups are read (specified before remain --group-concat-max-len=# The maximum length of the result of function GROUP_CONCAT() + --gtid-cleanup-batch-size=# + Normally does not need tuning. How many old rows must + accumulate in the mysql.gtid_slave_pos table before a + background job will be run to delete them. Can be + increased to reduce number of commits if using many + different engines with --gtid_pos_auto_engines, or to + reduce CPU overhead if using a huge number of different + gtid_domain_ids. Can be decreased to reduce number of old + rows in the table. --gtid-domain-id=# Used with global transaction ID to identify logically independent replication streams. When events can propagate through multiple parallel paths (for example @@ -1425,6 +1434,7 @@ gdb FALSE general-log FALSE getopt-prefix-matching FALSE group-concat-max-len 1048576 +gtid-cleanup-batch-size 64 gtid-domain-id 0 gtid-ignore-duplicates FALSE gtid-pos-auto-engines diff --git a/mysql-test/suite/rpl/r/rpl_gtid_mdev4484.result b/mysql-test/suite/rpl/r/rpl_gtid_mdev4484.result index aaeb0c8f119..55d2831dcf4 100644 --- a/mysql-test/suite/rpl/r/rpl_gtid_mdev4484.result +++ b/mysql-test/suite/rpl/r/rpl_gtid_mdev4484.result @@ -16,36 +16,32 @@ INSERT INTO t1 VALUES (1); connection slave; connection slave; include/stop_slave.inc +SET @old_gtid_cleanup_batch_size= @@GLOBAL.gtid_cleanup_batch_size; +SET GLOBAL gtid_cleanup_batch_size= 2; SET @old_dbug= @@GLOBAL.debug_dbug; SET GLOBAL debug_dbug="+d,gtid_slave_pos_simulate_failed_delete"; SET sql_log_bin= 0; -CALL mtr.add_suppression("Can't find file"); +CALL mtr.add_suppression("<DEBUG> Error deleting old GTID row"); SET sql_log_bin= 1; include/start_slave.inc connection master; -INSERT INTO t1 VALUES (2); -connection slave; -include/wait_for_slave_sql_error.inc [errno=1942] -STOP SLAVE IO_THREAD; -SELECT domain_id, server_id, seq_no FROM mysql.gtid_slave_pos -ORDER BY domain_id, sub_id DESC LIMIT 1; -domain_id server_id seq_no -0 1 3 +connection slave; +SELECT COUNT(*), MAX(seq_no) INTO @pre_count, @pre_max_seq_no +FROM mysql.gtid_slave_pos; +SELECT IF(@pre_count >= 20, "OK", CONCAT("Error: too few rows seen while errors injected: ", @pre_count)); +IF(@pre_count >= 20, "OK", CONCAT("Error: too few rows seen while errors injected: ", @pre_count)) +OK SET GLOBAL debug_dbug= @old_dbug; -include/start_slave.inc connection master; -INSERT INTO t1 VALUES (3); -connection slave; -connection slave; -SELECT domain_id, server_id, seq_no FROM mysql.gtid_slave_pos -ORDER BY domain_id, sub_id DESC LIMIT 1; -domain_id server_id seq_no -0 1 4 -SELECT * FROM t1 ORDER BY i; -i -1 -2 -3 +connection slave; +connection slave; +SELECT IF(COUNT(*) >= 1, "OK", CONCAT("Error: too few rows seen after errors no longer injected: ", COUNT(*))) +FROM mysql.gtid_slave_pos +WHERE seq_no <= @pre_max_seq_no; +IF(COUNT(*) >= 1, "OK", CONCAT("Error: too few rows seen after errors no longer injected: ", COUNT(*))) +OK connection master; DROP TABLE t1; +connection slave; +SET GLOBAL gtid_cleanup_batch_size= @old_gtid_cleanup_batch_size; include/rpl_end.inc diff --git a/mysql-test/suite/rpl/r/rpl_gtid_stop_start.result b/mysql-test/suite/rpl/r/rpl_gtid_stop_start.result index ff845794c22..b27ffed9f94 100644 --- a/mysql-test/suite/rpl/r/rpl_gtid_stop_start.result +++ b/mysql-test/suite/rpl/r/rpl_gtid_stop_start.result @@ -171,7 +171,7 @@ include/start_slave.inc *** MDEV-4692: mysql.gtid_slave_pos accumulates values for a domain *** SELECT domain_id, COUNT(*) FROM mysql.gtid_slave_pos GROUP BY domain_id; domain_id COUNT(*) -0 2 +0 3 1 2 connection server_1; INSERT INTO t1 VALUES (11); @@ -179,7 +179,7 @@ connection server_2; FLUSH NO_WRITE_TO_BINLOG TABLES; SELECT domain_id, COUNT(*) FROM mysql.gtid_slave_pos GROUP BY domain_id; domain_id COUNT(*) -0 2 +0 4 1 2 include/start_slave.inc connection server_1; @@ -189,8 +189,8 @@ connection server_2; FLUSH NO_WRITE_TO_BINLOG TABLES; SELECT domain_id, COUNT(*) FROM mysql.gtid_slave_pos GROUP BY domain_id; domain_id COUNT(*) -0 2 -1 2 +0 3 +1 1 *** MDEV-4650: show variables; ERROR 1946 (HY000): Failed to load replication slave GTID position *** connection server_2; SET sql_log_bin=0; diff --git a/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result b/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result index ca202a66b0e..83343e52cab 100644 --- a/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result +++ b/mysql-test/suite/rpl/r/rpl_parallel_optimistic.result @@ -12,6 +12,8 @@ SET GLOBAL slave_parallel_threads=10; CHANGE MASTER TO master_use_gtid=slave_pos; SET @old_parallel_mode=@@GLOBAL.slave_parallel_mode; SET GLOBAL slave_parallel_mode='optimistic'; +SET @old_gtid_cleanup_batch_size= @@GLOBAL.gtid_cleanup_batch_size; +SET GLOBAL gtid_cleanup_batch_size= 1000000; connection server_1; INSERT INTO t1 VALUES(1,1); BEGIN; @@ -131,6 +133,11 @@ c 204 205 206 +SELECT IF(COUNT(*) >= 30, "OK", CONCAT("Error: too few old rows found: ", COUNT(*))) +FROM mysql.gtid_slave_pos; +IF(COUNT(*) >= 30, "OK", CONCAT("Error: too few old rows found: ", COUNT(*))) +OK +SET GLOBAL gtid_cleanup_batch_size=1; *** Test @@skip_parallel_replication. *** connection server_2; include/stop_slave.inc @@ -651,9 +658,10 @@ DROP TABLE t1, t2, t3; include/save_master_gtid.inc connection server_2; include/sync_with_master_gtid.inc -Check that no more than the expected last four GTIDs are in mysql.gtid_slave_pos -select count(4) <= 4 from mysql.gtid_slave_pos order by domain_id, sub_id; -count(4) <= 4 +SELECT COUNT(*) <= 5*@@GLOBAL.gtid_cleanup_batch_size +FROM mysql.gtid_slave_pos; +COUNT(*) <= 5*@@GLOBAL.gtid_cleanup_batch_size 1 +SET GLOBAL gtid_cleanup_batch_size= @old_gtid_cleanup_batch_size; connection server_1; include/rpl_end.inc diff --git a/mysql-test/suite/rpl/t/rpl_gtid_mdev4484.test b/mysql-test/suite/rpl/t/rpl_gtid_mdev4484.test index e1f5696f5a1..a28bff3d27a 100644 --- a/mysql-test/suite/rpl/t/rpl_gtid_mdev4484.test +++ b/mysql-test/suite/rpl/t/rpl_gtid_mdev4484.test @@ -28,37 +28,79 @@ INSERT INTO t1 VALUES (1); # Inject an artificial error deleting entries, and check that the error handling code works. --connection slave --source include/stop_slave.inc +SET @old_gtid_cleanup_batch_size= @@GLOBAL.gtid_cleanup_batch_size; +SET GLOBAL gtid_cleanup_batch_size= 2; SET @old_dbug= @@GLOBAL.debug_dbug; SET GLOBAL debug_dbug="+d,gtid_slave_pos_simulate_failed_delete"; SET sql_log_bin= 0; -CALL mtr.add_suppression("Can't find file"); +CALL mtr.add_suppression("<DEBUG> Error deleting old GTID row"); SET sql_log_bin= 1; --source include/start_slave.inc

--connection master -INSERT INTO t1 VALUES (2); +--disable_query_log +let $i = 20; +while ($i) { + eval INSERT INTO t1 VALUES ($i+10); + dec $i; +} +--enable_query_log +--save_master_pos

--connection slave ---let $slave_sql_errno= 1942 ---source include/wait_for_slave_sql_error.inc -STOP SLAVE IO_THREAD; -SELECT domain_id, server_id, seq_no FROM mysql.gtid_slave_pos - ORDER BY domain_id, sub_id DESC LIMIT 1; +--sync_with_master + +# Now wait for the slave background thread to try to delete old rows and +# hit the error injection. +--let _TEST_MYSQLD_ERROR_LOG=$MYSQLTEST_VARDIR/log/mysqld.2.err +--perl + open F, '<', $ENV{'_TEST_MYSQLD_ERROR_LOG'} or die; + outer: while (1) { + inner: while (<F>) { + last outer if /<DEBUG> Error deleting old GTID row/; + } + # Easy way to do sub-second sleep without extra modules. + select(undef, undef, undef, 0.1); + } +EOF + +# Since we injected error in the cleanup code, the rows should remain in +# mysql.gtid_slave_pos. Check that we have at least 20 (more robust against +# non-deterministic cleanup and future changes than checking for exact number). +SELECT COUNT(*), MAX(seq_no) INTO @pre_count, @pre_max_seq_no + FROM mysql.gtid_slave_pos; +SELECT IF(@pre_count >= 20, "OK", CONCAT("Error: too few rows seen while errors injected: ", @pre_count)); SET GLOBAL debug_dbug= @old_dbug; ---source include/start_slave.inc

--connection master -INSERT INTO t1 VALUES (3); +--disable_query_log +let $i = 20; +while ($i) { + eval INSERT INTO t1 VALUES ($i+40); + dec $i; +} +--enable_query_log --sync_slave_with_master

--connection slave -SELECT domain_id, server_id, seq_no FROM mysql.gtid_slave_pos - ORDER BY domain_id, sub_id DESC LIMIT 1; -SELECT * FROM t1 ORDER BY i; - +# Now check that 1) rows are being deleted again after removing error +# injection, and 2) old rows are left that failed their delete while errors +# where injected (again compensating for non-deterministic deletion). +# Deletion is async and slightly non-deterministic, so we wait for at +# least 10 of the 20 new rows to be deleted. +let $wait_condition= + SELECT COUNT(*) <= 20-10 + FROM mysql.gtid_slave_pos + WHERE seq_no > @pre_max_seq_no; +--source include/wait_condition.inc +SELECT IF(COUNT(*) >= 1, "OK", CONCAT("Error: too few rows seen after errors no longer injected: ", COUNT(*))) + FROM mysql.gtid_slave_pos + WHERE seq_no <= @pre_max_seq_no;

# Clean up --connection master DROP TABLE t1; +--connection slave +SET GLOBAL gtid_cleanup_batch_size= @old_gtid_cleanup_batch_size;

--source include/rpl_end.inc diff --git a/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test b/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test index e08472d5f51..0060cf4416c 100644 --- a/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test +++ b/mysql-test/suite/rpl/t/rpl_parallel_optimistic.test @@ -21,6 +21,10 @@ SET GLOBAL slave_parallel_threads=10; CHANGE MASTER TO master_use_gtid=slave_pos; SET @old_parallel_mode=@@GLOBAL.slave_parallel_mode; SET GLOBAL slave_parallel_mode='optimistic'; +# Run the first part of the test with high batch size and see that +# old rows remain in the table. +SET @old_gtid_cleanup_batch_size= @@GLOBAL.gtid_cleanup_batch_size; +SET GLOBAL gtid_cleanup_batch_size= 1000000;

--connection server_1 @@ -108,7 +112,12 @@ SELECT * FROM t3 ORDER BY c; SELECT * FROM t1 ORDER BY a; SELECT * FROM t2 ORDER BY a; SELECT * FROM t3 ORDER BY c; -#SHOW STATUS LIKE 'Slave_retried_transactions'; +# Check that we have a bunch of old rows left-over - they were not deleted +# due to high @@gtid_cleanup_batch_size. Then set a low +# @@gtid_cleanup_batch_size so we can test that rows start being deleted. +SELECT IF(COUNT(*) >= 30, "OK", CONCAT("Error: too few old rows found: ", COUNT(*))) + FROM mysql.gtid_slave_pos; +SET GLOBAL gtid_cleanup_batch_size=1;

--echo *** Test @@skip_parallel_replication. *** @@ -557,25 +566,18 @@ DROP TABLE t1, t2, t3;

--connection server_2 --source include/sync_with_master_gtid.inc -# Check for left-over rows in table mysql.gtid_slave_pos (MDEV-12147). -# -# There was a bug when a transaction got a conflict and was rolled back. It -# might have also handled deletion of some old rows, and these deletions would -# then also be rolled back. And since the deletes were never re-tried, old no -# longer needed rows would accumulate in the table without limit. -# -# The earlier part of this test file have plenty of transactions being rolled -# back. But the last DROP TABLE statement runs on its own and should never -# conflict, thus at this point the mysql.gtid_slave_pos table should be clean. -# -# To support @@gtid_pos_auto_engines, when a row is inserted in the table, it -# is associated with the engine of the table at insertion time, and it will -# only be deleted during record_gtid from a table of the same engine. Since we -# alter the table from MyISAM to InnoDB at the start of this test, we should -# end up with 4 rows: two left-over from when the table was MyISAM, and two -# left-over from the InnoDB part. ---echo Check that no more than the expected last four GTIDs are in mysql.gtid_slave_pos -select count(4) <= 4 from mysql.gtid_slave_pos order by domain_id, sub_id; +# Check that old rows are deleted from mysql.gtid_slave_pos. +# Deletion is asynchronous, so use wait_condition.inc. +# Also, there is a small amount of non-determinism in the deletion of old +# rows, so it is not guaranteed that there can never be more than +# @@gtid_cleanup_batch_size rows in the table; so allow a bit of slack +# here. +let $wait_condition= + SELECT COUNT(*) <= 5*@@GLOBAL.gtid_cleanup_batch_size + FROM mysql.gtid_slave_pos; +--source include/wait_condition.inc +eval $wait_condition; +SET GLOBAL gtid_cleanup_batch_size= @old_gtid_cleanup_batch_size;

--connection server_1 --source include/rpl_end.inc diff --git a/mysql-test/suite/sys_vars/r/sysvars_server_notembedded.result b/mysql-test/suite/sys_vars/r/sysvars_server_notembedded.result index e8e4d671eb9..5c5ca8b66b2 100644 --- a/mysql-test/suite/sys_vars/r/sysvars_server_notembedded.result +++ b/mysql-test/suite/sys_vars/r/sysvars_server_notembedded.result @@ -1202,6 +1202,20 @@ NUMERIC_BLOCK_SIZE NULL ENUM_VALUE_LIST NULL READ_ONLY NO COMMAND_LINE_ARGUMENT NULL +VARIABLE_NAME GTID_CLEANUP_BATCH_SIZE +SESSION_VALUE NULL +GLOBAL_VALUE 64 +GLOBAL_VALUE_ORIGIN COMPILE-TIME +DEFAULT_VALUE 64 +VARIABLE_SCOPE GLOBAL +VARIABLE_TYPE INT UNSIGNED +VARIABLE_COMMENT Normally does not need tuning. How many old rows must accumulate in the mysql.gtid_slave_pos table before a background job will be run to delete them. Can be increased to reduce number of commits if using many different engines with --gtid_pos_auto_engines, or to reduce CPU overhead if using a huge number of different gtid_domain_ids. Can be decreased to reduce number of old rows in the table. +NUMERIC_MIN_VALUE 0 +NUMERIC_MAX_VALUE 2147483647 +NUMERIC_BLOCK_SIZE 1 +ENUM_VALUE_LIST NULL +READ_ONLY NO +COMMAND_LINE_ARGUMENT REQUIRED VARIABLE_NAME GTID_CURRENT_POS SESSION_VALUE NULL GLOBAL_VALUE diff --git a/sql/log_event.cc b/sql/log_event.cc index 8813d20578e..e10480fb015 100644 --- a/sql/log_event.cc +++ b/sql/log_event.cc @@ -5565,7 +5565,7 @@ int Query_log_event::do_apply_event(rpl_group_info *rgi, gtid= rgi->current_gtid; if (unlikely(rpl_global_gtid_slave_state->record_gtid(thd, >id, sub_id, - rgi, false, + true, false, &hton))) { int errcode= thd->get_stmt_da()->sql_errno(); @@ -8362,7 +8362,7 @@ Gtid_list_log_event::do_apply_event(rpl_group_info *rgi) { if ((ret= rpl_global_gtid_slave_state->record_gtid(thd, &list[i], sub_id_list[i], - NULL, false, &hton))) + false, false, &hton))) return ret; rpl_global_gtid_slave_state->update_state_hash(sub_id_list[i], &list[i], hton, NULL); @@ -8899,7 +8899,7 @@ int Xid_log_event::do_apply_event(rpl_group_info *rgi) rgi->gtid_pending= false;

gtid= rgi->current_gtid; - err= rpl_global_gtid_slave_state->record_gtid(thd, >id, sub_id, rgi, + err= rpl_global_gtid_slave_state->record_gtid(thd, >id, sub_id, true, false, &hton); if (unlikely(err)) { diff --git a/sql/mysqld.cc b/sql/mysqld.cc index afef4a5f52c..07bdd66f74c 100644 --- a/sql/mysqld.cc +++ b/sql/mysqld.cc @@ -580,6 +580,7 @@ ulong opt_binlog_commit_wait_count= 0; ulong opt_binlog_commit_wait_usec= 0; ulong opt_slave_parallel_max_queued= 131072; my_bool opt_gtid_ignore_duplicates= FALSE; +uint opt_gtid_cleanup_batch_size= 64;

const double log_10[] = { 1e000, 1e001, 1e002, 1e003, 1e004, 1e005, 1e006, 1e007, 1e008, 1e009, diff --git a/sql/mysqld.h b/sql/mysqld.h index d5cabd790b2..261748372f9 100644 --- a/sql/mysqld.h +++ b/sql/mysqld.h @@ -258,6 +258,7 @@ extern ulong opt_slave_parallel_mode; extern ulong opt_binlog_commit_wait_count; extern ulong opt_binlog_commit_wait_usec; extern my_bool opt_gtid_ignore_duplicates; +extern uint opt_gtid_cleanup_batch_size; extern ulong back_log; extern ulong executed_events; extern char language[FN_REFLEN]; diff --git a/sql/rpl_gtid.cc b/sql/rpl_gtid.cc index fabd09adaa7..196c2fe3d16 100644 --- a/sql/rpl_gtid.cc +++ b/sql/rpl_gtid.cc @@ -79,7 +79,7 @@ rpl_slave_state::record_and_update_gtid(THD *thd, rpl_group_info *rgi) rgi->gtid_pending= false; if (rgi->gtid_ignore_duplicate_state!=rpl_group_info::GTID_DUPLICATE_IGNORE) { - if (record_gtid(thd, &rgi->current_gtid, sub_id, NULL, false, &hton)) + if (record_gtid(thd, &rgi->current_gtid, sub_id, false, false, &hton)) DBUG_RETURN(1); update_state_hash(sub_id, &rgi->current_gtid, hton, rgi); } @@ -244,7 +244,7 @@ rpl_slave_state_free_element(void *arg)

rpl_slave_state::rpl_slave_state() - : last_sub_id(0), gtid_pos_tables(0), loaded(false) + : pending_gtid_count(0), last_sub_id(0), gtid_pos_tables(0), loaded(false) { mysql_mutex_init(key_LOCK_slave_state, &LOCK_slave_state, MY_MUTEX_INIT_SLOW); @@ -331,14 +331,11 @@ rpl_slave_state::update(uint32 domain_id, uint32 server_id, uint64 sub_id, } } rgi->gtid_ignore_duplicate_state= rpl_group_info::GTID_DUPLICATE_NULL; - -#ifdef HAVE_REPLICATION - rgi->pending_gtid_deletes_clear(); -#endif }

if (!(list_elem= (list_element *)my_malloc(sizeof(*list_elem), MYF(MY_WME)))) return 1; + list_elem->domain_id= domain_id; list_elem->server_id= server_id; list_elem->sub_id= sub_id; list_elem->seq_no= seq_no; @@ -348,6 +345,15 @@ rpl_slave_state::update(uint32 domain_id, uint32 server_id, uint64 sub_id, if (last_sub_id < sub_id) last_sub_id= sub_id;

+#ifdef HAVE_REPLICATION + ++pending_gtid_count; + if (pending_gtid_count >= opt_gtid_cleanup_batch_size) + { + pending_gtid_count = 0; + slave_background_gtid_pending_delete_request(); + } +#endif + return 0; }

@@ -382,20 +388,22 @@ rpl_slave_state::get_element(uint32 domain_id)

int -rpl_slave_state::put_back_list(uint32 domain_id, list_element *list) +rpl_slave_state::put_back_list(list_element *list) { - element *e; + element *e= NULL; int err= 0;

mysql_mutex_lock(&LOCK_slave_state); - if (!(e= (element *)my_hash_search(&hash, (const uchar *)&domain_id, 0))) - { - err= 1; - goto end; - } while (list) { list_element *next= list->next; + + if ((!e || e->domain_id != list->domain_id) && + !(e= (element *)my_hash_search(&hash, (const uchar *)&list->domain_id, 0))) + { + err= 1; + goto end; + } e->add(list); list= next; } @@ -572,12 +580,12 @@ rpl_slave_state::select_gtid_pos_table(THD *thd, LEX_CSTRING *out_tablename) /* Write a gtid to the replication slave state table.

+ Do it as part of the transaction, to get slave crash safety, or as a separate + transaction if !in_transaction (eg. MyISAM or DDL). + gtid The global transaction id for this event group. sub_id Value allocated within the sub_id when the event group was read (sub_id must be consistent with commit order in master binlog). - rgi rpl_group_info context, if we are recording the gtid transactionally - as part of replicating a transactional event. NULL if called from - outside of a replicated transaction.

Note that caller must later ensure that the new gtid and sub_id is inserted into the appropriate HASH element with rpl_slave_state.add(), so that it can @@ -585,16 +593,13 @@ rpl_slave_state::select_gtid_pos_table(THD *thd, LEX_CSTRING *out_tablename) */ int rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id, - rpl_group_info *rgi, bool in_statement, + bool in_transaction, bool in_statement, void **out_hton) { TABLE_LIST tlist; int err= 0, not_sql_thread; bool table_opened= false; TABLE *table; - list_element *delete_list= 0, *next, *cur, **next_ptr_ptr, **best_ptr_ptr; - uint64 best_sub_id; - element *elem; ulonglong thd_saved_option= thd->variables.option_bits; Query_tables_list lex_backup; wait_for_commit* suspended_wfc; @@ -684,7 +689,7 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id, thd->wsrep_ignore_table= true; #endif

- if (!rgi) + if (!in_transaction) { DBUG_PRINT("info", ("resetting OPTION_BEGIN")); thd->variables.option_bits&= @@ -716,168 +721,280 @@ rpl_slave_state::record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id, my_error(ER_OUT_OF_RESOURCES, MYF(0)); goto end; } +end:

- mysql_mutex_lock(&LOCK_slave_state); - if ((elem= get_element(gtid->domain_id)) == NULL) +#ifdef WITH_WSREP + thd->wsrep_ignore_table= false; +#endif + + if (table_opened) { - mysql_mutex_unlock(&LOCK_slave_state); - my_error(ER_OUT_OF_RESOURCES, MYF(0)); - err= 1; - goto end; + if (err || (err= ha_commit_trans(thd, FALSE))) + ha_rollback_trans(thd, FALSE); + + close_thread_tables(thd); + if (in_transaction) + thd->mdl_context.release_statement_locks(); + else + thd->mdl_context.release_transactional_locks(); } + thd->lex->restore_backup_query_tables_list(&lex_backup); + thd->variables.option_bits= thd_saved_option; + thd->resume_subsequent_commits(suspended_wfc); + DBUG_EXECUTE_IF("inject_record_gtid_serverid_100_sleep", + { + if (gtid->server_id == 100) + my_sleep(500000); + }); + DBUG_RETURN(err); +}

- /* Now pull out all GTIDs that were recorded in this engine. */ - delete_list = NULL; - next_ptr_ptr= &elem->list; - cur= elem->list; - best_sub_id= 0; - best_ptr_ptr= NULL; - while (cur) + +/* + Return a list of all old GTIDs in any mysql.gtid_slave_pos* table that are + no longer needed and can be deleted from the table. + + Within each domain, we need to keep around the latest GTID (the one with the + highest sub_id), but any others in that domain can be deleted. +*/ +rpl_slave_state::list_element * +rpl_slave_state::gtid_grab_pending_delete_list() +{ + uint32 i; + list_element *full_list; + + mysql_mutex_lock(&LOCK_slave_state); + full_list= NULL; + for (i= 0; i < hash.records; ++i) { - list_element *next= cur->next; - if (cur->hton == hton) - { - /* Belongs to same engine, so move it to the delete list. */ - cur->next= delete_list; - delete_list= cur; - if (cur->sub_id > best_sub_id) + element *elem= (element *)my_hash_element(&hash, i); + list_element *elist= elem->list; + list_element *last_elem, **best_ptr_ptr, *cur, *next; + uint64 best_sub_id; + + if (!elist) + continue; /* Nothing here */ + + /* Delete any old stuff, but keep around the most recent one. */ + cur= elist; + best_sub_id= cur->sub_id; + best_ptr_ptr= &elist; + last_elem= cur; + while ((next= cur->next)) { + last_elem= next; + if (next->sub_id > best_sub_id) { - best_sub_id= cur->sub_id; - best_ptr_ptr= &delete_list; - } - else if (best_ptr_ptr == &delete_list) + best_sub_id= next->sub_id; best_ptr_ptr= &cur->next; - } - else - { - /* Another engine, leave it in the list. */ - if (cur->sub_id > best_sub_id) - { - best_sub_id= cur->sub_id; - /* Current best is not on the delete list. */ - best_ptr_ptr= NULL; } - *next_ptr_ptr= cur; - next_ptr_ptr= &cur->next; + cur= next; } - cur= next; - } - *next_ptr_ptr= NULL; - /* - If the highest sub_id element is on the delete list, put it back on the - original list, to preserve the highest sub_id element in the table for - GTID position recovery. - */ - if (best_ptr_ptr) - { + /* + Append the new elements to the full list. Note the order is important; + we do it here so that we do not break the list if best_sub_id is the + last of the new elements. + */ + last_elem->next= full_list; + /* + Delete the highest sub_id element from the old list, and put it back as + the single-element new list. + */ cur= *best_ptr_ptr; *best_ptr_ptr= cur->next; - cur->next= elem->list; + cur->next= NULL; elem->list= cur; + + /* + Collect the full list so far here. Note that elist may have moved if we + deleted the first element, so order is again important. + */ + full_list= elist; } mysql_mutex_unlock(&LOCK_slave_state);

- if (!delete_list) - goto end; + return full_list; +} +

- /* Now delete any already committed GTIDs. */ - bitmap_set_bit(table->read_set, table->field[0]->field_index); - bitmap_set_bit(table->read_set, table->field[1]->field_index); +/* Find the mysql.gtid_slave_posXXX table associated with a given hton. */ +LEX_CSTRING * +rpl_slave_state::select_gtid_pos_table(void *hton) +{ + struct gtid_pos_table *table_entry;

- if ((err= table->file->ha_index_init(0, 0))) + /* + See comments on rpl_slave_state::gtid_pos_tables for rules around proper + access to the list. + */ + table_entry= (struct gtid_pos_table *) + my_atomic_loadptr_explicit(>id_pos_tables, MY_MEMORY_ORDER_ACQUIRE); + + while (table_entry) { - table->file->print_error(err, MYF(0)); - goto end; + if (table_entry->table_hton == hton) + { + if (likely(table_entry->state == GTID_POS_AVAILABLE)) + return &table_entry->table_name; + } + table_entry= table_entry->next; } - cur = delete_list; - while (cur) - { - uchar key_buffer[4+8];

- DBUG_EXECUTE_IF("gtid_slave_pos_simulate_failed_delete", - { err= ENOENT; - table->file->print_error(err, MYF(0)); - /* `break' does not work inside DBUG_EXECUTE_IF */ - goto dbug_break; }); + table_entry= (struct gtid_pos_table *) + my_atomic_loadptr_explicit(&default_gtid_pos_table, MY_MEMORY_ORDER_ACQUIRE); + return &table_entry->table_name; +}

- next= cur->next;

- table->field[1]->store(cur->sub_id, true); - /* domain_id is already set in table->record[0] from write_row() above. */ - key_copy(key_buffer, table->record[0], &table->key_info[0], 0, false); - if (table->file->ha_index_read_map(table->record[1], key_buffer, - HA_WHOLE_KEY, HA_READ_KEY_EXACT)) - /* We cannot find the row, assume it is already deleted. */ - ; - else if ((err= table->file->ha_delete_row(table->record[1]))) - table->file->print_error(err, MYF(0)); - /* - In case of error, we still discard the element from the list. We do - not want to endlessly error on the same element in case of table - corruption or such. - */ - cur= next; - if (err) - break; - } -IF_DBUG(dbug_break:, ) - table->file->ha_index_end(); +void +rpl_slave_state::gtid_delete_pending(THD *thd, + rpl_slave_state::list_element **list_ptr) +{ + int err= 0; + ulonglong thd_saved_option;

-end: + if (unlikely(!loaded)) + return;

#ifdef WITH_WSREP - thd->wsrep_ignore_table= false; + /* + Updates in slave state table should not be appended to galera transaction + writeset. + */ + thd->wsrep_ignore_table= true; #endif

- if (table_opened) + thd_saved_option= thd->variables.option_bits; + thd->variables.option_bits&= + ~(ulonglong)(OPTION_NOT_AUTOCOMMIT |OPTION_BEGIN |OPTION_BIN_LOG | + OPTION_GTID_BEGIN); + + while (*list_ptr) { - if (err || (err= ha_commit_trans(thd, FALSE))) - { - /* - If error, we need to put any remaining delete_list back into the HASH - so we can do another delete attempt later. - */ - if (delete_list) - { - put_back_list(gtid->domain_id, delete_list); - delete_list = 0; - } + LEX_CSTRING *gtid_pos_table_name, *tmp_table_name; + Query_tables_list lex_backup; + TABLE_LIST tlist; + TABLE *table; + handler::Table_flags direct_pos; + list_element *cur, **cur_ptr_ptr; + bool table_opened= false; + void *hton= (*list_ptr)->hton;

- ha_rollback_trans(thd, FALSE); + thd->reset_for_next_command(); + + /* + Only the SQL thread can call select_gtid_pos_table without a mutex + Other threads needs to use a mutex and take into account that the + result may change during execution, so we have to make a copy. + */ + mysql_mutex_lock(&LOCK_slave_state); + tmp_table_name= select_gtid_pos_table(hton); + gtid_pos_table_name= thd->make_clex_string(tmp_table_name->str, + tmp_table_name->length); + mysql_mutex_unlock(&LOCK_slave_state); + if (!gtid_pos_table_name) + { + /* Out of memory - we can try again later. */ + break; } - close_thread_tables(thd); - if (rgi) + + thd->lex->reset_n_backup_query_tables_list(&lex_backup); + tlist.init_one_table(&MYSQL_SCHEMA_NAME, gtid_pos_table_name, NULL, TL_WRITE); + if ((err= open_and_lock_tables(thd, &tlist, FALSE, 0))) + goto end; + table_opened= true; + table= tlist.table; + + if ((err= gtid_check_rpl_slave_state_table(table))) + goto end; + + direct_pos= table->file->ha_table_flags() & HA_PRIMARY_KEY_REQUIRED_FOR_POSITION; + bitmap_set_all(table->write_set); + table->rpl_write_set= table->write_set; + + /* Now delete any already committed GTIDs. */ + bitmap_set_bit(table->read_set, table->field[0]->field_index); + bitmap_set_bit(table->read_set, table->field[1]->field_index); + + if (!direct_pos && (err= table->file->ha_index_init(0, 0))) { - thd->mdl_context.release_statement_locks(); - /* - Save the list of old gtid entries we deleted. If this transaction - fails later for some reason and is rolled back, the deletion of those - entries will be rolled back as well, and we will need to put them back - on the to-be-deleted list so we can re-do the deletion. Otherwise - redundant rows in mysql.gtid_slave_pos may accumulate if transactions - are rolled back and retried after record_gtid(). - */ -#ifdef HAVE_REPLICATION - rgi->pending_gtid_deletes_save(gtid->domain_id, delete_list); -#endif + table->file->print_error(err, MYF(0)); + goto end; } - else + + cur = *list_ptr; + cur_ptr_ptr = list_ptr; + do { - thd->mdl_context.release_transactional_locks(); -#ifdef HAVE_REPLICATION - rpl_group_info::pending_gtid_deletes_free(delete_list); -#endif + uchar key_buffer[4+8]; + list_element *next= cur->next; + + if (cur->hton == hton) + { + int res; + + table->field[0]->store((ulonglong)cur->domain_id, true); + table->field[1]->store(cur->sub_id, true); + if (direct_pos) + { + res= table->file->ha_rnd_pos_by_record(table->record[0]); + } + else + { + key_copy(key_buffer, table->record[0], &table->key_info[0], 0, false); + res= table->file->ha_index_read_map(table->record[0], key_buffer, + HA_WHOLE_KEY, HA_READ_KEY_EXACT); + } + DBUG_EXECUTE_IF("gtid_slave_pos_simulate_failed_delete", + { res= 1; + err= ENOENT; + sql_print_error("<DEBUG> Error deleting old GTID row"); + }); + if (res) + /* We cannot find the row, assume it is already deleted. */ + ; + else if ((err= table->file->ha_delete_row(table->record[0]))) + { + sql_print_error("Error deleting old GTID row: %s", + thd->get_stmt_da()->message()); + /* + In case of error, we still discard the element from the list. We do + not want to endlessly error on the same element in case of table + corruption or such. + */ + } + *cur_ptr_ptr= next; + my_free(cur); + } + else + { + /* Leave this one in the list until we get to the table for its hton. */ + cur_ptr_ptr= &cur->next; + } + cur= next; + if (err) + break; + } while (cur); +end: + if (table_opened) + { + if (!direct_pos) + table->file->ha_index_end(); + + if (err || (err= ha_commit_trans(thd, FALSE))) + ha_rollback_trans(thd, FALSE); } + close_thread_tables(thd); + thd->mdl_context.release_transactional_locks(); + thd->lex->restore_backup_query_tables_list(&lex_backup); + + if (err) + break; } - thd->lex->restore_backup_query_tables_list(&lex_backup); thd->variables.option_bits= thd_saved_option; - thd->resume_subsequent_commits(suspended_wfc); - DBUG_EXECUTE_IF("inject_record_gtid_serverid_100_sleep", - { - if (gtid->server_id == 100) - my_sleep(500000); - }); - DBUG_RETURN(err); + +#ifdef WITH_WSREP + thd->wsrep_ignore_table= false; +#endif }

@@ -1251,7 +1368,7 @@ rpl_slave_state::load(THD *thd, const char *state_from_master, size_t len,

if (gtid_parser_helper(&state_from_master, end, >id) || !(sub_id= next_sub_id(gtid.domain_id)) || - record_gtid(thd, >id, sub_id, NULL, in_statement, &hton) || + record_gtid(thd, >id, sub_id, false, in_statement, &hton) || update(gtid.domain_id, gtid.server_id, sub_id, gtid.seq_no, hton, NULL)) return 1; if (state_from_master == end) diff --git a/sql/rpl_gtid.h b/sql/rpl_gtid.h index 0fc92d5e33c..60d822f7b0d 100644 --- a/sql/rpl_gtid.h +++ b/sql/rpl_gtid.h @@ -118,8 +118,9 @@ struct rpl_slave_state { struct list_element *next; uint64 sub_id; - uint64 seq_no; + uint32 domain_id; uint32 server_id; + uint64 seq_no; /* hton of mysql.gtid_slave_pos* table used to record this GTID. Can be NULL if the gtid table failed to load (eg. missing @@ -191,6 +192,8 @@ struct rpl_slave_state

/* Mapping from domain_id to its element. */ HASH hash; + /* GTIDs added since last purge of old mysql.gtid_slave_pos rows. */ + uint32 pending_gtid_count; /* Mutex protecting access to the state. */ mysql_mutex_t LOCK_slave_state; /* Auxiliary buffer to sort gtid list. */ @@ -233,7 +236,10 @@ struct rpl_slave_state int truncate_state_table(THD *thd); void select_gtid_pos_table(THD *thd, LEX_CSTRING *out_tablename); int record_gtid(THD *thd, const rpl_gtid *gtid, uint64 sub_id, - rpl_group_info *rgi, bool in_statement, void **out_hton); + bool in_transaction, bool in_statement, void **out_hton); + list_element *gtid_grab_pending_delete_list(); + LEX_CSTRING *select_gtid_pos_table(void *hton); + void gtid_delete_pending(THD *thd, rpl_slave_state::list_element **list_ptr); uint64 next_sub_id(uint32 domain_id); int iterate(int (*cb)(rpl_gtid *, void *), void *data, rpl_gtid *extra_gtids, uint32 num_extra, @@ -245,7 +251,7 @@ struct rpl_slave_state bool is_empty();

element *get_element(uint32 domain_id); - int put_back_list(uint32 domain_id, list_element *list); + int put_back_list(list_element *list);

void update_state_hash(uint64 sub_id, rpl_gtid *gtid, void *hton, rpl_group_info *rgi); diff --git a/sql/rpl_rli.cc b/sql/rpl_rli.cc index b275ad884bd..2d91620c898 100644 --- a/sql/rpl_rli.cc +++ b/sql/rpl_rli.cc @@ -1820,6 +1820,7 @@ rpl_load_gtid_slave_state(THD *thd) int err= 0; uint32 i; load_gtid_state_cb_data cb_data; + rpl_slave_state::list_element *old_gtids_list; DBUG_ENTER("rpl_load_gtid_slave_state");

mysql_mutex_lock(&rpl_global_gtid_slave_state->LOCK_slave_state); @@ -1905,6 +1906,13 @@ rpl_load_gtid_slave_state(THD *thd) rpl_global_gtid_slave_state->loaded= true; mysql_mutex_unlock(&rpl_global_gtid_slave_state->LOCK_slave_state);

+ /* Clear out no longer needed elements now. */ + old_gtids_list= + rpl_global_gtid_slave_state->gtid_grab_pending_delete_list(); + rpl_global_gtid_slave_state->gtid_delete_pending(thd, &old_gtids_list); + if (old_gtids_list) + rpl_global_gtid_slave_state->put_back_list(old_gtids_list); + end: if (array_inited) delete_dynamic(&array); @@ -2086,7 +2094,6 @@ rpl_group_info::reinit(Relay_log_info *rli) long_find_row_note_printed= false; did_mark_start_commit= false; gtid_ev_flags2= 0; - pending_gtid_delete_list= NULL; last_master_timestamp = 0; gtid_ignore_duplicate_state= GTID_DUPLICATE_NULL; speculation= SPECULATE_NO; @@ -2217,12 +2224,6 @@ void rpl_group_info::cleanup_context(THD *thd, bool error) erroneously update the GTID position. */ gtid_pending= false; - - /* - Rollback will have undone any deletions of old rows we might have made - in mysql.gtid_slave_pos. Put those rows back on the list to be deleted. - */ - pending_gtid_deletes_put_back(); } m_table_map.clear_tables(); slave_close_thread_tables(thd); @@ -2448,78 +2449,6 @@ rpl_group_info::unmark_start_commit() }

-/* - When record_gtid() has deleted any old rows from the table - mysql.gtid_slave_pos as part of a replicated transaction, save the list of - rows deleted here. - - If later the transaction fails (eg. optimistic parallel replication), the - deletes will be undone when the transaction is rolled back. Then we can - put back the list of rows into the rpl_global_gtid_slave_state, so that - we can re-do the deletes and avoid accumulating old rows in the table. -*/ -void -rpl_group_info::pending_gtid_deletes_save(uint32 domain_id, - rpl_slave_state::list_element *list) -{ - /* - We should never get to a state where we try to save a new pending list of - gtid deletes while we still have an old one. But make sure we handle it - anyway just in case, so we avoid leaving stray entries in the - mysql.gtid_slave_pos table. - */ - DBUG_ASSERT(!pending_gtid_delete_list); - if (unlikely(pending_gtid_delete_list)) - pending_gtid_deletes_put_back(); - - pending_gtid_delete_list= list; - pending_gtid_delete_list_domain= domain_id; -} - - -/* - Take the list recorded by pending_gtid_deletes_save() and put it back into - rpl_global_gtid_slave_state. This is needed if deletion of the rows was - rolled back due to transaction failure. -*/ -void -rpl_group_info::pending_gtid_deletes_put_back() -{ - if (pending_gtid_delete_list) - { - rpl_global_gtid_slave_state->put_back_list(pending_gtid_delete_list_domain, - pending_gtid_delete_list); - pending_gtid_delete_list= NULL; - } -} - - -/* - Free the list recorded by pending_gtid_deletes_save(). Done when the deletes - in the list have been permanently committed. -*/ -void -rpl_group_info::pending_gtid_deletes_clear() -{ - pending_gtid_deletes_free(pending_gtid_delete_list); - pending_gtid_delete_list= NULL; -} - - -void -rpl_group_info::pending_gtid_deletes_free(rpl_slave_state::list_element *list) -{ - rpl_slave_state::list_element *next; - - while (list) - { - next= list->next; - my_free(list); - list= next; - } -} - - rpl_sql_thread_info::rpl_sql_thread_info(Rpl_filter *filter) : rpl_filter(filter) { diff --git a/sql/rpl_rli.h b/sql/rpl_rli.h index d9f0e0e5d3b..b8b153c34be 100644 --- a/sql/rpl_rli.h +++ b/sql/rpl_rli.h @@ -757,11 +757,6 @@ struct rpl_group_info /* Needs room for "Gtid D-S-N\x00". */ char gtid_info_buf[5+10+1+10+1+20+1];

- /* List of not yet committed deletions in mysql.gtid_slave_pos. */ - rpl_slave_state::list_element *pending_gtid_delete_list; - /* Domain associated with pending_gtid_delete_list. */ - uint32 pending_gtid_delete_list_domain; - /* The timestamp, from the master, of the commit event. Used to do delayed update of rli->last_master_timestamp, for getting @@ -903,12 +898,6 @@ struct rpl_group_info char *gtid_info(); void unmark_start_commit();

- static void pending_gtid_deletes_free(rpl_slave_state::list_element *list); - void pending_gtid_deletes_save(uint32 domain_id, - rpl_slave_state::list_element *list); - void pending_gtid_deletes_put_back(); - void pending_gtid_deletes_clear(); - longlong get_row_stmt_start_timestamp() { return row_stmt_start_timestamp; diff --git a/sql/slave.cc b/sql/slave.cc index bb1300d36e6..f8499513dd6 100644 --- a/sql/slave.cc +++ b/sql/slave.cc @@ -465,6 +465,8 @@ static struct slave_background_gtid_pos_create_t { void *hton; } *slave_background_gtid_pos_create_list;

+static volatile bool slave_background_gtid_pending_delete_flag; +

pthread_handler_t handle_slave_background(void *arg __attribute__((unused))) @@ -499,6 +501,7 @@ handle_slave_background(void *arg __attribute__((unused))) { slave_background_kill_t *kill_list; slave_background_gtid_pos_create_t *create_list; + bool pending_deletes;

thd->ENTER_COND(&COND_slave_background, &LOCK_slave_background, &stage_slave_background_wait_request, @@ -508,13 +511,15 @@ handle_slave_background(void *arg __attribute__((unused))) stop= abort_loop || thd->killed || slave_background_thread_stop; kill_list= slave_background_kill_list; create_list= slave_background_gtid_pos_create_list; - if (stop || kill_list || create_list) + pending_deletes= slave_background_gtid_pending_delete_flag; + if (stop || kill_list || create_list || pending_deletes) break; mysql_cond_wait(&COND_slave_background, &LOCK_slave_background); }

slave_background_kill_list= NULL; slave_background_gtid_pos_create_list= NULL; + slave_background_gtid_pending_delete_flag= false; thd->EXIT_COND(&old_stage);

while (kill_list) @@ -541,6 +546,17 @@ handle_slave_background(void *arg __attribute__((unused))) create_list= next; }

+ if (pending_deletes) + { + rpl_slave_state::list_element *list; + + slave_background_gtid_pending_delete_flag= false; + list= rpl_global_gtid_slave_state->gtid_grab_pending_delete_list(); + rpl_global_gtid_slave_state->gtid_delete_pending(thd, &list); + if (list) + rpl_global_gtid_slave_state->put_back_list(list); + } + mysql_mutex_lock(&LOCK_slave_background); } while (!stop);

@@ -615,6 +631,23 @@ slave_background_gtid_pos_create_request(

/* + Request the slave background thread to delete no longer used rows from the + mysql.gtid_slave_pos* tables. + + This is called from time-critical rpl_slave_state::update(), so we avoid + taking any locks here. This means we may race with the background thread + to occasionally lose a signal. This is not a problem; any pending rows to + be deleted will just be deleted a bit later as part of the next batch. +*/ +void +slave_background_gtid_pending_delete_request(void) +{ + slave_background_gtid_pending_delete_flag= true; + mysql_cond_signal(&COND_slave_background); +} + + +/* Start the slave background thread.

This thread is currently used for two purposes: diff --git a/sql/slave.h b/sql/slave.h index 649d55b45b9..12d569b0333 100644 --- a/sql/slave.h +++ b/sql/slave.h @@ -276,6 +276,7 @@ bool net_request_file(NET* net, const char* fname); void slave_background_kill_request(THD *to_kill); void slave_background_gtid_pos_create_request (rpl_slave_state::gtid_pos_table *table_entry); +void slave_background_gtid_pending_delete_request(void);

extern bool volatile abort_loop; extern Master_info *active_mi; /* active_mi for multi-master */ diff --git a/sql/sys_vars.cc b/sql/sys_vars.cc index 6d4c135683a..9348f4e5c98 100644 --- a/sql/sys_vars.cc +++ b/sql/sys_vars.cc @@ -1942,6 +1942,19 @@ Sys_var_last_gtid::session_value_ptr(THD *thd, const LEX_CSTRING *base) }

+static Sys_var_uint Sys_gtid_cleanup_batch_size( + "gtid_cleanup_batch_size", + "Normally does not need tuning. How many old rows must accumulate in " + "the mysql.gtid_slave_pos table before a background job will be run to " + "delete them. Can be increased to reduce number of commits if " + "using many different engines with --gtid_pos_auto_engines, or to " + "reduce CPU overhead if using a huge number of different " + "gtid_domain_ids. Can be decreased to reduce number of old rows in the " + "table.", + GLOBAL_VAR(opt_gtid_cleanup_batch_size), CMD_LINE(REQUIRED_ARG), + VALID_RANGE(0,2147483647), DEFAULT(64), BLOCK_SIZE(1)); + + static bool check_slave_parallel_threads(sys_var *self, THD *thd, set_var *var) { diff --git a/storage/rocksdb/mysql-test/rocksdb_rpl/r/mdev12179.result b/storage/rocksdb/mysql-test/rocksdb_rpl/r/mdev12179.result index 9c20fea97ae..a1e501f78f4 100644 --- a/storage/rocksdb/mysql-test/rocksdb_rpl/r/mdev12179.result +++ b/storage/rocksdb/mysql-test/rocksdb_rpl/r/mdev12179.result @@ -2,6 +2,7 @@ include/master-slave.inc [connection master] connection server_2; include/stop_slave.inc +SET GLOBAL gtid_cleanup_batch_size = 999999999; CHANGE MASTER TO master_use_gtid=slave_pos; SET sql_log_bin=0; CREATE TABLE mysql.gtid_slave_pos_innodb LIKE mysql.gtid_slave_pos; @@ -41,6 +42,8 @@ a 1 SELECT * FROM mysql.gtid_slave_pos ORDER BY sub_id; domain_id sub_id server_id seq_no +0 1 1 1 +0 2 1 2 0 3 1 3 0 4 1 4 SELECT * FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb @@ -121,6 +124,21 @@ Transactions_multi_engine 6 DELETE FROM t1 WHERE a >= 100; DELETE FROM t2 WHERE a >= 100; DELETE FROM t3 WHERE a >= 100; +connection server_1; +include/save_master_gtid.inc +connection server_2; +include/sync_with_master_gtid.inc +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos; +COUNT(*)>=10 +1 +SELECT COUNT(*)>=10 FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb +UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select; +COUNT(*)>=10 +1 +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos_rocksdb; +COUNT(*)>=10 +1 +SET GLOBAL gtid_cleanup_batch_size = 3; connection server_2; include/stop_slave.inc SET sql_log_bin=0; diff --git a/storage/rocksdb/mysql-test/rocksdb_rpl/t/mdev12179.test b/storage/rocksdb/mysql-test/rocksdb_rpl/t/mdev12179.test index e0d16e7f242..631d9ca533f 100644 --- a/storage/rocksdb/mysql-test/rocksdb_rpl/t/mdev12179.test +++ b/storage/rocksdb/mysql-test/rocksdb_rpl/t/mdev12179.test @@ -4,6 +4,12 @@

--connection server_2 --source include/stop_slave.inc + +# Set GTID cleanup limit high enough that cleanup will not run and we +# can rely on consistent table output in .result. +--let $old_gtid_cleanup_batch_size=`SELECT @@GLOBAL.gtid_cleanup_batch_size` +SET GLOBAL gtid_cleanup_batch_size = 999999999; + CHANGE MASTER TO master_use_gtid=slave_pos; SET sql_log_bin=0; CREATE TABLE mysql.gtid_slave_pos_innodb LIKE mysql.gtid_slave_pos; @@ -89,6 +95,82 @@ DELETE FROM t2 WHERE a >= 100; DELETE FROM t3 WHERE a >= 100;

+# Create a bunch more GTIDs in mysql.gtid_slave_pos* tables to test with. +--connection server_1 +--disable_query_log +let $i=10; +while ($i) { + eval INSERT INTO t1 VALUES (300+$i); + eval INSERT INTO t2 VALUES (300+$i); + eval INSERT INTO t3 VALUES (300+$i); + dec $i; +} +--enable_query_log +--source include/save_master_gtid.inc + +--connection server_2 +--source include/sync_with_master_gtid.inc + +# Check that we have many rows in mysql.gtid_slave_pos now (since +# @@gtid_cleanup_batch_size was set to a huge value). No need to check +# for an exact number, since that will require changing .result if +# anything changes prior to this point, and we just need to know that +# we have still have some data in the tables to make the following +# test effective. +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos; +SELECT COUNT(*)>=10 FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb + UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select; +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos_rocksdb; + +# Check that old GTID rows will be deleted when batch delete size is +# set reasonably. Old row deletion is not 100% deterministic (by design), so +# we must wait for it to occur, but it should occur eventually. +SET GLOBAL gtid_cleanup_batch_size = 3; +let $i=40; +--disable_query_log +--let $keep_include_silent=1 +while ($i) { + let N=`SELECT 1+($i MOD 3)`; + --connection server_1 + eval UPDATE t$N SET a=a+1 WHERE a=(SELECT MAX(a) FROM t$N); + --source include/save_master_gtid.inc + --connection server_2 + --source include/sync_with_master_gtid.inc + let $j=50; + while ($j) { + let $is_done=`SELECT SUM(a)=1 FROM ( + SELECT COUNT(*) AS a FROM mysql.gtid_slave_pos + UNION ALL + SELECT COUNT(*) AS a FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb + UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select + UNION ALL + SELECT COUNT(*) AS a FROM mysql.gtid_slave_pos_rocksdb) outer_select`; + if ($is_done) { + let $j=0; + } + if (!$is_done) { + real_sleep 0.1; + dec $j; + } + } + dec $i; + if ($is_done) { + let $i=0; + } +} +--enable_query_log +--let $keep_include_silent=0 +if (!$is_done) { + --echo Timed out waiting for mysql.gtid_slave_pos* tables to be cleaned up +} + +--disable_query_log +DELETE FROM t1 WHERE a >= 100; +DELETE FROM t2 WHERE a >= 100; +DELETE FROM t3 WHERE a >= 100; +--enable_query_log + + # Test status variables Rpl_transactions_multi_engine and Transactions_gtid_foreign_engine. # Have mysql.gtid_slave_pos* for myisam and innodb but not rocksdb. --connection server_2 @@ -223,6 +305,9 @@ SHOW STATUS LIKE "%transactions%engine"; SET sql_log_bin=0; DROP TABLE mysql.gtid_slave_pos_innodb; SET sql_log_bin=1; +--disable_query_log +eval SET GLOBAL gtid_cleanup_batch_size = $old_gtid_cleanup_batch_size; +--enable_query_log

--connection server_1 DROP TABLE t1; diff --git a/storage/tokudb/mysql-test/tokudb_rpl/r/mdev12179.result b/storage/tokudb/mysql-test/tokudb_rpl/r/mdev12179.result index d4532eec4e2..d79e7e59aa4 100644 --- a/storage/tokudb/mysql-test/tokudb_rpl/r/mdev12179.result +++ b/storage/tokudb/mysql-test/tokudb_rpl/r/mdev12179.result @@ -2,6 +2,7 @@ include/master-slave.inc [connection master] connection server_2; include/stop_slave.inc +SET GLOBAL gtid_cleanup_batch_size = 999999999; CHANGE MASTER TO master_use_gtid=slave_pos; SET sql_log_bin=0; CREATE TABLE mysql.gtid_slave_pos_innodb LIKE mysql.gtid_slave_pos; @@ -41,6 +42,8 @@ a 1 SELECT * FROM mysql.gtid_slave_pos ORDER BY sub_id; domain_id sub_id server_id seq_no +0 1 1 1 +0 2 1 2 0 3 1 3 0 4 1 4 SELECT * FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb @@ -121,6 +124,21 @@ Transactions_multi_engine 6 DELETE FROM t1 WHERE a >= 100; DELETE FROM t2 WHERE a >= 100; DELETE FROM t3 WHERE a >= 100; +connection server_1; +include/save_master_gtid.inc +connection server_2; +include/sync_with_master_gtid.inc +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos; +COUNT(*)>=10 +1 +SELECT COUNT(*)>=10 FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb +UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select; +COUNT(*)>=10 +1 +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos_tokudb; +COUNT(*)>=10 +1 +SET GLOBAL gtid_cleanup_batch_size = 3; connection server_2; include/stop_slave.inc SET sql_log_bin=0; diff --git a/storage/tokudb/mysql-test/tokudb_rpl/t/mdev12179.test b/storage/tokudb/mysql-test/tokudb_rpl/t/mdev12179.test index ceb119cd0dc..1d19a25889e 100644 --- a/storage/tokudb/mysql-test/tokudb_rpl/t/mdev12179.test +++ b/storage/tokudb/mysql-test/tokudb_rpl/t/mdev12179.test @@ -4,6 +4,12 @@

--connection server_2 --source include/stop_slave.inc + +# Set GTID cleanup limit high enough that cleanup will not run and we +# can rely on consistent table output in .result. +--let $old_gtid_cleanup_batch_size=`SELECT @@GLOBAL.gtid_cleanup_batch_size` +SET GLOBAL gtid_cleanup_batch_size = 999999999; + CHANGE MASTER TO master_use_gtid=slave_pos; SET sql_log_bin=0; CREATE TABLE mysql.gtid_slave_pos_innodb LIKE mysql.gtid_slave_pos; @@ -89,6 +95,82 @@ DELETE FROM t2 WHERE a >= 100; DELETE FROM t3 WHERE a >= 100;

+# Create a bunch more GTIDs in mysql.gtid_slave_pos* tables to test with. +--connection server_1 +--disable_query_log +let $i=10; +while ($i) { + eval INSERT INTO t1 VALUES (300+$i); + eval INSERT INTO t2 VALUES (300+$i); + eval INSERT INTO t3 VALUES (300+$i); + dec $i; +} +--enable_query_log +--source include/save_master_gtid.inc + +--connection server_2 +--source include/sync_with_master_gtid.inc + +# Check that we have many rows in mysql.gtid_slave_pos now (since +# @@gtid_cleanup_batch_size was set to a huge value). No need to check +# for an exact number, since that will require changing .result if +# anything changes prior to this point, and we just need to know that +# we have still have some data in the tables to make the following +# test effective. +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos; +SELECT COUNT(*)>=10 FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb + UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select; +SELECT COUNT(*)>=10 FROM mysql.gtid_slave_pos_tokudb; + +# Check that old GTID rows will be deleted when batch delete size is +# set reasonably. Old row deletion is not 100% deterministic (by design), so +# we must wait for it to occur, but it should occur eventually. +SET GLOBAL gtid_cleanup_batch_size = 3; +let $i=40; +--disable_query_log +--let $keep_include_silent=1 +while ($i) { + let N=`SELECT 1+($i MOD 3)`; + --connection server_1 + eval UPDATE t$N SET a=a+1 WHERE a=(SELECT MAX(a) FROM t$N); + --source include/save_master_gtid.inc + --connection server_2 + --source include/sync_with_master_gtid.inc + let $j=50; + while ($j) { + let $is_done=`SELECT SUM(a)=1 FROM ( + SELECT COUNT(*) AS a FROM mysql.gtid_slave_pos + UNION ALL + SELECT COUNT(*) AS a FROM ( SELECT * FROM mysql.gtid_slave_pos_innodb + UNION ALL SELECT * FROM mysql.gtid_slave_pos_innodb_redundant) inner_select + UNION ALL + SELECT COUNT(*) AS a FROM mysql.gtid_slave_pos_tokudb) outer_select`; + if ($is_done) { + let $j=0; + } + if (!$is_done) { + real_sleep 0.1; + dec $j; + } + } + dec $i; + if ($is_done) { + let $i=0; + } +} +--enable_query_log +--let $keep_include_silent=0 +if (!$is_done) { + --echo Timed out waiting for mysql.gtid_slave_pos* tables to be cleaned up +} + +--disable_query_log +DELETE FROM t1 WHERE a >= 100; +DELETE FROM t2 WHERE a >= 100; +DELETE FROM t3 WHERE a >= 100; +--enable_query_log + + # Test status variables Rpl_transactions_multi_engine and Transactions_gtid_foreign_engine. # Have mysql.gtid_slave_pos* for myisam and innodb but not tokudb. --connection server_2 @@ -223,6 +305,9 @@ SHOW STATUS LIKE "%transactions%engine"; SET sql_log_bin=0; DROP TABLE mysql.gtid_slave_pos_innodb; SET sql_log_bin=1; +--disable_query_log +eval SET GLOBAL gtid_cleanup_batch_size = $old_gtid_cleanup_batch_size; +--enable_query_log

--connection server_1 DROP TABLE t1;