Adding option to enable global commit ordering on secondaries

Summary:
Extending the variable mts_dependency_order_commits to allow
three values NONE, DB and GLOBAL. When the variabale it set to GLOBAL we
order commits even across databases.

Reviewed By: bhatvinay

Differential Revision: D23366447

fbshipit-source-id: dad1c58e504

mysql-test/r/mysqld--help-notwin-profiling.result

@@ -794,10 +794,9 @@ The following options may be given as the first argument:
  Max number of keys in a transaction after which it will
  be executed in isolation. This limits the amount of
  metadata we'll need to maintain.
- --mts-dependency-order-commits 
- Commit trxs in the same order as the master (per
- database)
- (Defaults to on; use --skip-mts-dependency-order-commits to disable.)
+ --mts-dependency-order-commits[=name] 
+ Commit trxs in the same order as the master (per database
+ or globally)
  --mts-dependency-refill-threshold[=#] 
  Capacity in percentage at which to start refilling the
  dependency buffer
@@ -2523,7 +2522,7 @@ min-examined-row-limit 0
 minimum-hlc-ns 0
 mts-dependency-cond-wait-timeout 5000
 mts-dependency-max-keys 100000
-mts-dependency-order-commits TRUE
+mts-dependency-order-commits DB
 mts-dependency-refill-threshold 60
 mts-dependency-replication NONE
 mts-dependency-size 1000
@@ -2606,7 +2605,7 @@ performance-schema-max-rwlock-instances -1
 performance-schema-max-socket-classes 10
 performance-schema-max-socket-instances -1
 performance-schema-max-stage-classes 150
-performance-schema-max-statement-classes 188
+performance-schema-max-statement-classes 191
 performance-schema-max-table-handles -1
 performance-schema-max-table-instances -1
 performance-schema-max-thread-classes 50

mysql-test/r/mysqld--help-notwin.result

@@ -794,10 +794,9 @@ The following options may be given as the first argument:
  Max number of keys in a transaction after which it will
  be executed in isolation. This limits the amount of
  metadata we'll need to maintain.
- --mts-dependency-order-commits 
- Commit trxs in the same order as the master (per
- database)
- (Defaults to on; use --skip-mts-dependency-order-commits to disable.)
+ --mts-dependency-order-commits[=name] 
+ Commit trxs in the same order as the master (per database
+ or globally)
  --mts-dependency-refill-threshold[=#] 
  Capacity in percentage at which to start refilling the
  dependency buffer
@@ -2521,7 +2520,7 @@ min-examined-row-limit 0
 minimum-hlc-ns 0
 mts-dependency-cond-wait-timeout 5000
 mts-dependency-max-keys 100000
-mts-dependency-order-commits TRUE
+mts-dependency-order-commits DB
 mts-dependency-refill-threshold 60
 mts-dependency-replication NONE
 mts-dependency-size 1000
@@ -2604,7 +2603,7 @@ performance-schema-max-rwlock-instances -1
 performance-schema-max-socket-classes 10
 performance-schema-max-socket-instances -1
 performance-schema-max-stage-classes 150
-performance-schema-max-statement-classes 188
+performance-schema-max-statement-classes 191
 performance-schema-max-table-handles -1
 performance-schema-max-table-instances -1
 performance-schema-max-thread-classes 50

mysql-test/suite/rpl_mts/r/rpl_mts_dependency_order_commits.result

@@ -7,6 +7,10 @@ stop slave;
 set @save.slave_parallel_workers= @@global.slave_parallel_workers;
 set @@global.slave_parallel_workers= 8;
 start slave;
+select @@global.mts_dependency_order_commits;
+@@global.mts_dependency_order_commits
+DB
+set @save.mts_dependency_order_commits= @@global.mts_dependency_order_commits;
 create database d1;
 create database d2;
 create table d1.t1(a int primary key) engine=innodb;
@@ -46,7 +50,44 @@ select * from d2.t1;
 a
 1
 10
+stop slave;
+set @@global.mts_dependency_order_commits = GLOBAL;
+start slave;
+delete from d1.t1;
+delete from d1.t2;
+delete from d2.t1;
+insert into d1.t1 values(1);
+begin;
+update d1.t1 set a = 2 where a = 1;
+update d1.t1 set a = 3 where a = 1;
+insert into d2.t1 values(1);
+begin;
+insert into d1.t1 values(10);
+insert into d2.t1 values(10);
+commit;
+"d1.t1 should contain a single row with value 1"
+select * from d1.t1;
+a
+1
+"d2.t1 should also be empty"
+select * from d2.t1;
+a
+rollback;
+"Final states"
+select * from d1.t1;
+a
+3
+10
+select * from d1.t2;
+a
+select * from d2.t1;
+a
+1
+10
+stop slave;
 set @@global.slave_parallel_workers= @save.slave_parallel_workers;
+set @@global.mts_dependency_order_commits= @save.mts_dependency_order_commits;
+start slave;
 drop database d1;
 drop database d2;
 include/rpl_end.inc

mysql-test/suite/rpl_mts/t/rpl_mts_dependency_order_commits.test

@@ -1,5 +1,5 @@
 # Tests if mts_dependency_order_commits= TRUE master's commit order is honored
-# on the slave (per database)
+# on the slave
 
 source include/have_innodb.inc;
 source include/have_mts_dependency_replication.inc;
@@ -11,6 +11,10 @@ set @save.slave_parallel_workers= @@global.slave_parallel_workers;
 set @@global.slave_parallel_workers= 8;
 start slave;
 
+# Case 1: Per DB commit ordering (default behavior)
+select @@global.mts_dependency_order_commits;
+set @save.mts_dependency_order_commits= @@global.mts_dependency_order_commits;
+
 connection master;
 create database d1;
 create database d2;
@@ -70,8 +74,67 @@ select * from d1.t1;
 select * from d1.t2;
 select * from d2.t1;
 
+
+# Case 2: Global commit ordering
+connection slave;
+stop slave;
+set @@global.mts_dependency_order_commits = GLOBAL;
+start slave;
+connection master;
+delete from d1.t1;
+delete from d1.t2;
+delete from d2.t1;
+insert into d1.t1 values(1);
+sync_slave_with_master;
+
+connection slave;
+# create a blocking trx on the slave
+begin;
+update d1.t1 set a = 2 where a = 1;
+
+connection master;
+update d1.t1 set a = 3 where a = 1; # this will be blocked due to engine lock
+insert into d2.t1 values(1); # this will be blocked because we're ordering globally
+begin;
+insert into d1.t1 values(10);
+insert into d2.t1 values(10);
+commit; # this will wait for all workers to finish (to execute in isolation)
+
+
+connection slave1;
+# Wait for 2nd trx above to block on commit order
+let $wait_condition= SELECT COUNT(*) = 1 FROM INFORMATION_SCHEMA.PROCESSLIST
+WHERE STATE LIKE "%Waiting for preceding transaction to commit%";
+let $wait_timeout= 120;
+source include/wait_condition.inc;
+
+# Wait for 3rd trx to wait for workers to finish
+let $wait_condition= SELECT COUNT(*) = 1 FROM INFORMATION_SCHEMA.PROCESSLIST
+WHERE STATE LIKE "%Waiting for dependency workers to finish%";
+let $wait_timeout= 120;
+source include/wait_condition.inc;
+
+echo "d1.t1 should contain a single row with value 1";
+select * from d1.t1;
+echo "d2.t1 should also be empty";
+select * from d2.t1;
+
+connection slave;
+rollback; # rollback the conflicting trx
+
+connection master;
+sync_slave_with_master;
+echo "Final states";
+select * from d1.t1;
+select * from d1.t2;
+select * from d2.t1;
+
 # cleanup
+connection slave;
+stop slave;
 set @@global.slave_parallel_workers= @save.slave_parallel_workers;
+set @@global.mts_dependency_order_commits= @save.mts_dependency_order_commits;
+start slave;
 connection master;
 drop database d1;
 drop database d2;

mysql-test/suite/sys_vars/r/mts_dependency_order_commits_basic.result

@@ -8,7 +8,7 @@ variable_name
 set @@global.mts_dependency_order_commits= false;
 select @@global.mts_dependency_order_commits;
 @@global.mts_dependency_order_commits
-0
+NONE
 set @@global.mts_dependency_order_commits= 1.1;
 ERROR 42000: Incorrect argument type to variable 'mts_dependency_order_commits'
 set @@global.mts_dependency_order_commits= "foo";
@@ -17,5 +17,5 @@ set @@global.mts_dependency_order_commits= false;
 set @@global.mts_dependency_order_commits= true;
 select @@global.mts_dependency_order_commits as "truncated to the maximum";
 truncated to the maximum
-1
+DB
 set @@global.mts_dependency_order_commits= @save.mts_dependency_order_commits;

sql/dependency_slave_worker.cc

@@ -43,7 +43,9 @@ Dependency_slave_worker::get_begin_event(Commit_order_manager *co_mngr)
   if (ret && co_mngr != NULL)
   {
     DBUG_ASSERT(c_rli->mts_dependency_order_commits);
-    set_current_db(ret->get_db());
+    // case: if we need to order commits by DB we set current DB
+    if (c_rli->mts_dependency_order_commits == DEP_RPL_ORDER_DB)
+      set_current_db(ret->get_db());
     co_mngr->register_trx(this);
   }
 

sql/mysqld.cc

@@ -801,7 +801,7 @@ ulong opt_mts_dependency_replication;
 ulonglong opt_mts_dependency_size;
 double opt_mts_dependency_refill_threshold;
 ulonglong opt_mts_dependency_max_keys;
-my_bool opt_mts_dependency_order_commits;
+ulong opt_mts_dependency_order_commits;
 ulonglong opt_mts_dependency_cond_wait_timeout;
 my_bool opt_mts_dynamic_rebalance;
 double opt_mts_imbalance_threshold;

sql/mysqld.h

@@ -975,7 +975,7 @@ extern ulong opt_mts_dependency_replication;
 extern ulonglong opt_mts_dependency_size;
 extern double opt_mts_dependency_refill_threshold;
 extern ulonglong opt_mts_dependency_max_keys;
-extern my_bool opt_mts_dependency_order_commits;
+extern ulong opt_mts_dependency_order_commits;
 extern ulonglong opt_mts_dependency_cond_wait_timeout;
 extern my_bool opt_mts_dynamic_rebalance;
 extern double opt_mts_imbalance_threshold;

sql/rpl_rli.h

@@ -1137,7 +1137,7 @@ class Relay_log_info : public Rpl_info
   ulonglong mts_dependency_size= 0;
   double mts_dependency_refill_threshold= 0;
   ulonglong mts_dependency_max_keys= 0;
-  my_bool mts_dependency_order_commits= TRUE;
+  ulong mts_dependency_order_commits= 0;
   ulonglong mts_dependency_cond_wait_timeout= 0;
 
   std::deque<std::shared_ptr<Log_event_wrapper>> dep_queue;

sql/sql_class.h

@@ -143,6 +143,12 @@ enum enum_mts_dependency_replication {
   DEP_RPL_STATEMENT,
 };
 
+enum enum_mts_dependency_order_commits {
+  DEP_RPL_ORDER_NONE,
+  DEP_RPL_ORDER_DB,
+  DEP_RPL_ORDER_GLOBAL,
+};
+
 enum enum_slave_use_idempotent_for_recovery {
   SLAVE_USE_IDEMPOTENT_FOR_RECOVERY_NO,
   SLAVE_USE_IDEMPOTENT_FOR_RECOVERY_YES};

sql/sys_vars.cc

@@ -5543,6 +5543,7 @@ static Sys_var_double Sys_mts_imbalance_threshold(
        CMD_LINE(OPT_ARG), VALID_RANGE(0, 100), DEFAULT(90));
 
 static const char *dep_rpl_type_names[]= { "NONE", "TBL", "STMT", NullS };
+static const char *commit_order_type_names[]= { "NONE", "DB", "GLOBAL", NullS };
 
 static Sys_var_enum Sys_mts_dependency_replication(
        "mts_dependency_replication",
@@ -5572,11 +5573,11 @@ static Sys_var_ulonglong Sys_mts_dependency_max_keys(
        GLOBAL_VAR(opt_mts_dependency_max_keys), CMD_LINE(OPT_ARG),
        VALID_RANGE(0, ULONGLONG_MAX), DEFAULT(100000), BLOCK_SIZE(1));
 
-static Sys_var_mybool Sys_mts_dependency_order_commits(
+static Sys_var_enum Sys_mts_dependency_order_commits(
        "mts_dependency_order_commits",
-       "Commit trxs in the same order as the master (per database)",
+       "Commit trxs in the same order as the master (per database or globally)",
        GLOBAL_VAR(opt_mts_dependency_order_commits),
-       CMD_LINE(OPT_ARG), DEFAULT(TRUE));
+       CMD_LINE(OPT_ARG), commit_order_type_names, DEFAULT(DEP_RPL_ORDER_DB));
 
 static Sys_var_ulonglong Sys_mts_dependency_cond_wait_timeout(
        "mts_dependency_cond_wait_timeout",