Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-31482

Lock wait timeout with INSERT-SELECT, autoinc, and statement-based replication

    XMLWordPrintable

Details

    Description

      Debugging a lock wait timeout in optimistic parallel replication, found a problem with the way lock wait reporting is done wrt. auto-increment.

      Comments on thd_rpl_deadlock_check() say:

      "This call need only receive reports about waits for locks that will remain
      until the holding transaction commits. InnoDB auto-increment locks,
      for example, are released earlier, and so need not be reported."

      The assumption was that if the locks are released at the end of the statement, they will not be able to cause a deadlock involving a transaction waiting for an earlier transaction to commit.

      However, this assumption turns out to be false, as can be seen by considering 3 transactions T1 T2 T3:

      1. T1 is waiting for T3 on an autoinc lock
      2. T2 is waiting for T1 to commit
      3. T3 is waiting on a normal row lock held by T2

      Here there is a wait cycle, and we need to deadlock kill T3 as it is blocking T1. However, as autoinc lock waits are currently not reported to the replication layer, this deadlock kill can be missed, and parallel replication will hang until an innodb lock wait timeout occurs.

      Here is a test case that demonstrates the problem:

      --source include/have_binlog_format_statement.inc
      --source include/have_innodb.inc
      --source include/master-slave.inc
       
      --echo MDEV-31482: Lock wait timeout with INSERT-SELECT, autoing, and statement-based replication
       
      # The scenario is transactions T1, T2, T3:
      #
      #   T1 is waiting for T3 on an autoinc lock
      #   T2 is waiting for T1 to commit
      #   T3 is waiting on a normal row lock held by T2
      #
      # This caused a hang until innodb_lock_wait_timeout, because autoinc
      # locks were not reported to the in-order parallel replication, so T3
      # was not deadlock killed.
       
      --let $lock_wait_timeout=20
       
      --let $rpl_connection_name= slave2
      --let $rpl_server_number= 2
      --source include/rpl_connect.inc
       
      --let $rpl_connection_name= slave3
      --let $rpl_server_number= 2
      --source include/rpl_connect.inc
       
      --connection master
      ALTER TABLE mysql.gtid_slave_pos ENGINE=InnoDB;
       
      # A table as destination for INSERT-SELECT
      CREATE TABLE t1 (a INT PRIMARY KEY AUTO_INCREMENT, b INT, c INT, INDEX (c)) ENGINE=InnoDB;
      INSERT INTO t1 (b,c) VALUES (0, 1), (0, 1), (0, 2), (0,3), (0, 5), (0, 7), (0, 8);
       
      # A table as source for INSERT-SELECT.
      CREATE TABLE t2 (a INT PRIMARY KEY, b INT) ENGINE=InnoDB;
      INSERT INTO t2 VALUES (10,1), (20,2), (30,3), (40,4), (50,5);
       
      # A table to help order slave worker threads to setup the desired scenario.
      CREATE TABLE t3 (a VARCHAR(20) PRIMARY KEY, b INT) ENGINE=InnoDB;
      INSERT INTO t3 VALUES ('row for T1', 0), ('row for T2', 0), ('row for T3', 0);
      --source include/save_master_gtid.inc
       
      --connection slave
      --source include/sync_with_master_gtid.inc
      --source include/stop_slave.inc
      --let $save_innodb_lock_wait_timeout= `SELECT @@global.innodb_lock_wait_timeout`
      --let $save_slave_parallel_threads= `SELECT @@global.slave_parallel_threads`
      --let $save_slave_parallel_mode= `SELECT @@global.slave_parallel_mode`
      set @@global.slave_parallel_threads= 3;
      set @@global.slave_parallel_mode= OPTIMISTIC;
      eval set @@global.innodb_lock_wait_timeout= $lock_wait_timeout;
       
      --connection master
      # Transaction T1.
      BEGIN;
      UPDATE t3 SET b=b+1 where a="row for T1";
      INSERT INTO t1(b, c) SELECT 1, t2.b FROM t2 WHERE a=10;
      COMMIT;
       
      # Transaction T2.
      DELETE FROM t1 WHERE c >= 4 and c < 6;
       
      # Transaction T3.
      BEGIN;
      UPDATE t3 SET b=b+1 where a="row for T3";
      INSERT INTO t1(b, c) SELECT 3, t2.b FROM t2 WHERE a >= 20 AND a <= 40;
      COMMIT;
       
      --source include/save_master_gtid.inc
       
      --connection slave1
      # Temporarily block T1 to create the scheduling that triggers the bug.
      BEGIN;
      SELECT * FROM t3 WHERE a="row for T1" FOR UPDATE;
       
      --connection slave2
      # Temporarily block T3 from starting (so T2 can reach commit).
      BEGIN;
      SELECT * FROM t3 WHERE a="row for T3" FOR UPDATE;
       
      --connection slave3
      # This critical step blocks T3 after it has inserted its first row,
      # and thus taken the auto-increment lock, but before it has reached
      # the point where it gets a row lock wait on T2. Even though
      # auto-increment lock waits were not reported due to the bug,
      # transitive lock waits (T1 waits on autoinc of T3 which waits on row
      # on T2) _were_ reported as T1 waiting on T2, and thus a deadlock kill
      # happened and the bug was not triggered.
      BEGIN;
      DELETE FROM t2 WHERE a=30;
       
      --connection slave
      --source include/start_slave.inc
       
      # First let T2 complete until it is waiting for T1 to commit.
      --let $wait_condition= SELECT count(*)=1 FROM information_schema.processlist WHERE state='Waiting for prior transaction to commit' and command LIKE 'Slave_worker';
      --source include/wait_condition.inc
       
      # Then let T3 reach the point where it has obtained the autoinc lock,
      # but it is not yet waiting for a row lock held by T2.
      --connection slave2
      ROLLBACK;
      --let $wait_condition= SELECT count(*)=1 FROM information_schema.processlist WHERE state='Sending data' and info LIKE 'INSERT INTO t1(b, c) SELECT 3, t2.b%' and time_ms > 500 and command LIKE 'Slave_worker';
      --source include/wait_condition.inc
       
      # Now let T1 continue, while T3 is holding the autoinc lock but before
      # it is waiting for T2. Wait a short while to give the hang a chance to
      # happen; T1 needs to get to request the autoinc lock before we let T3
      # continue. (There's a small chance the sleep will be too small, which will
      # let the test occasionally pass on non-fixed server).
      --connection slave1
      ROLLBACK;
      --sleep 0.5
       
      # Now let T3 continue; the bug was that this lead to an undetected
      # deadlock that remained until innodb lock wait timeout.
      --connection slave3
      ROLLBACK;
       
      --connection slave
      --let $slave_timeout= `SELECT $lock_wait_timeout/2`
      --source include/sync_with_master_gtid.inc
      SELECT * FROM t1 ORDER BY a;
      SELECT * FROM t2 ORDER BY a;
      SELECT * FROM t3 ORDER BY a;
       
      # Cleanup.
      --connection master
      CALL mtr.add_suppression("Unsafe statement written to the binary log using statement format");
      DROP TABLE t1, t2, t3;
       
      --connection slave
      --source include/stop_slave.inc
      eval SET @@global.slave_parallel_threads= $save_slave_parallel_threads;
      eval SET @@global.slave_parallel_mode= $save_slave_parallel_mode;
      eval SET @@global.innodb_lock_wait_timeout= $save_innodb_lock_wait_timeout;
      --source include/start_slave.inc
      SELECT @@GLOBAL.innodb_autoinc_lock_mode;
      --source include/rpl_end.inc
      

      I think the fix is simply to remove the condition in DeadlockChecker::search() that skips calling thd_rpl_deadlock_check() for autoinc locks. Only issue is to consider if this can cause performance regression from excessive deadlock kills and transaction retries. This needs to be analyzed, but I'm hopeful it will not be much of an issue. In many (most?) cases no autoincrement lock will be taken in replication, depending on innodb_autoinc_lock_mode, and in mixed mode some unsafe autoinc statements (eg. INSERT-SELECT as in this test case) will be logged in row format and skip the autoinc locks.

      Attachments

        Activity

          People

            knielsen Kristian Nielsen
            knielsen Kristian Nielsen
            Votes:
            0 Vote for this issue
            Watchers:
            5 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved:

              Git Integration

                Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.