Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-15740

Galera does not recover prepared XA-transactions correctly




      A stripped down version of galera.galera_gcache_recover:

      mysql-test/suite/galera/t/foo.test (no need for a .result file):
      --source include/galera_cluster.inc
      --connection node_1
      INSERT INTO t1 VALUES (1);
      --connection node_2
      --let $wait_condition = SELECT COUNT(*) > 0 FROM t1;
      --source include/wait_condition.inc
      # If we don't sleep here, the tail command below prints:
      #   WSREP: Recovered position: ...:1
      # If we do sleep then we recover to seqno=2:
      #   WSREP: Recovered position: ...:2
      #--sleep 1
      --source include/kill_galera.inc
      --connection node_2
      --let $galera_wsrep_recover_server_id=2
      --source suite/galera/include/galera_wsrep_recover.inc
      --echo Recovered node2
      --exec tail -1 $MYSQL_TMP_DIR/galera_wsrep_recover.log
      --echo sleep 1000...
      --sleep 1000

      The problem with that test is that (without the sleep) the contents of the t1 table on node2 goes out of sync with the recorded wsrep xid. Wsrep xid seqno=2 is lost and when it recovers later to 1 it would try to re-apply the changes behind seqno=2 from another node, BUT that would cause a duplicate key error because the row is already in the table t1.

      The reason for this is that the mini-transaction which writes wsrep xid commits, but the redo log is never flushed up to that LSN before the node is killed. This bug is not present in MySQL because there the log is flushed properly during commit:

      trx_sys_update_wsrep_checkpoint() writes seqno=2 using mtr=0x7fffdf9b4ff8
      mtr_commit(mtr=0x7fffdf9b4ff8) end_lsn=1631956
          trx_flush_log_if_needed()  // this is innodb_flush_log_at_trx_commit handling

      In MariaDB this code path is cancelled at trx_commit_complete_for_mysql():

      1d0f70c2f894 (Michael Widenius 2243) trx_commit_complete_for_mysql(
      1d0f70c2f894 (Michael Widenius 2244) /*==========================*/
      068c61978e3a (Michael Widenius 2245)     trx_t*    trx)    /*!< in/out: transaction */
      1d0f70c2f894 (Michael Widenius 2246) {
      2e814d4702d7 (Jan Lindström    2247)     if (trx->id != 0
      2e814d4702d7 (Jan Lindström    2248)         || !trx->must_flush_log_later
      36e81a23c567 (Kristian Nielsen 2249)         || (srv_flush_log_at_trx_commit == 1 && trx->active_commit_ordered)) {
      1d0f70c2f894 (Michael Widenius 2250) 
      068c61978e3a (Michael Widenius 2251)         return;
      1d0f70c2f894 (Michael Widenius 2252)     }
      1d0f70c2f894 (Michael Widenius 2253) 
      068c61978e3a (Michael Widenius 2254)     trx_flush_log_if_needed(trx->commit_lsn, trx);

      The return; on line 2251 is executed and trx_flush_log_if_needed() is never called. This is because line 2249 evaluates to true - both srv_flush_log_at_trx_commit and trx->active_commit_ordered are 1.

      commit 36e81a23c56
      Parent: 5ae598390aa
      Author:     Kristian Nielsen <knielsen@knielsen-hq.org>
      AuthorDate: Mon Aug 7 12:38:47 2017 +0200
      Commit:     Kristian Nielsen <knielsen@knielsen-hq.org>
      CommitDate: Mon Aug 7 18:23:55 2017 +0200
          MDEV-11937: InnoDB flushes redo log too often
          Problem was introduced with the InnoDB 5.7 merge, the code related to
          avoiding extra fsync at the end of commit when binlog is enabled. The
          MariaDB method for this was removed, but the replacement MySQL method
          based on thd_get_durability_property() is not functional in MariaDB.
          This commit reverts the offending parts of the merge and adds a test
          case, to fix the problem for InnoDB. But other storage engines are
          likely to have a similar problem.
      -           || thd_requested_durability(trx->mysql_thd)
      -              == HA_IGNORE_DURABILITY) {
      +           || (srv_flush_log_at_trx_commit == 1 && trx->active_commit_ordered)) {

      This looks particularly strange because changing innodb_flush_log_at_trx_commit from 1 to 2 "fixes" the bug due to the condition srv_flush_log_at_trx_commit == 1 not being true anymore. But 1 is supposed to be more durable and slow compared to 2. Now, with the current code 2 is more durable!


        Issue Links



              teemu.ollakka Teemu Ollakka
              Dimov Vasil (Inactive)
              2 Vote for this issue
              11 Start watching this issue



                Git Integration

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