refs #181

- Galera related errors in mysql slave applying will now cause slave to abort
codership · Sep 6, 2015 · a35c12f · a35c12f
1 parent 3b80cdd
commit a35c12f
Showing 1 changed file with 47 additions and 18 deletions.
diff --git a/sql/rpl_slave.cc b/sql/rpl_slave.cc
@@ -3858,6 +3858,14 @@ apply_event_and_update_pos(Log_event** ptr_ev, THD* thd, Relay_log_info* rli)
       DBUG_RETURN(SLAVE_APPLY_EVENT_AND_UPDATE_POS_OK);
 
     exec_res= ev->apply_event(rli);
+#ifdef WITH_WSREP
+    if (exec_res && thd->wsrep_conflict_state != NO_CONFLICT)
+    {
+      WSREP_DEBUG("SQL apply failed, res %d conflict state: %d", 
+                 exec_res, thd->wsrep_conflict_state);
+      rli->abort_slave = 1;
+    }
+#endif
 
     if (!exec_res && (ev->worker != rli))
     {
@@ -4365,6 +4373,7 @@ static int exec_relay_log_event(THD* thd, Relay_log_info* rli)
         Hence deferred events wont be deleted here.
         They will be deleted in Deferred_log_events::rewind() funciton.
     */
+    WSREP_DEBUG("apply_event_and_update_pos result: %d", exec_res);
     if (*ptr_ev)
     {
       DBUG_ASSERT(*ptr_ev == ev); // event remains to belong to Coordinator
@@ -4408,6 +4417,12 @@ static int exec_relay_log_event(THD* thd, Relay_log_info* rli)
       DBUG_RETURN(1);
     }
 
+#ifdef WITH_WSREP
+    mysql_mutex_lock(&thd->LOCK_wsrep_thd);
+    if (thd->wsrep_conflict_state == NO_CONFLICT)
+    {
+      mysql_mutex_unlock(&thd->LOCK_wsrep_thd);
+#endif /* WITH_WSREP */
     if (slave_trans_retries)
     {
       int UNINIT_VAR(temp_err);
@@ -4486,6 +4501,10 @@ static int exec_relay_log_event(THD* thd, Relay_log_info* rli)
                             rli->trans_retries));
       }
     }
+#ifdef WITH_WSREP
+    } else mysql_mutex_unlock(&thd->LOCK_wsrep_thd);
+#endif /* WITH_WSREP */
+
     if (exec_res)
       delete ev;
     DBUG_RETURN(exec_res);
@@ -6218,6 +6237,13 @@ log '%s' at position %s, relay log '%s' position: %s", rli->get_rpl_log_name(),
 
     if (exec_relay_log_event(thd,rli))
     {
+#ifdef WITH_WSREP
+      if (thd->wsrep_conflict_state != NO_CONFLICT)
+      {
+        wsrep_node_dropped = 1;
+        rli->abort_slave   = 1;
+      }
+#endif /* WITH_WSREP */
       DBUG_PRINT("info", ("exec_relay_log_event() failed"));
       // do not scare the user if SQL thread was simply killed or stopped
       if (!sql_slave_killed(thd,rli))
@@ -6289,8 +6315,8 @@ llstr(rli->get_group_master_log_pos(), llbuff));
 #ifdef WITH_WSREP
         if (WSREP_ON && last_errno == ER_UNKNOWN_COM_ERROR)
         {
-	  wsrep_node_dropped= TRUE;
-	}
+          wsrep_node_dropped= TRUE;
+        }
 #endif /* WITH_WSREP */
       }
       goto err;
@@ -6379,22 +6405,25 @@ llstr(rli->get_group_master_log_pos(), llbuff));
   /* if slave stopped due to node going non primary, we set global flag to
      trigger automatic restart of slave when node joins back to cluster
   */
-   if (wsrep_node_dropped && wsrep_restart_slave)
-   {
-     if (wsrep_ready)
-     {
-       WSREP_INFO("Slave error due to node temporarily non-primary"
-		  "SQL slave will continue");
-       wsrep_node_dropped= FALSE;
-       mysql_mutex_unlock(&rli->run_lock);
-       goto wsrep_restart_point;
-     } else {
-       WSREP_INFO("Slave error due to node going non-primary");
-       WSREP_INFO("wsrep_restart_slave was set and therefore slave will be "
-		  "automatically restarted when node joins back to cluster");
-       wsrep_restart_slave_activated= TRUE;
-     }
-   }
+  if (wsrep_node_dropped && wsrep_restart_slave)
+  {
+    if (wsrep_ready)
+    {
+      WSREP_INFO("Slave error due to node temporarily non-primary"
+                 "SQL slave will continue");
+      wsrep_node_dropped= FALSE;
+      mysql_mutex_unlock(&rli->run_lock);
+      WSREP_DEBUG("wsrep_conflict_state now: %d", thd->wsrep_conflict_state);
+      WSREP_INFO("slave restart: %d", thd->wsrep_conflict_state);
+      thd->wsrep_conflict_state = NO_CONFLICT;
+      goto wsrep_restart_point;
+    } else {
+      WSREP_INFO("Slave error due to node going non-primary");
+      WSREP_INFO("wsrep_restart_slave was set and therefore slave will be "
+                 "automatically restarted when node joins back to cluster");
+      wsrep_restart_slave_activated= TRUE;
+    }
+  }
 #endif /* WITH_WSREP */
  /*
   Note: the order of the broadcast and unlock calls below (first broadcast, then unlock)