Fixing calculation seconds behind master when IO thread is caught up

Summary:
When IO has finished downloading all binlogs but SQL thread is still
executing events a HB event from the master used to update the last master
timestamp on the slave which lead to SBM being reset. This change updates last
master timestamp from a HB event only when the SQL thread has caught up to the
IO thread.

Reference Patch: 257d463

Reviewed By: Pushapgl

Differential Revision: D27885260

fbshipit-source-id: 3642d8fca88

sql/rpl_slave.cc

@@ -8316,8 +8316,19 @@ QUEUE_EVENT_RESULT queue_event(Master_info *mi, const char *buf,
         is always monotonically increasing
       */
       mysql_mutex_lock(&rli->data_lock);
-      rli->set_last_master_timestamp(hb.common_header->when.tv_sec,
-                                     hb.common_header->when.tv_sec * 1000);
+      auto io_thread_file = mi->get_master_log_name();
+      auto io_thread_pos = mi->get_master_log_pos();
+      auto sql_thread_file = mi->rli->get_group_master_log_name();
+      auto sql_thread_pos = mi->rli->get_group_master_log_pos();
+
+      // find out if the SQL thread has caught up with the IO thread by
+      // comparing their coordinates
+      bool caughtup = (sql_thread_pos == io_thread_pos) &&
+                      (!strcmp(sql_thread_file, io_thread_file));
+      // case: update last master ts only if we've caughtup
+      if (caughtup)
+        rli->set_last_master_timestamp(hb.common_header->when.tv_sec,
+                                       hb.common_header->when.tv_sec * 1000);
       mysql_mutex_unlock(&rli->data_lock);
 
       /*