SHOW SLAVE STATUS shows real error

Summary: In parallel replication SHOW SLAVE STATUS doesn't show the actual error. It actually creates the real error (Slave_reporting_capability::va_report) and set m_last_error right after setting coordinator error so we just need to reverse the order of the two and show the real error in coordinator. Accessing the last error is protected by err_lock and also fallback to previous message in case the error number is 0 and can't be trusted, just in case.

Reviewed By: hermanlee, Pushapgl

Differential Revision: D28925760

fbshipit-source-id: f351b8ca3e6

sql/rpl_rli_pdb.cc

@@ -1692,33 +1692,60 @@ void Slave_worker::do_report(loglevel level, int err_code, const char *msg,
 
   gtid_next->to_string(global_sid_map, buff_gtid, true);
 
+  if (is_group_replication_applier_channel) {
+    snprintf(buff_coord, sizeof(buff_coord),
+             "Worker %u failed executing transaction '%s'", internal_id,
+             buff_gtid);
+  } else {
+    snprintf(buff_coord, sizeof(buff_coord),
+             "Worker %u failed executing transaction '%s' at "
+             "master log %s, end_log_pos %llu",
+             internal_id, buff_gtid, log_name, log_pos);
+  }
+
+  /*
+    Error reporting by the worker. The worker updates its error fields as well
+    as reports the error in the log.
+  */
+  this->va_report(level, err_code, buff_coord, msg, args);
+
   if (level == ERROR_LEVEL && (!has_temporary_error(thd, err_code) ||
                                thd->get_transaction()->cannot_safely_rollback(
                                    Transaction_ctx::SESSION))) {
     char coordinator_errmsg[MAX_SLAVE_ERRMSG];
+    const char *err_msg = nullptr;
+
+    mysql_mutex_lock(&err_lock);
+    if (m_last_error.number) {
+      /* We know for sure there is an valid error */
+      err_msg = m_last_error.message;
+    } else {
+      /* Fallback to generic error message just in case */
+      err_msg = buff_coord;
+    }
 
     if (is_group_replication_applier_channel) {
       snprintf(coordinator_errmsg, MAX_SLAVE_ERRMSG,
                "Coordinator stopped because there were error(s) in the "
                "worker(s). "
-               "The most recent failure being: Worker %u failed executing "
-               "transaction '%s'. See error log and/or "
+               "The most recent failure being: %s; "
+               "See error log and/or "
                "performance_schema.replication_applier_status_by_worker "
                "table for "
                "more details about this failure or others, if any.",
-               internal_id, buff_gtid);
+               err_msg);
     } else {
       snprintf(coordinator_errmsg, MAX_SLAVE_ERRMSG,
                "Coordinator stopped because there were error(s) in the "
                "worker(s). "
-               "The most recent failure being: Worker %u failed executing "
-               "transaction '%s' at master log %s, end_log_pos %llu. "
+               "The most recent failure being: %s; "
                "See error log and/or "
                "performance_schema.replication_applier_status_by_worker "
                "table for "
                "more details about this failure or others, if any.",
-               internal_id, buff_gtid, log_name, log_pos);
+               err_msg);
     }
+    mysql_mutex_unlock(&err_lock);
 
     /*
       We want to update the errors in coordinator as well as worker.
@@ -1730,23 +1757,6 @@ void Slave_worker::do_report(loglevel level, int err_code, const char *msg,
     */
     c_rli->fill_coord_err_buf(level, err_code, coordinator_errmsg);
   }
-
-  if (is_group_replication_applier_channel) {
-    snprintf(buff_coord, sizeof(buff_coord),
-             "Worker %u failed executing transaction '%s'", internal_id,
-             buff_gtid);
-  } else {
-    snprintf(buff_coord, sizeof(buff_coord),
-             "Worker %u failed executing transaction '%s' at "
-             "master log %s, end_log_pos %llu",
-             internal_id, buff_gtid, log_name, log_pos);
-  }
-
-  /*
-    Error reporting by the worker. The worker updates its error fields as well
-    as reports the error in the log.
-  */
-  this->va_report(level, err_code, buff_coord, msg, args);
 }
 
 #ifndef DBUG_OFF