Fix for BUG#1870

    "CHANGE MASTER makes SQL thread restart from coordinates of I/O thread".
    So, in CHANGE MASTER:
    when it seems reasonable that the user did not want to discontinue
    its replication (i.e. when he is not specifying host or port or master_log_file
    or master_log_pos; this will be documented), set the coordinates of the
    I/O thread to those of the SQL thread. This way, the SQL thread will see
    no discontinuity in the relay log (i.e. will skip no events), because
    the I/O thread will fill the brand new relay log with the events which
    are just after the position where the SQL thread had stopped
    (before CHANGE MASTER was issued).
    And a new test for this bug.


mysql-test/r/rpl_loaddata.result:
  Now, after CHANGE MASTER the coordinates of the I/O thread are the last ones of the SQL thread, so result update.
sql/sql_repl.cc:
  Fix for BUG#1870
  "CHANGE MASTER makes SQL thread restart from coordinates of I/O thread".
  So, in CHANGE MASTER:
  when it seems reasonable that the user did not want to discontinue
  its replication (i.e. when he is not specifying host or port or master_log_file
  or master_log_pos; this will be documented), set the coordinates of the
  I/O thread to those of the SQL thread. This way, the SQL thread will see
  no discontinuity in the relay log (i.e. will skip no events), because
  the I/O thread will fill the brand new relay log with the events which
  are just after the position where the SQL thread had stopped
  (before CHANGE MASTER was issued).

1 files changed, 45 insertions, 4 deletions

diff --git a/sql/sql_repl.cc b/sql/sql_repl.cc
index 10581431c72..c95cdc1b04e 100644
--- a/sql/sql_repl.cc
+++ b/sql/sql_repl.cc
@@ -853,8 +853,8 @@ void kill_zombie_dump_threads(uint32 slave_server_id)
 int change_master(THD* thd, MASTER_INFO* mi)
 {
   int thread_mask;
-  const char* errmsg=0;
-  bool need_relay_log_purge=1;
+  const char* errmsg= 0;
+  bool need_relay_log_purge= 1;
   DBUG_ENTER("change_master");
 
   lock_slave_threads(mi);
@@ -928,6 +928,36 @@ int change_master(THD* thd, MASTER_INFO* mi)
     mi->rli.relay_log_pos=lex_mi->relay_log_pos;
   }
 
+  /*
+    If user did specify neither host nor port nor any log name nor any log
+    pos, i.e. he specified only user/password/master_connect_retry, he probably
+    wants replication to resume from where it had left, i.e. from the
+    coordinates of the **SQL** thread (imagine the case where the I/O is ahead
+    of the SQL; restarting from the coordinates of the I/O would lose some
+    events which is probably unwanted when you are just doing minor changes
+    like changing master_connect_retry).
+    A side-effect is that if only the I/O thread was started, this thread may
+    restart from ''/4 after the CHANGE MASTER. That's a minor problem (it is a
+    much more unlikely situation than the one we are fixing here).
+    Note: coordinates of the SQL thread must be read here, before the
+    'if (need_relay_log_purge)' block which resets them.
+  */
+  if (!lex_mi->host && !lex_mi->port &&
+      !lex_mi->log_file_name && !lex_mi->pos &&
+      need_relay_log_purge)
+   {
+     /*
+       Sometimes mi->rli.master_log_pos == 0 (it happens when the SQL thread is
+       not initialized), so we use a max().
+       What happens to mi->rli.master_log_pos during the initialization stages
+       of replication is not 100% clear, so we guard against problems using
+       max().
+      */
+     mi->master_log_pos = max(BIN_LOG_HEADER_SIZE, mi->rli.master_log_pos);
+     strmake(mi->master_log_name,mi->rli.master_log_name,
+             sizeof(mi->master_log_name)-1);
+  }
+
   flush_master_info(mi);
   if (need_relay_log_purge)
   {
@@ -959,10 +989,21 @@ int change_master(THD* thd, MASTER_INFO* mi)
     }
   }
   DBUG_PRINT("info", ("master_log_pos: %d", (ulong) mi->master_log_pos));
-  /* If changing RELAY_LOG_FILE or RELAY_LOG_POS, this will be nonsense: */
+
+  /*
+    Coordinates in rli were spoilt by the 'if (need_relay_log_purge)' block,
+    so restore them to good values. If we left them to ''/0, that would work;
+    but that would fail in the case of 2 successive CHANGE MASTER (without a
+    START SLAVE in between): because first one would set the coords in mi to
+    the good values of those in rli, the set those in rli to ''/0, then
+    second CHANGE MASTER would set the coords in mi to those of rli, i.e. to
+    ''/0: we have lost all copies of the original good coordinates.
+    That's why we always save good coords in rli.
+  */
   mi->rli.master_log_pos = mi->master_log_pos;
   strmake(mi->rli.master_log_name,mi->master_log_name,
-	  sizeof(mi->rli.master_log_name)-1);
+          sizeof(mi->rli.master_log_name)-1);
+
   if (!mi->rli.master_log_name[0]) // uninitialized case
     mi->rli.master_log_pos=0;