Bug#18436 (RBR: Replication to partition engine triggers assertion on slave side):

Partial fix for resolving the problem. Swapping contents of record[0]
and record[1] since this is what some storage engines expect.


sql/handler.cc:
  Adding assertion to get early failure.
sql/log_event.cc:
  Correcting code so that new record is passed in record[0] and
  old record is passed in record[1] when calling update_row().
mysql-test/r/rpl_row_basic_8partition.result:
  New BitKeeper file ``mysql-test/r/rpl_row_basic_8partition.result''
mysql-test/t/rpl_row_basic_8partition.test:
  New BitKeeper file ``mysql-test/t/rpl_row_basic_8partition.test''

1 files changed, 16 insertions, 10 deletions

diff --git a/sql/log_event.cc b/sql/log_event.cc
index d7ad9501ffd..a5f9b28de32 100644
--- a/sql/log_event.cc
+++ b/sql/log_event.cc
@@ -6402,7 +6402,7 @@ static int find_and_fetch_row(TABLE *table, byte *key)
       table->record[0] if the engine allows it.  We first compute a
       row reference using the position() member function (it will be
       stored in table->file->ref) and the use rnd_pos() to position
-      the "cursor" at the correct row.
+      the "cursor" (i.e., record[0] in this case) at the correct row.
     */
     table->file->position(table->record[0]);
     DBUG_RETURN(table->file->rnd_pos(table->record[0], table->file->ref));
@@ -6797,19 +6797,25 @@ int Update_rows_log_event::do_exec_row(TABLE *table)
     return error;
 
   /*
-    This is only a precaution to make sure that the call to
-    ha_update_row is using record[1].
-
-    If this is not needed/required, then we could use m_after_image in
-    that call instead.
+    We have to ensure that the new record (i.e., the after image) is
+    in record[0] and the old record (i.e., the before image) is in
+    record[1].  This since some storage engines require this (for
+    example, the partition engine).
+
+    Since find_and_fetch_row() puts the fetched record (i.e., the old
+    record) in record[0], we have to move it out of the way and into
+    record[1]. After that, we can put the new record (i.e., the after
+    image) into record[0].
   */
-  bmove_align(table->record[1], m_after_image,(size_t) table->s->reclength);
+  bmove_align(table->record[1], table->record[0], table->s->reclength);
+  bmove_align(table->record[0], m_after_image, table->s->reclength);
 
   /*
-    Now we should have the right row to update.  The record that has
-    been fetched is guaranteed to be in record[0], so we use that.
+    Now we should have the right row to update.  The old row (the one
+    we're looking for) has to be in record[1] and the new row has to
+    be in record[0] for all storage engines to work correctly.
   */
-  error= table->file->ha_update_row(table->record[0], table->record[1]);
+  error= table->file->ha_update_row(table->record[1], table->record[0]);
 
   return error;
 }