BUG#29288: myisam transactions replicated to a transactional

slave leaves slave unstable

Problem: when replicating from non-transactional to
transactional engine with autocommit off, no BEGIN/COMMIT
is written to the binlog. When the slave replicates, it
will start a transaction that never ends.

Fix: Force autocommit=on on slave by always replicating
autocommit=1 from the master.

1 files changed, 22 insertions, 6 deletions

diff --git a/sql/log_event.cc b/sql/log_event.cc
index 0cda724b698..fb6a5230fda 100644
--- a/sql/log_event.cc
+++ b/sql/log_event.cc
@@ -2389,13 +2389,29 @@ Query_log_event::Query_log_event(THD* thd_arg, const char* query_arg,
     charset_database_number= thd_arg->variables.collation_database->number;
   
   /*
-    If we don't use flags2 for anything else than options contained in
-    thd_arg->options, it would be more efficient to flags2=thd_arg->options
-    (OPTIONS_WRITTEN_TO_BIN_LOG would be used only at reading time).
-    But it's likely that we don't want to use 32 bits for 3 bits; in the future
-    we will probably want to reclaim the 29 bits. So we need the &.
+    We only replicate over the bits of flags2 that we need: the rest
+    are masked out by "& OPTIONS_WRITTEN_TO_BINLOG".
+
+    We also force AUTOCOMMIT=1.  Rationale (cf. BUG#29288): After
+    fixing BUG#26395, we always write BEGIN and COMMIT around all
+    transactions (even single statements in autocommit mode).  This is
+    so that replication from non-transactional to transactional table
+    and error recovery from XA to non-XA table should work as
+    expected.  The BEGIN/COMMIT are added in log.cc. However, there is
+    one exception: MyISAM bypasses log.cc and writes directly to the
+    binlog.  So if autocommit is off, master has MyISAM, and slave has
+    a transactional engine, then the slave will just see one long
+    never-ending transaction.  The only way to bypass explicit
+    BEGIN/COMMIT in the binlog is by using a non-transactional table.
+    So setting AUTOCOMMIT=1 will make this work as expected.
+
+    Note: explicitly replicate AUTOCOMMIT=1 from master. We do not
+    assume AUTOCOMMIT=1 on slave; the slave still reads the state of
+    the autocommit flag as written by the master to the binlog. This
+    behavior may change after WL#4162 has been implemented.
   */
-  flags2= (uint32) (thd_arg->options & OPTIONS_WRITTEN_TO_BIN_LOG);
+  flags2= (uint32) (thd_arg->options &
+                    (OPTIONS_WRITTEN_TO_BIN_LOG & ~OPTION_NOT_AUTOCOMMIT));
   DBUG_ASSERT(thd_arg->variables.character_set_client->number < 256*256);
   DBUG_ASSERT(thd_arg->variables.collation_connection->number < 256*256);
   DBUG_ASSERT(thd_arg->variables.collation_server->number < 256*256);