Replication: new code to not modify in-memory log positions until the COMMIT

is executed, even if the transaction spans on >=2 relay logs (bug #53).
New variable relay_log_purge =0|1
New test to verify bug #53


sql/log.cc:
  Now we purge a relay log only when we are sure we won't need it,
  i.e. we have executed the final query (if autocommit=1) or the COMMIT.
sql/log_event.cc:
  Better tracking of the relay log's name and position
  lastly executed, even if we are in a transaction which spans on
  2 or more relay logs.
sql/mysql_priv.h:
  new option relay_log_purge (the user can now decide himself
  if he wants his relay logs to be automatically purged or not,
  we don't make unsafe guesses like before)
sql/mysqld.cc:
  new option --innodb (replaces --skip-innodb).
  Useful for the test suite : we have skip-innodb in mysql-test-run,
  but we can ('-opt.info' file) choose to start the server with
  InnoDB for this test only.
  New option --bdb
sql/repl_failsafe.cc:
  Better tracking of the relay log's name and position
  lastly executed, even if we are in a transaction which spans on
  2 or more relay logs.
sql/set_var.cc:
  new variable relay_log_purge
sql/slave.cc:
  Better tracking of the relay log's name and position
  lastly executed, even if we are in a transaction which spans on
  2 or more relay logs.
  Now we purge a relay log only when we are sure we won't need it,
  i.e. we have executed the final query (if autocommit=1) or the COMMIT
sql/slave.h:
  Better tracking of the relay log's name and position
  lastly executed, even if we are in a transaction which spans on
  2 or more relay logs.
sql/sql_class.h:
  prototypes change
sql/sql_parse.cc:
  removed thd argument (was not used in the function's body)
sql/sql_repl.cc:
  Better tracking of the relay log's name and position
  lastly executed, even if we are in a transaction which spans on
  2 or more relay logs.
  Turn relay_log_purge silently off when someone does CHANGE
  MASTER TO RELAY_LOG_*

1 files changed, 61 insertions, 0 deletions

diff --git a/mysql-test/t/rpl_relayrotate.test b/mysql-test/t/rpl_relayrotate.test
new file mode 100644
index 00000000000..4c330b8c9a2
--- /dev/null
+++ b/mysql-test/t/rpl_relayrotate.test
@@ -0,0 +1,61 @@
+# When the relay log gets rotated while the I/O thread
+# is reading a transaction, the transaction spans on two or more
+# relay logs. If STOP SLAVE occurs while the SQL thread is
+# executing a part of the transaction in the non-first relay logs,
+# we test if START SLAVE will resume in the beginning of the
+# transaction (i.e., step back to the first relay log)
+  
+# The slave is started with max_binlog_size=16384 bytes,
+# to force many rotations (approximately 30 rotations)
+
+# If the master or slave does not support InnoDB, this test will pass
+
+source include/master-slave.inc;
+connection slave;
+stop slave;
+connection master;
+create table t1 (a int) type=innodb;
+let $1=8000;
+disable_query_log;
+begin;
+while ($1)
+{
+# eval means expand $ expressions
+ eval insert into t1 values( $1 );
+ dec $1;
+}
+commit;
+# This will generate a 500kB master's binlog,
+# which corresponds to 30 slave's relay logs.
+enable_query_log;
+save_master_pos;
+connection slave;
+reset slave;
+start slave;
+# We wait 1 sec for the SQL thread to be somewhere in
+# the middle of the transaction, hopefully not in
+# the first relay log, and hopefully before the COMMIT.
+# Usually it stops when the SQL thread is around the 15th relay log.
+# We cannot use MASTER_POS_WAIT() as master's position
+# increases only when the slave executes the COMMIT.
+system sleep 1;
+stop slave;
+# We suppose the SQL thread stopped before COMMIT.
+# If so the transaction was rolled back
+# and the table is now empty.
+# Now restart
+start slave;
+# And see if the table contains '8000'
+# which proves that the transaction restarted at
+# the right place.
+# We must wait for the transaction to commit before
+# reading, MASTER_POS_WAIT() will do it for sure
+# (the only statement with position>=3000 is COMMIT).
+# Older versions of MySQL would hang forever in MASTER_POS_WAIT
+# because COMMIT was said to be position 0 in the master's log (bug).
+# Detect this with timeout.
+select master_pos_wait('master-bin.001',3000,120)=-1;
+select * from t1 where a=8000;
+# Note that the simple fact to have less than around 30 slave's binlogs
+# (the slave is started with --log-slave-updates) is already
+# a proof that the transaction was not properly resumed.