1# ==== Purpose ==== 2# 3# Tests that an autocommitted XA transaction where the master crashes 4# just before writing the XID log event is executed correctly. The 5# master rolls back, so the slave should not execute statement. 6# 7# This test was previously part of rpl_ndb_transaction.test 8# 9# 10# ==== Method ==== 11# 12# We want master to be alive so that it can replicate the statement to 13# the slave. So in the test case, we must not crash the 14# master. Instead, we fake the crash by just not writing the XID event 15# to the binlog. This is done by the @@debug='d,do_not_write_xid' 16# flag. This, in turn, requires us to do 'source 17# include/have_debug.inc' 18# 19# So, unlike if the master had crashed, the master *will* execute the 20# statement. But the slave should not execute it. Hence, after the 21# test is executed, the expected result on master is a table with one 22# row, and on slave a table with no rows. 23# 24# To simulate the slave correctly, we wait until everything up to but 25# not including the XID is replicated. This has to be done with 26# include/sync_slave_io_with_master.inc, not sync_slave_with_master, 27# since the latter waits until the slave *SQL* thread has caught up 28# with the master's position, which it will never do. 29# 30# 31# ==== Related bugs ==== 32# 33# BUG#26395: if crash during autocommit update to transactional table on master, slave fails 34 35--source include/not_group_replication_plugin.inc 36source include/have_innodb.inc; 37# have_debug is needed since we use the @@debug variable on master 38source include/have_debug.inc; 39source include/master-slave.inc; 40 41# test adapts simulation of incomplete transaction that MTS does not tolerate 42# when is stopped. So it reacts with an error whereas the single-threaded is fine. 43-- source include/not_mts_slave_parallel_workers.inc 44 45 46--echo ==== Initialize ==== 47 48--echo [on master] 49--connection master 50 51CREATE TABLE tinnodb (a INT) ENGINE = INNODB; 52SHOW CREATE TABLE tinnodb; 53 54# do_not_write_xid stops the master from writing an XID event. 55set @old_debug= @@debug; 56set @@debug= 'd,do_not_write_xid'; 57 58 59--echo ==== Test ==== 60 61# Save the position up to which the slave SQL thread should execute. 62save_master_pos; 63 64# Execute query and check that the row made it to the table. 65INSERT INTO tinnodb VALUES (1); 66SELECT * FROM tinnodb ORDER BY a; 67 68# Sync slave's IO thread. 69--echo [on slave] 70# After the implementation of the transaction boundary parser, 71# this sync cannot use GTIDs or else sync will timeout because only 72# GTIDs of fully replicated transactions are in the 73# Retrieved_Gtid_Set of the slave and, as the last event of the 74# transaction never arrives, syncing with GTIDs won't work. 75--let ignore_gtids_on_sync= 1 76source include/sync_slave_io_with_master.inc; 77 78# Sync slave's SQL thread. 79sync_with_master 0; 80 81--echo ==== Verify results on slave ==== 82 83source include/stop_slave.inc; 84 85--let $status_items=Slave_IO_State,Last_SQL_Error,Last_IO_Error 86--source include/show_slave_status.inc 87 88--let $assert_text= Assert that the slave table has no rows 89--let $assert_cond= `SELECT COUNT(*) = 0 FROM tinnodb` 90--source include/assert.inc 91 92--echo ==== Clean up ==== 93 94# Easiest to clean up master and slave separately, without 95# replication, since master and slave are out of sync. 96 97--echo [on master] 98connection master; 99DROP TABLE tinnodb; 100set @@debug= @old_debug; 101 102--echo [on slave] 103connection slave; 104DROP TABLE tinnodb; 105 106# Warning: do not add more tests here. The binlog is in a bad state. 107--let $rpl_only_running_threads= 1 108--source include/rpl_end.inc 109