when running sysbench oltp test ,XA transaction can't restore correctly after restarting dble

[funnyAnt]

• dble version：
  2.19.01.0
• preconditions ：
  • 1、 sudo yum install sysbench
  • 2、edit $SYSBENCH_SCRIPT_HOME/lua-tests/db/oltp.lua ,add db_query("set xa=1") below db_query(begin_query)
  • 3、edit $SYSBENCH_SCRIPT_HOME/oltp.sh， add --mysql-ignore-errors=1205,2013,1213,1003,1062

• 4、edit $SYSBENCH_SCRIPT_HOME/conf.sh, configure the correct database information

• configs：

schema.xml

<schema name="dbtest">
		<table dataNode="dn1,dn2" name="sbtest1" rule="mod2" />
   </schema>

	<dataNode name="dn1" dataHost="localhost1" database="sysbench_0" />
	<dataNode name="dn2" dataHost="localhost1" database="sysbench_1" />

rule.xml

<tableRule name="mod2">
		<rule>
			<columns>id</columns>
			<algorithm>hashmod2</algorithm>
		</rule>
	</tableRule>
	<function name="hashmod2" class="Hash">
		<property name="partitionCount">2</property>
		<property name="partitionLength">1</property>
	</function>

server.xml

• steps：
  step1. execute shell sh oltp.sh on the installed sysbench machine
  step2. restart dble when sysbench is running
  step3. stop sysbench
  step4. execute xa recover on the mysql which is used by dble
• expect result：
  1. empty result, dble can restore all not completed xa transactions.
• real result：
  1. There are still some transactions that have not been released
• supplements：
  1.

/label ~BUG

[funnyAnt]

reason:

• One of the xa logs without normal restore
  {"id":"'Dble_Server.1.194'","state":"12","participants":[{"host":"10.45.61.34","port":"32305","p_state":"2","expires":0,"schema":"sysbench_0"},{"host":"10.45.61.34","port":"32305","p_state":" 3","expires":0,"schema":"sysbench_1"}]}

the transaction with p_state":"3" state will be rolled back, the transaction with "p_state":"2" will be ignored, but the transaction with "p_state":" 2" is already 3 on mysql , just because dble has no time to write correct state to disk.

[PanternBao]

In my opinion, state of participants are written to file when dble has received mysql response. Now, when you restart dble, dble miss mysql response partly,so state of part participants are not entirely written to file.