1.故障现象
20191113-22:32 datax全量同步t_shop_info表到 eorder所在的实例,t_shop_info表有两个唯一约束。总数据量不超过1w行,同步完成后MGR从库复制线程死锁。导致同步停止,影响业务。
从库processlist 截图
2.故障复现
2.1 集群环境
MGR集群ip
|
端口
|
---|---|
10.204.55.205 | 3320 |
10.204.55.206 | 3320 |
10.204.55.207 | 3320 |
2.2 复现方法1 ,使用datax全量同步。
集群主要参数配置
SET GLOBAL transaction_isolation = 'READ-COMMITTED'
SET GLOBAL slave_parallel_type = 'LOGICAL_CLOCK';
SET GLOBAL slave_parallel_workers = 16;
SET GLOBAL slave_preserve_commit_order = ON;
用datax启动 replace into方式全量同步t_shop_info表到MGR测试集群。
结果:任务启动第二次从库复制线程必定hangs。
用datax启动 DUPLICATE KEY UPDATE 方式全量同步t_shop_info表到MGR测试集群
结果:任务启动5次都没有出现从库复制线程hangs。
2.3 复现方法2,使用 jmeter 并发40线程并行执行sql。
集群主要参数
SET GLOBAL slave_parallel_type = 'LOGICAL_CLOCK';
SET GLOBAL slave_parallel_workers = 16;
SET GLOBAL slave_preserve_commit_order = ON;
表结构
CREATE TABLE `replaceinto_deadlocktest` ( `fid` BIGINT (20) NOT NULL AUTO_INCREMENT, `fname` VARCHAR (10) DEFAULT NULL , `fdesc` VARCHAR (10) DEFAULT NULL , PRIMARY KEY (`fid`), UNIQUE KEY `fname` (`fname`) ) ENGINE=INNODB AUTO_INCREMENT=10754 DEFAULT CHARSET=utf8mb4 |
#随机数 replace into REPLACE INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES (@fid:= '${__Random(1,200000,)}' , @fname:=CONCAT( 'b' ,@fid), 'a2' ); #递增id replace into REPLACE INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES ( '${__intSum(${__counter(TRUE,)},1)}' , 'b${__intSum(${__counter(TRUE,)},1)}' , 'a2' ) #随机id DUPLICATE update INSERT INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES (@fid:= '${__Random(1,200000,)}' ,@fname:=CONCAT( 'b' ,@fid), 'a2' ) ON DUPLICATE KEY UPDATE fid=@fid,fname=@fname,fdesc= 'a2' ; #递增id DUPLICATE update INSERT INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES ( '${__intSum(${__counter(TRUE,)},1)}' , 'b${__intSum(${__counter(TRUE,)},1)}' , 'a2' ) ON DUPLICATE KEY UPDATE fid= '${__intSum(${__counter(TRUE,)},1)}' ,fname= 'b${__intSum(${__counter(TRUE,)},1)}' ,fdesc= 'a2' ; |
结果:
随机数replace into 与 递增id replace into 都会出现从库复制线程hangs。并且jmeter会有一定的比例报线程死锁。
随机数DUPLICATE update 与 递增id DUPLICATE update 都没出现从库复制线程hangs。
jmeter死锁日志 Response code:40001 1213 Response message:com.mysql.jdbc.exceptions.jdbc4.MySQLTransactionRollbackException: Deadlock found when trying to get lock; try restarting transaction mysql死锁日志 TRANSACTION 10244303573, ACTIVE 0 sec inserting mysql TABLES IN USE 1, locked 1 LOCK WAIT 5 LOCK struct(s), HEAP size 1136, 4 ROW LOCK(s), UNDO LOG entries 2 MySQL thread id 39582, OS thread handle 140372828473088, QUERY id 11662118 10.201.3.1 WORK UPDATE REPLACE INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES (@fid:= '49' , @fname:=CONCAT( 'b' ,@fid), 'a2' ) *** (1) WAITING FOR THIS LOCK TO BE GRANTED: RECORD LOCKS SPACE id 734 page NO 4 n bits 120 INDEX fname of TABLE `dayutest`.`replaceinto_deadlocktest` trx id 10244303573 lock_mode X waiting *** (2) TRANSACTION : TRANSACTION 10244303574, ACTIVE 0 sec inserting mysql TABLES IN USE 1, locked 1 4 LOCK struct(s), HEAP size 1136, 3 ROW LOCK(s), UNDO LOG entries 2 MySQL thread id 39560, OS thread handle 140349632673536, QUERY id 11662119 10.201.3.1 WORK UPDATE REPLACE INTO replaceinto_deadlocktest (fid,fname,fdesc) VALUES (@fid:= '5' , @fname:=CONCAT( 'b' ,@fid), 'a2' ) *** (2) HOLDS THE LOCK(S): RECORD LOCKS SPACE id 734 page NO 4 n bits 120 INDEX fname of TABLE `dayutest`.`replaceinto_deadlocktest` trx id 10244303574 lock_mode X LOCKS rec but NOT gap *** (2) WAITING FOR THIS LOCK TO BE GRANTED: RECORD LOCKS SPACE id 734 page NO 4 n bits 120 INDEX fname of TABLE `dayutest`.`replaceinto_deadlocktest` trx id 10244303574 lock_mode X waiting *** WE ROLL BACK TRANSACTION (2) |
2.4 复现方法3,jmeter+(pt-osc,gh-ost)
jmeter并发写入带唯一索引的表,并且使用pt-osc 和gh-ost 分别修改表结构
结果:jmeter并发写入时,同时pt-osc出现从库复制线程hangs。
jmeter并发写入时,同时gh-ost没有出现从库复制线程hangs。
pt-online-schema-change --user=work --password=111111 --host=10.204.55.205 --port 3320 --alter="add index idx_fdesc(fdesc);" D=dayutest,t=replaceinto_deadlocktest --chunk-size=20 --nodrop-old-table --chunk-time=0.5 --max-lag=2 --check-interval=2 --charset=utf8 --critical-load="Threads_running=64" --max-load="Threads_running=32" --print --execute gh-ost --host='10.204.55.205' --port 3320 --user='work' --password='111111' --database='dayutest' --table='replaceinto_deadlocktest' --alter="add index idx_fdesc(fdesc);" --allow-on-master --verbose --initially-drop-ghost-table --throttle-flag-file /tmp/1.log --allow-master-master --ok-to-drop-table --approve-renamed-columns --max-lag-millis=2000 --chunk-size=2000 --max-load=Threads_running=20 --switch-to-rbr --execute
3.测试结论
MGR开启并行16线程复制,并且对带唯一约束表做replace into操作很高的几率导致MGR从库hangs。
replace into 场景包括(pt-osc,datax全量同步)。
后续改进措施,使用gh-ost替代pt-osc,datax全量同步调整为ON DUPLICATE KEY UPDATE 插入。