1.背景
某一零售业后端使用了分布式中间件+MySQL数据库作为后端存储。但是因为历史问题存在两种分布式中间件,分别是Mycat和DBLE,共用一组后端MySQL实例。分片规则以及后端数据完全一致。最近碰到了一个比较有意思的场景,财务结算单来往明细和业务来往单据的关联查询。一条跨节点join查询在DBLE、Mycat的查询得到的结果不一致。究竟谁对谁错?
2.DBLE项目介绍
DBLE官方网站: https://opensource.actionsky.com 可以详细了解DBLE的背景和应用场景,本文不涉及到的细节都可在官方文档获得更细节都信息;对于刚了解到同学,可以以本文为快速入门基础
DBLE 官方项目: https://github.com/actiontech/dble 如对源码有兴趣或者需要定制的功能的可以通过源码编译安装
DBLE 下载地址: https://github.com/actiontech/dble/releases
DBLE 官方社区交流:669663113
3.环境准备
在虚拟机搭建类似架构,模拟场景,比较Mycat-DBLE在跨节点join上的异同点。
3.1测试架构
测试环境架构比较简单,DBLE与Mycat共用数据库。
3.2测试软件版本
软件名称 | 软件版本 | 端口 | 管理端口 |
---|---|---|---|
DBLE | DBLE-2.18.10.1-cb392c3-20181106093917 | 3309 | 3310 |
Mycat | Mycat-1.6-RELEASE-20161028204710 | 8066 | 9066 |
MySQL | 5.7.21-log | 3306 | 无 |
3.3表结构
结算单来往明细表
CREATE TABLE `t_bl_detail` ( `unit_num` int(11) DEFAULT NULL, `tenantid` int(11) DEFAULT NULL, `detail_num` int(11) DEFAULT NULL, `balance_date` datetime DEFAULT NULL ) ENGINE=InnoDB DEFAULT CHARSET=utf8;
业务来往单据表
CREATE TABLE `t_bl_super_detail` ( `unit_num` int(11) DEFAULT NULL, `sup_id` int(11) DEFAULT NULL, `tenantid` int(11) DEFAULT NULL, `bl_unit_num` int(11) DEFAULT NULL ) ENGINE=InnoDB DEFAULT CHARSET=utf8;
3.4分片规则配置
配置表t_bl_detail、t_bl_super_detail,使用取模算法,数据分布在db1-db4四个database中。
3.41schema配置
DBLE:
<schema name="testdb" sqlMaxLimit="100" dataNode="dn01"> <table name="t_bl_detail" rule="mod4Series" dataNode="dn01,dn02,dn03,dn04"></table> <table name="t_bl_super_detail" rule="mod4Series" dataNode="dn01,dn02,dn03,dn04"></table> </schema> <dataNode name="dn01" dataHost="group1" database="db1"></dataNode> <dataNode name="dn02" dataHost="group2" database="db2"></dataNode> <dataNode name="dn03" dataHost="group1" database="db3"></dataNode> <dataNode name="dn04" dataHost="group2" database="db4"></dataNode> <dataHost name="group1" maxCon="100" minCon="5" balance="0" switchType="-1"> <heartbeat>select user()</heartbeat> <writeHost host="mysql-w9zhkr" url="1.1.1.26:3306" user="ushard" password="VRDkwsUlq++O3HeamsDcuW/2K22si3RIhcTpAdjalbkiinNAeUUDWk11ttls1Z3PXY5TxGJToeK4LsJE3+k0Wg==" id="mysql-w9zhkr" usingDecrypt="1"></writeHost> </dataHost> <dataHost name="group2" maxCon="100" minCon="5" balance="0" switchType="-1"> <heartbeat>select user()</heartbeat> <writeHost host="mysql-ci2va0" url="1.1.1.27:3306" user="ushard" password="O/iCi0F+9ts5YR78ZcQnKzpb5GqG7xSUzJcO44yZ3BhNkT5E7aiZakmz1tbKIQylnqh20vu5Z9egXBUC3GOKow==" id="mysql-ci2va0" usingDecrypt="1"></writeHost> </dataHost>
Mycat:
<schema name="testdb" sqlMaxLimit="100" dataNode="dn01"> <table name="t1" rule="mod4Series" dataNode="dn01,dn02,dn03,dn04"></table> <table name="t2" rule="mod4Series" dataNode="dn01,dn02,dn03,dn04"></table> </schema> <dataNode name="dn01" dataHost="group1" database="db1"></dataNode> <dataNode name="dn02" dataHost="group2" database="db2"></dataNode> <dataNode name="dn03" dataHost="group1" database="db3"></dataNode> <dataNode name="dn04" dataHost="group2" database="db4"></dataNode> <dataHost name="group1" maxCon="100" minCon="5" balance="0" switchType="-1" dbType="mysql" dbDriver="native"> <heartbeat>select user()</heartbeat> <writeHost host="mysql-w9zhkr" url="1.1.1.26:3306" user="ushard" password="ushard"></writeHost> </dataHost> <dataHost name="group2" maxCon="100" minCon="5" balance="0" switchType="-1" dbType="mysql" dbDriver="native"> <heartbeat>select user()</heartbeat> <writeHost host="mysql-ci2va0" url="1.1.1.27:3306" user="ushard" password="ushard"></writeHost> </dataHost>
3.42rule配置
DBLE:
<tableRule name="mod4Series"> <rule> <columns>unit_num</columns> <algorithm>mod4DB</algorithm> </rule> </tableRule> <function name="mod4DB" class="Hash"> <property name="partitionCount">4</property> <property name="partitionLength">1</property> </function>
Mycat:
<tableRule name="mod4Series"> <rule> <columns>id</columns> <algorithm>mod4DB</algorithm> </rule> </tableRule> <function name="mod4DB" class="io.mycat.route.function.PartitionByMod"> <property name="count">4</property> </function>
4.比对开始
4.1准备测试数据
随便登录哪一台中间件写入测试数据
insert into t_bl_detail values(1,3,123443,'2019-01-01 00:00:00'); insert into t_bl_detail values(2,3,3423524,'2019-01-01 00:00:00'); insert into t_bl_detail values(3,3,245245,'2019-01-01 00:00:00'); insert into t_bl_detail values(4,4,356356,'2019-01-01 00:00:00'); insert into t_bl_super_detail values(1,10342,3,2); insert into t_bl_super_detail values(2,12355,3,2); insert into t_bl_super_detail values(3,62542,3,3); insert into t_bl_super_detail values(4,74235,4,1);
4.2执行跨节点join查询
select m.unit_num, n.sup_id, n.tenantid, m.detail_num, n.bl_unit_num, m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid = n.tenantid where n.tenantid = 3 and m.unit_num = n.bl_unit_num;
在通过中间件之前,现在MySQL中执行一遍看下结果,作为预期结果供后续案例使用。
分别通过DBLE、mycat执行跨节点join语句
可看到相同的查询语句,DBLE执行结果符合预期,Mycat执行结果缺失。数据差异在于DBLE查询结果相较于Mycat多了跨节点的结果。虽然Mycat执行跨节点join不报错,但是查询结果却和预期不一致。
4.3执行计划
只从结果上判断并没有办法知道是什么原因导致了Mycat结果缺失,查看查询计划,比较两者差异。
4.31DBLE
通过DBLE的执行计划可看出,DBLE内部分别对结算明细表、业务单据表做了各自的数据查询,将查询结果在中间层做了merge。最后获得跨节点join的结果
- DBLE的执行计划
mysql> explain select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num; +-----------------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | DATA_NODE | TYPE | SQL/REF | +-----------------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | dn01_0 | BASE SQL | select `m`.`unit_num`,`m`.`detail_num`,`m`.`balance_date`,`m`.`tenantid` from `t_bl_detail` `m` where m.tenantid = 3 ORDER BY `m`.`tenantid` ASC,`m`.`unit_num` ASC | | dn02_0 | BASE SQL | select `m`.`unit_num`,`m`.`detail_num`,`m`.`balance_date`,`m`.`tenantid` from `t_bl_detail` `m` where m.tenantid = 3 ORDER BY `m`.`tenantid` ASC,`m`.`unit_num` ASC | | dn03_0 | BASE SQL | select `m`.`unit_num`,`m`.`detail_num`,`m`.`balance_date`,`m`.`tenantid` from `t_bl_detail` `m` where m.tenantid = 3 ORDER BY `m`.`tenantid` ASC,`m`.`unit_num` ASC | | dn04_0 | BASE SQL | select `m`.`unit_num`,`m`.`detail_num`,`m`.`balance_date`,`m`.`tenantid` from `t_bl_detail` `m` where m.tenantid = 3 ORDER BY `m`.`tenantid` ASC,`m`.`unit_num` ASC | | merge_1 | MERGE | dn01_0; dn02_0; dn03_0; dn04_0 | | shuffle_field_1 | SHUFFLE_FIELD | merge_1 | | dn01_1 | BASE SQL | select `n`.`sup_id`,`n`.`tenantid`,`n`.`bl_unit_num` from `t_bl_super_detail` `n` where n.tenantid = 3 ORDER BY `n`.`tenantid` ASC,`n`.`bl_unit_num` ASC | | dn02_1 | BASE SQL | select `n`.`sup_id`,`n`.`tenantid`,`n`.`bl_unit_num` from `t_bl_super_detail` `n` where n.tenantid = 3 ORDER BY `n`.`tenantid` ASC,`n`.`bl_unit_num` ASC | | dn03_1 | BASE SQL | select `n`.`sup_id`,`n`.`tenantid`,`n`.`bl_unit_num` from `t_bl_super_detail` `n` where n.tenantid = 3 ORDER BY `n`.`tenantid` ASC,`n`.`bl_unit_num` ASC | | dn04_1 | BASE SQL | select `n`.`sup_id`,`n`.`tenantid`,`n`.`bl_unit_num` from `t_bl_super_detail` `n` where n.tenantid = 3 ORDER BY `n`.`tenantid` ASC,`n`.`bl_unit_num` ASC | | merge_2 | MERGE | dn01_1; dn02_1; dn03_1; dn04_1 | | shuffle_field_3 | SHUFFLE_FIELD | merge_2 | | join_1 | JOIN | shuffle_field_1; shuffle_field_3 | | shuffle_field_2 | SHUFFLE_FIELD | join_1 | +-----------------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------+ 14 rows in set (0.00 sec)
4.32Mycat
Mycat对于跨节点join的处理则相对暴力,直接将查询语句下发到各个节点,最后将结果进行汇总,如果表连接涉及到跨节点。则跨节点的数据无法进行join。
- Mycat的执行计划
mysql> explain select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num; +-----------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | DATA_NODE | SQL | +-----------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ | dn01 | select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num | | dn02 | select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num | | dn03 | select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num | | dn04 | select m.unit_num,n.sup_id,n.tenantid,m.detail_num,n.bl_unit_num,m.balance_date from t_bl_detail m join t_bl_super_detail n on m.tenantid=n.tenantid where n.tenantid = 3 and m.unit_num=n.bl_unit_num | +-----------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ 4 rows in set (0.00 sec)
5.总结
Mycat是一款非常优秀的分布式中间件,但是在某些细节方面处理的不尽人意。在跨节点关联查询场景下,Mycat采取的策略是直接将语句透传到各个节点上,将获取到的结果整合后返回,得到的结果集和预期结果有出入,缺失了跨节点关联的数据。
DBLE处理跨节点的关联查询是先获取到关联需要的数据,提取到中间件进行融合,得到关联查询的结果并返回。得到的结果集符合预期,与MySQL执行结果一致。可见DBLE在跨节点关联查询方面做了优化,能够提供准确的查询结果。
来源:oschina
链接:https://my.oschina.net/u/3883885/blog/3007553