TiKV & TiFlash accelerate complex business queries丨TiFlash application practice

Complex business queries are a test for traditional relational databases,而通过 TiKV row storage and TiFlash The combination of column storage can be well dealt with.本文根据 TUG User Biancheng Yuanyuan is here TiDB The sharing and sorting of the Shijiazhuang Station of the Community Technology Exchange,详细介绍了 TiKV & TiFlash Principles and practical solutions for accelerating complex business queries.

背景

In Internet companies or traditional companies CRM 系统中,One of the most frequently used functions is customer filtering.通过不同的角度、维度、A combination of tags to box select customers,for subsequent business operations.

This is undoubtedly the traditional relational database,Or the architecture of relational database plus column storage database is a test,There are mainly the following pain points：

• Traditional relational databases cannot optimize and speed up queries by adding indexes,业务无法正常开展;

• Column-stored databases need to put filter-related data into the column database,And need to do real-time synchronization of data;

• Data cannot be read from the database level,It is often necessary to read data from a column database and then to a relational database for data merging and output,Performance is not optimistic.

TiDB 数据库的 TiKV 和 TiFlash The combination of theoretically solves several of the above pain points.

TiKV 行存 与 TiFlash 列存混合使用

TiDB 中 query 执行的示意图,可以看到在 TiDB 中一个 query 的执行会被分成两部分,一部分在 TiDB 执行,一部分下推给存储层（ TiFlash/TiKV ）执行.

混用原理

• TiDB 的行列混合并不是传统设计上的行存列存二选一,而是 TiDB 可以在同一张表同时拥有行存和列存,且两者永远保持数据强一致（而非最终一致）.

• 多表查询分别使用不同的引擎 TiKV 或 TiFlash.

• TiFlash 支持 MPP 模式的查询执行,即在计算中引入跨节点的数据交换（data shuffle 过程）.

混用优化

标签系统高级筛选

通过标签（从宽表里不确定字段）和窄表特定字段组合查询客户并分页

Read from TiKV

SELECT 
/*+ READ_FROM_STORAGE(tikv[b], tikv[c],tikv[d]) */
	a.*,
	b.CUST_NAME,b.CERT_TYPE,b.CERT_NUM,b.CUST_TYPE,b.SEX,b.AGE,b.BIRTH_DT,
	c.ORG_ID,c.ORG_NAME,
	d.ASSET,d.ASSET_MON_AVG 
FROM
	(
	SELECT /*+ READ_FROM_STORAGE(tikv[m],tikv[n]) */
		m.cust_id 
	FROM
		m_cust_label m
		RIGHT JOIN m_cust_org n ON m.CUST_ID = n.CUST_ID 
	WHERE
		m.cat1 IN ( 516, 710, 230,3301 ) 
		AND n.ORG_ID IN ( '133','8716', '7162') ORDER BY	n.cust_id ASC 	LIMIT 100 
	) a
	LEFT JOIN m_cust_main b ON a.cust_id = b.cust_id
	LEFT JOIN m_cust_org c ON a.cust_id = c.cust_id
	LEFT JOIN m_cust_data d ON a.cust_id = d.cust_id ;
	

4G,2c 虚拟机 300 万数据,首次执行 48 s 二次执行 0.7s

Read From TiKV & TiFlash

 
 SELECT 
/*+ READ_FROM_STORAGE(tikv[b], tikv[c],tikv[d]) */
	a.*,
	b.CUST_NAME,b.CERT_TYPE,b.CERT_NUM,b.CUST_TYPE,b.SEX,b.AGE,b.BIRTH_DT,
	c.ORG_ID,c.ORG_NAME,
	d.ASSET,d.ASSET_MON_AVG 
FROM
	(
	SELECT /*+ READ_FROM_STORAGE(tiflash[m],tikv[n]) */
		m.cust_id 
	FROM
		m_cust_label m
		RIGHT JOIN m_cust_org n ON m.CUST_ID = n.CUST_ID 
	WHERE
		m.cat1 IN ( 516, 710, 230,3301 ) 
		AND n.ORG_ID IN ( '133','8716', '7162') ORDER BY	n.cust_id ASC 	LIMIT 100 
	) a
	LEFT JOIN m_cust_main b ON a.cust_id = b.cust_id
	LEFT JOIN m_cust_org c ON a.cust_id = c.cust_id
	LEFT JOIN m_cust_data d ON a.cust_id = d.cust_id 

4G,2c 虚拟机 300 万数据,首次执行 3s 二次执行 0.3s

TiFlash & MPP

Controls whether to select MPP 模式

变量 tidb_allow_mpp 控制 TiDB 能否选择 MPP 模式执行查询.变量 tidb_enforce_mpp Controls whether to ignore optimizer cost estimates,强制使用 TiFlash 的 MPP 模式执行查询.

The results corresponding to all the values ​​of these two variables are as follows：

	tidb_allow_mpp=off	tidb_allow_mpp=on（默认）
tidb_enforce_mpp=off（默认）	不使用 MPP 模式.	The optimizer is chosen based on the cost estimate.（默认）
tidb_enforce_mpp=on	不使用 MPP 模式.	TiDB 无视代价估算,选择 MPP 模式.

set @@session.tidb_allow_mpp=1;
set @@session.tidb_enforce_mpp=1;

SELECT 
/*+ READ_FROM_STORAGE(tikv[b], tikv[c],tikv[d]) */
	a.*,
	b.CUST_NAME,b.CERT_TYPE,b.CERT_NUM,b.CUST_TYPE,b.SEX,b.AGE,b.BIRTH_DT,
	c.ORG_ID,c.ORG_NAME,
	d.ASSET,d.ASSET_MON_AVG 
FROM
	(
	SELECT /*+ READ_FROM_STORAGE(tiflash[m],tiflash[n]) */
		m.cust_id 
	FROM
		m_cust_label m
		RIGHT JOIN m_cust_org n ON m.CUST_ID = n.CUST_ID 
	WHERE
		m.cat1 IN ( 516, 710, 230,3301 ) 
		AND n.ORG_ID IN ( '133','8716', '7162') ORDER BY	n.cust_id ASC 	LIMIT 100 
	) a
	LEFT JOIN m_cust_main b ON a.cust_id = b.cust_id
	LEFT JOIN m_cust_org c ON a.cust_id = c.cust_id
	LEFT JOIN m_cust_data d ON a.cust_id = d.cust_id 

使用 MPP 模式来执行查询后基本秒开,4G 2c 虚拟机 300 万数据,首次执行 1s 二次执行 0.15s

2.4 SPM 固定执行计划

CREATE GLOBAL|SESSION  BINDING for	<BindableStmt > USING <BindableStmt2>
SHOW GLOBAL|SESSION BINDINGS ; -- 查看绑定计划explain format = 'verbose' <BindableStmt2>;
show warnings; -- 通过执行 show warnings 了解该 SQL 语句使用了哪一条 binding

固定特定查询走 TiFlash 列存查询.

标签下价值机构排名

根据选中的属性（多值）

使用这些值最多的排名前 3 的机构,并统计出总额

执行计划

table:c 走 TiFlash ;table:a, table:b 走 TiKV ,同时使用了列存和行存的优势.

总结

使用 TiKV 和 TiFlash 可以加速复杂查询,下面简单增加了使用使用场景.

组件	适用场景说明
TiKV	检索条件固定,且有索引
TiFlash	检索条件不固定,无法加索引
TiKV + TiFlash	部分表检索条件不固定,部分表有索引

如果有描述不当的地方欢迎评论指正！

谢谢 PingCAP 社区的大力支持！