摘要:通常跑批加工场景下,都是大数量做关联操作,通常不建议使用索引。有些时候因为计划误判导致使用索引的可能会导致严重的性能问题。本文从一个典型的索引导致性能的场景重发,剖析此类问题的特征,定位方法和解决方法

本文分享自华为云社区《GaussDB(DWS)性能调优:indexscan导致的性能问题识别与优化 #【玩转PB级数仓GaussDB(DWS)】》,作者: 譡里个檔 。

通常跑批加工场景下,都是大数量做关联操作,通常不建议使用索引。有些时候因为计划误判导致使用索引的可能会导致严重的性能问题。本文从一个典型的索引导致性能的场景重发,剖析此类问题的特征,定位方法和解决方法。

1)在某局点POC测试时发现某SQL语句比较慢,原始SQL如下

WITH
/**
etl_116583_7960703_994644
**/
LOADABLE as (select "boq_rel_type_id","to_pu_id","to_version","cycle_id",
"part_offset_flag","to_boq_id","descr","from_contract_id",
"from_version","from_pu_id","ss_id","to_contract_id",
"from_boq_id","enable_flag","last_update_date"
from (SELECT /*+ PARALLEL(4)*/
BOQ_REL.FROM_BOQ_ID,
BOQ_REL.TO_BOQ_ID,
BOQ_REL.FROM_PU_ID,
BOQ_REL.TO_PU_ID,
BOQ_REL.PART_OFFSET_FLAG,
BOQ_REL.DESCR,
BOQ_REL.SS_ID,
BOQ_REL.CYCLE_ID,
NVL(BOQ_REL.FROM_VERSION, 'SNULL') FROM_VERSION,
NVL(BOQ_REL.TO_VERSION, 'SNULL') TO_VERSION,
BOQ_REL.LAST_UPDATE_DATE,
FROM_CON.CONTRACT_ID AS FROM_CONTRACT_ID,
TO_CON.CONTRACT_ID AS TO_CONTRACT_ID,
CLA.CLASS_ID AS BOQ_REL_TYPE_ID,
BOQ_REL.ENABLE_FLAG
FROM (SELECT A.FROM_BOQ_ID,
A.TO_BOQ_ID,
A.FROM_PU_ID,
A.TO_PU_ID,
A.FROM_CONTRACT_NUMBER,
A.TO_CONTRACT_NUMBER,
A.BOQ_REL_TYPE_CODE,
A.PART_OFFSET_FLAG,
A.DESCR,
A.SS_ID,
A.FROM_VERSION,
A.TO_VERSION,
A.LAST_UPDATE_DATE,
A.CYCLE_ID,
A.ENABLE_FLAG,
DECODE(A.SS_ID, 2820, 2600, A.SS_ID) SS_ID_TMP,
ROW_NUMBER() OVER(PARTITION BY FROM_BOQ_ID, TO_BOQ_ID, FROM_PU_ID, TO_PU_ID, FROM_CONTRACT_NUMBER, TO_CONTRACT_NUMBER, BOQ_REL_TYPE_CODE, FROM_VERSION, TO_VERSION
ORDER BY DECODE(A.SS_ID, 2820, 1, 2600, 2, 3)) RN
FROM LDB_MD_BOQ_REL A) BOQ_REL,
(SELECT CONTRACT_ID,
HW_CONTRACT_NUM,
SS_ID
FROM DWI_MD_CONTRACT
WHERE CONTRACT_ID IS NOT NULL
AND END_TIME = TO_DATE('4712-12-31', 'YYYY-MM-DD')) FROM_CON,
(SELECT CONTRACT_ID,
HW_CONTRACT_NUM,
SS_ID
FROM DWI_MD_CONTRACT
WHERE CONTRACT_ID IS NOT NULL
AND END_TIME = TO_DATE('4712-12-31', 'YYYY-MM-DD')) TO_CON,
(SELECT CLASS_ID,
CODE,
CLASS_TYPE_ID,
SS_ID
FROM DWI_MD_CLASS
WHERE CLASS_TYPE_ID = 193) CLA
WHERE BOQ_REL.RN = 1
AND BOQ_REL.FROM_CONTRACT_NUMBER = FROM_CON.HW_CONTRACT_NUM
AND BOQ_REL.SS_ID = FROM_CON.SS_ID
AND BOQ_REL.TO_CONTRACT_NUMBER = TO_CON.HW_CONTRACT_NUM
AND BOQ_REL.SS_ID = TO_CON.SS_ID
AND BOQ_REL.BOQ_REL_TYPE_CODE = CLA.CODE
AND BOQ_REL.SS_ID_TMP = CLA.SS_ID
) t
),
BEFORE_TARGET as (select "from_contract_id","from_pu_id","ss_id","from_boq_id","from_version","to_version",
"crt_cycle_id","to_pu_id","to_boq_id","del_flag","last_upd_cycle_id","last_update_date",
"descr","enable_flag","crt_job_instance_id","dq_improve_flag","upd_job_instance_id",
"to_contract_id","part_offset_flag","boq_rel_type_id"
from (SELECT /*+PARALLEL(4)*/
FROM_BOQ_ID,
TO_BOQ_ID,
FROM_PU_ID,
TO_PU_ID,
FROM_CONTRACT_ID,
TO_CONTRACT_ID,
BOQ_REL_TYPE_ID,
PART_OFFSET_FLAG,
DESCR,
SS_ID,
CRT_CYCLE_ID,
LAST_UPD_CYCLE_ID,
DEL_FLAG,
DQ_IMPROVE_FLAG,
CRT_JOB_INSTANCE_ID,
UPD_JOB_INSTANCE_ID,
NVL(FROM_VERSION, 'SNULL') FROM_VERSION,
NVL(TO_VERSION, 'SNULL') TO_VERSION,
LAST_UPDATE_DATE,
ENABLE_FLAG
FROM DWI_MD_BOQ_REL
) t
),
CDC as (select LOADABLE."ss_id",LOADABLE."from_version",LOADABLE."from_boq_id",
LOADABLE."part_offset_flag",LOADABLE."from_pu_id",
case when BEFORE_TARGET.BOQ_REL_TYPE_ID is null and BEFORE_TARGET.FROM_BOQ_ID is null
and BEFORE_TARGET.FROM_CONTRACT_ID is null and BEFORE_TARGET.FROM_PU_ID is null
and BEFORE_TARGET.FROM_VERSION is null
and BEFORE_TARGET.TO_BOQ_ID is null and BEFORE_TARGET.TO_CONTRACT_ID is null
and BEFORE_TARGET.TO_PU_ID is null and BEFORE_TARGET.TO_VERSION is null
then 1
else 3
end as "change_code",
LOADABLE."to_version",LOADABLE."boq_rel_type_id",
LOADABLE."from_contract_id",LOADABLE."to_contract_id",
LOADABLE."descr",LOADABLE."last_update_date",
LOADABLE."to_pu_id",LOADABLE."enable_flag",LOADABLE."cycle_id",
LOADABLE."to_boq_id"
from LOADABLE
left join BEFORE_TARGET on LOADABLE.BOQ_REL_TYPE_ID = BEFORE_TARGET.BOQ_REL_TYPE_ID
and LOADABLE.FROM_BOQ_ID = BEFORE_TARGET.FROM_BOQ_ID and LOADABLE.FROM_CONTRACT_ID = BEFORE_TARGET.FROM_CONTRACT_ID
and LOADABLE.FROM_PU_ID = BEFORE_TARGET.FROM_PU_ID and LOADABLE.FROM_VERSION = BEFORE_TARGET.FROM_VERSION
and LOADABLE.TO_BOQ_ID = BEFORE_TARGET.TO_BOQ_ID and LOADABLE.TO_CONTRACT_ID = BEFORE_TARGET.TO_CONTRACT_ID
and LOADABLE.TO_PU_ID = BEFORE_TARGET.TO_PU_ID and LOADABLE.TO_VERSION = BEFORE_TARGET.TO_VERSION
),
TFM_FILTER_DATA_TARGET_OUTPUT_U as (select CDC."to_pu_id",CDC."boq_rel_type_id",CDC."ss_id",
Current_Timestamp() as "dw_last_update_date",CDC."to_version",
CDC."from_version",20230104000000 as "last_upd_cycle_id",
CDC."from_contract_id",CDC."last_update_date",CDC."descr",
'N' as "del_flag",CDC."from_boq_id",CDC."to_boq_id",
CDC."enable_flag",CDC."from_pu_id",-1 as "upd_job_instance_id",
'N' as "dq_improve_flag",CDC."to_contract_id",
CDC."part_offset_flag"
from CDC where CDC.change_code=3
)
update DWI_MD_BOQ_REL TARGET_U
set "dq_improve_flag" = TFM_FILTER_DATA_TARGET_OUTPUT_U."dq_improve_flag",
"dw_last_update_date" = TFM_FILTER_DATA_TARGET_OUTPUT_U."dw_last_update_date",
"upd_job_instance_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."upd_job_instance_id",
"descr" = TFM_FILTER_DATA_TARGET_OUTPUT_U."descr",
"part_offset_flag" = TFM_FILTER_DATA_TARGET_OUTPUT_U."part_offset_flag",
"last_update_date" = TFM_FILTER_DATA_TARGET_OUTPUT_U."last_update_date",
"del_flag" = TFM_FILTER_DATA_TARGET_OUTPUT_U."del_flag",
"last_upd_cycle_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."last_upd_cycle_id",
"enable_flag" = TFM_FILTER_DATA_TARGET_OUTPUT_U."enable_flag",
"ss_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."ss_id"
from TFM_FILTER_DATA_TARGET_OUTPUT_U
where TARGET_U."boq_rel_type_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."boq_rel_type_id"
and TARGET_U."to_version" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_version"
and TARGET_U."to_version" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_version"
and TARGET_U."to_pu_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_pu_id"
and TARGET_U."to_pu_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_pu_id"
and TARGET_U."to_contract_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_contract_id"
and TARGET_U."to_contract_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_contract_id"
and TARGET_U."to_boq_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_boq_id"
and TARGET_U."to_boq_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."to_boq_id"
and TARGET_U."from_version" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_version"
and TARGET_U."from_version" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_version"
and TARGET_U."from_pu_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_pu_id"
and TARGET_U."from_pu_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_pu_id"
and TARGET_U."from_contract_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_contract_id"
and TARGET_U."from_contract_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_contract_id"
and TARGET_U."from_boq_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_boq_id"
and TARGET_U."from_boq_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."from_boq_id"
and TARGET_U."boq_rel_type_id" = TFM_FILTER_DATA_TARGET_OUTPUT_U."boq_rel_type_id"
;

2)查询此query的topSQL信息的warning字段,发现SQL自诊断信息中有索引相关告警信息。

3)查询此query的topSQL信息(如下图),分析历史执行信息,发现id=20的CStore Index Scan算子的耗时为90796.980ms,SQL执行总时长137135.658ms。CStore Index Scan算子的耗时占比为 66%

4)找到原始SQL语句,对查询语句中出现的表dwimd.dwi_md_contract 进行hint,强制其走顺序扫描,避免走indexscan(全量语句见附件)

5)对语句进行explain verbose,查看计划,发现计划符合预期(即表dwimd.dwi_md_contract走tablescan,对于列存表计划上显式为CStore Scan)

6)对语句执行EXPLAIN ANALYZE操作(即实际执行语句),查看实际执行时间如下,发现SQL语句性能提升近10倍。全量的执行信息见附件

点击关注,第一时间了解华为云新鲜技术~

GaussDB(DWS)性能调优:indexscan导致的性能问题识别与优化的更多相关文章

  1. [Spark性能调优] 第一章:性能调优的本质、Spark资源使用原理和调优要点分析

    本課主題 大数据性能调优的本质 Spark 性能调优要点分析 Spark 资源使用原理流程 Spark 资源调优最佳实战 Spark 更高性能的算子 引言 我们谈大数据性能调优,到底在谈什么,它的本质 ...

  2. JVM性能调优2:JVM性能调优参数整理

    序号 参数名 说明 JDK 默认值 使用过 1 JVM执行模式 2 -client-server 设置该JVM运行与Client 或者Server Hotspot模式,这两种模式从本质上来说是在JVM ...

  3. spark 性能调优(一) 性能调优的本质、spark资源使用原理、调优要点分析

    转载:http://www.cnblogs.com/jcchoiling/p/6440709.html 一.大数据性能调优的本质 编程的时候发现一个惊人的规律,软件是不存在的!所有编程高手级别的人无论 ...

  4. JVM性能调优(4) —— 性能调优工具

    前序文章: JVM性能调优(1) -- JVM内存模型和类加载运行机制 JVM性能调优(2) -- 垃圾回收器和回收策略 JVM性能调优(3) -- 内存分配和垃圾回收调优 一.JDK工具 先来看看有 ...

  5. MySql(九):MySQL性能调优——Schema设计的性能优化

    一.高效的模型设计 先了解下数据库设计的三大范式 第一范式:要求有主键,并且要求每一个字段原子性不可再分 第二范式:要求所有非主键字段完全依赖主键,不能产生部分依赖 第三范式:所有非主键字段和主键字段 ...

  6. 图灵学院Java架构师-VIP-【性能调优-Mysql索引数据结构详解与索引优化】

    最近报名了图灵学院的架构专题的付费课程,没有赶上6月份开课,中途加入的.错过了多线程的直播课程,只能看录播了

  7. 成为Java GC专家(5)—Java性能调优原则

    并不是每个程序都需要调优.如果一个程序性能表现和预期一样,你不必付出额外的精力去提高它的性能.然而,在程序调试完成之后,很难马上就满足它的性能需求,于是就有了调优这项工作.无论哪种编程语言,对应用程序 ...

  8. MySQL 性能调优之存储引擎

    原文:http://bbs.landingbj.com/t-0-246222-1.html        http://bbs.landingbj.com/t-0-245851-1.html MySQ ...

  9. 性能调优的本质、Spark资源使用原理和调优要点分析

    本课主题 大数据性能调优的本质 Spark 性能调优要点分析 Spark 资源使用原理流程 Spark 资源调优最佳实战 Spark 更高性能的算子 引言 我们谈大数据性能调优,到底在谈什么,它的本质 ...

  10. hbase性能调优(1)

    hbase性能调优 标签: hbase 性能调优 | 发表时间:2014-05-17 15:10 | 作者:无尘道长 分享到: 出处:http://www.iteye.com 一.服务端调优 1.参数 ...

随机推荐

  1. CF431C

    题目简化和分析: k叉树,乍一看好像是树论,但我们通过分析条件,发现它每个阶段要做的事情一样,皆为:\(1\sim k\) 中选数字,这就很明显是DP. \(\mathit{f}_{i,0}\) 表示 ...

  2. client-go实战之八:更新资源时的冲突错误处理

    欢迎访问我的GitHub 这里分类和汇总了欣宸的全部原创(含配套源码):https://github.com/zq2599/blog_demos 本篇概览 本文是<client-go实战> ...

  3. Python 利用pandas和matplotlib绘制饼图

    这段代码使用了Pandas和Matplotlib库来绘制店铺销售数量占比的饼图.通过读取Excel文件中的数据,对店铺名称进行聚合并按销售数量降序排列,然后使用Matplotlib绘制饼图展示销售数量 ...

  4. k 分算法是 k 越大越好吗?

    引入 我们有二分算法,就是: 定义 二分查找(英语:binary search),也称折半搜索(英语:half-interval search).对数搜索(英语:logarithmic search) ...

  5. 洛谷1451(BFS)

    #include"bits/stdc++.h" using namespace std; int mp[110][110]; bool vis[110][110]; int dx[ ...

  6. PostgreSQL 提升子连接与 ORACLE 子查询非嵌套

    查询优化器对子查询一般采用嵌套执行的方式,即父查询中的每一行,都要执行一次子查询,这样子查询会执行很多次,效率非常低. 例如 exists.not exists 逐行取出经行匹配处理,项目中使用子查询 ...

  7. fileinclude

    这里的话需要注意几个点 cookie中的变量赋值不适用冒号而是等于号 如果后面有路径的话就直接写文件名就好了,不用写后缀

  8. windows10 使用 USB 无线网卡的热点功能

    一.概述 在某宝上买了一个 COMFAST CF-727B 的无线模块,由于笔记本电脑一直使用不上,所以放了很久.多年后我来到了一个公司,遇到了我此生最想吐槽的网管,简直不敢想象几十人的办公室,居然能 ...

  9. L3-013 非常弹的球

    初速度:\(v = \sqrt{\cfrac{2 * E}{m}};\) 竖直速度:\(v_y = v \,sin(ans) = g * t\) 水平距离:\(s = v_x * 2t = v \, ...

  10. vue-router重写push方法,解决相同路径跳转报错,解决点击菜单栏打开外部链接

    修改vue-router的配置文件,默认位置router/index.js import Vue from 'vue' import Router from 'vue-router' /** * 重写 ...