oralce之 10046对Hash Join分析
前两天解决了一个优化SQL的case，SQL语句如下，big_table为150G大小，small_table很小，9000多条记录，不到1M大小，hash_area_size, sort_area_size均设置足够大，可以进行optimal hash join和memory sort。
|
1
2
3
4
5
6
|
select /*+ leading(b) use_hash(a b) */ distinct a.IDfrom BIG_TABLE a, SMALL_TABLE bwhere (a.category = b.from_cat or a.category2 = b.from_cat) and a.site_id = b.site_id and a.sale_end >= sysdate; |
执行计划如下：
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
|
--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | 2 | 174 | 18 (17)|| 1 | SORT UNIQUE | | 2 | 174 | 18 (17)||* 2 | HASH JOIN | | 2 | 174 | 17 (12)|| 3 | TABLE ACCESS FULL | SMALL_TABLE | 1879 | 48854 | 14 (8)||* 4 | TABLE ACCESS FULL | BIG_TABLE | 4 | 244 | 3 (34)|--------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 2 - access("A"."SITE_ID"="B"."SITE_ID") filter("A"."CATEGORY"="B"."FROM_CAT" OR "A"."CATEGORY2"="B"."FROM_CAT") 4 - filter("A"."SALE_END">=SYSDATE@!) |
粗略来看，PLAN非常的完美，SQL HINT写的也很到位，小表在内build hash table,大表在外进行probe操作，根据经验来看，整个SQL执行的时间应该和FTS（Full Table Scan） BIG_TABLE的时间差不多。
但是FTS BIG_TABLE的时间大约是8分钟，而真个SQL执行的时间长达3~4小时。
那么问题究竟出在哪里?
FTS时间应该不会有太大变化，那么问题应该在hash join，设置event来trace一下hash join的过程：
|
1
|
alter session set events '10104 trace name context forever, level 2'; |
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
|
### Hash table #### NOTE: The calculated number of rows in non-empty buckets may be smaller# than the true number.Number of buckets with 0 rows: 16373Number of buckets with 1 rows: 0Number of buckets with 2 rows: 0Number of buckets with 3 rows: 1Number of buckets with 4 rows: 0Number of buckets with 5 rows: 0Number of buckets with 6 rows: 0Number of buckets with 7 rows: 1Number of buckets with 8 rows: 0Number of buckets with 9 rows: 0Number of buckets with between 10 and 19 rows: 1Number of buckets with between 20 and 29 rows: 1Number of buckets with between 30 and 39 rows: 3Number of buckets with between 40 and 49 rows: 0Number of buckets with between 50 and 59 rows: 0Number of buckets with between 60 and 69 rows: 0Number of buckets with between 70 and 79 rows: 0Number of buckets with between 80 and 89 rows: 0Number of buckets with between 90 and 99 rows: 0Number of buckets with 100 or more rows: 4### Hash table overall statistics ###Total buckets: 16384 Empty buckets: 16373 Non-empty buckets: 11Total number of rows: 9232Maximum number of rows in a bucket: 2531Average number of rows in non-empty buckets: 839.272705 |
仔细看，在一个bucket中最多的行数竟然有2531行，因为bucket中是一个链表的结构，所以这几千行都是串在一个链表上。
由这一点想到这个Hash Table所依赖的hash key的distinct value可能太少，重复值太多。否则不应该会有这么多行在同一个bucket里面。
因为Join条件里面有两个列from_cat和site_id,穷举法有三种情况：
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
|
SQL> select site_id,from_cat,count(*) from SMALL_TABLE group by site_id,from_cat having count(*)>100;no rows selected2. Build hash table based on (from_cat):SQL> select from_cat,count(*) from SMALL_TABLE group by from_cat having count(*)>100;no rows selected3. Build hash table based on (site_id):SQL> select site_id,count(*) from SMALL_TABLE group by site_id having count(*)>100; SITE_ID COUNT(*)---------- ---------- 0 2531 2 2527 146 1490 210 2526 |
到这里可以发现，基于site_id这种情况和trace file中这两行很相符：
|
1
2
|
Number of buckets with 100 or more rows: 4Maximum number of rows in a bucket: 2531 |
注：这判断过程可以从执行计划的“Predicate Information”部分看出：
|
1
|
access("A"."SITE_ID"="B"."SITE_ID") |
所以推断这个hash table是基于site_id而建的，而Big_Table中大量的行site_id=0，都落在这个linked list最长的bucket中，而大部分行都会扫描完整个链表而最后被丢弃掉,所以这个Hash Join的操作效率非常差，几乎变为了Nest Loop操作。
找到了根本原因，问题也就迎刃而解了。
理想状况下，hash table应当建立于(site_id,from_cat)上，那么问题肯定出在这个OR上，把OR用UNION改写：
|
1
2
3
4
5
6
7
8
9
10
11
|
select /*+ leading(b) use_hash(a b) */ distinct a.IDfrom BIG_TABLE a, SMALL_TABLE bwhere a.category = b.from_cat and a.site_id = b.site_id and a.sale_end >= sysdateUNIONselect /*+ leading(b) use_hash(a b) */ distinct a.IDfrom BIG_TABLE a, SMALL_TABLE bwhere a.category2 = b.from_cat and a.site_id = b.site_id and a.sale_end >= sysdate; |
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
|
--------------------------------------------------------------------------| Id | Operation | Name | Rows | Bytes | Cost (%CPU)|--------------------------------------------------------------------------| 0 | SELECT STATEMENT | | 2 | 148 | 36 (59)|| 1 | SORT UNIQUE | | 2 | 148 | 36 (59)|| 2 | UNION-ALL | | | | ||* 3 | HASH JOIN | | 1 | 74 | 17 (12)|| 4 | TABLE ACCESS FULL| SMALL_TABLE | 1879 | 48854 | 14 (8)||* 5 | TABLE ACCESS FULL| BIG_TABLE | 4 | 192 | 3 (34)||* 6 | HASH JOIN | | 1 | 74 | 17 (12)|| 7 | TABLE ACCESS FULL| SMALL_TABLE | 1879 | 48854 | 14 (8)||* 8 | TABLE ACCESS FULL| BIG_TABLE | 4 | 192 | 3 (34)|--------------------------------------------------------------------------Predicate Information (identified by operation id):--------------------------------------------------- 3 - access("A"."CATEGORY"="B"."FROM_CAT" AND "A"."SITE_ID"="B"."SITE_ID") 5 - filter("A"."SALE_END">=SYSDATE@!) 6 - access("A"."CATEGORY2"="B"."FROM_CAT" AND "A"."SITE_ID"="B"."SITE_ID") 8 - filter("A"."SALE_END">=SYSDATE@!) |
初看这个PLAN好像不如第一个PLAN，因为执行了两次BIG_TABLE的FTS，但是让我们在来看看HASH TABLE的结构
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
|
### Hash table #### NOTE: The calculated number of rows in non-empty buckets may be smaller# than the true number.Number of buckets with 0 rows: 9306Number of buckets with 1 rows: 5310Number of buckets with 2 rows: 1436Number of buckets with 3 rows: 285Number of buckets with 4 rows: 43Number of buckets with 5 rows: 4Number of buckets with 6 rows: 0Number of buckets with 7 rows: 0Number of buckets with 8 rows: 0Number of buckets with 9 rows: 0Number of buckets with between 10 and 19 rows: 0Number of buckets with between 20 and 29 rows: 0Number of buckets with between 30 and 39 rows: 0Number of buckets with between 40 and 49 rows: 0Number of buckets with between 50 and 59 rows: 0Number of buckets with between 60 and 69 rows: 0Number of buckets with between 70 and 79 rows: 0Number of buckets with between 80 and 89 rows: 0Number of buckets with between 90 and 99 rows: 0Number of buckets with 100 or more rows: 0### Hash table overall statistics ###Total buckets: 16384 Empty buckets: 9306 Non-empty buckets: 7078Total number of rows: 9232Maximum number of rows in a bucket: 5Average number of rows in non-empty buckets: 1.304323 |
这就是我们所需要的Hash Table，最长的链表只有五行数据。
整个SQL的执行时间从三四个小时缩短为16分钟，大大超出了developer的预期。
这个SQL单纯从PLAN上很难看出问题所在，需要了解Hash Join的机制，进行更深一步的分析。
source:http://www.itpub.net/thread-955209-1-1.html
oralce之 10046对Hash Join分析的更多相关文章
- minhash pyspark 源码分析——hash join table是关键
从下面分析可以看出,是先做了hash计算,然后使用hash join table来讲hash值相等的数据合并在一起.然后再使用udf计算距离,最后再filter出满足阈值的数据: 参考:https:/ ...
- Merge join、Hash join、Nested loop join对比分析
简介 我们所常见的表与表之间的Inner Join,Outer Join都会被执行引擎根据所选的列,数据上是否有索引,所选数据的选择性转化为Loop Join,Merge Join,Hash Join ...
- SQL Tuning 基础概述06 - 表的关联方式：Nested Loops Join，Merge Sort Join & Hash Join
nested loops join(嵌套循环) 驱动表返回几条结果集,被驱动表访问多少次,有驱动顺序,无须排序,无任何限制. 驱动表限制条件有索引,被驱动表连接条件有索引. hints:use_n ...
- Sort merge join、Nested loops、Hash join（三种连接类型）
目前为止,典型的连接类型有3种: Sort merge join(SMJ排序-合并连接):首先生产driving table需要的数据,然后对这些数据按照连接操作关联列进行排序:然后生产probed ...
- 视图合并、hash join连接列数据分布不均匀引发的惨案
表大小 SQL> select count(*) from agent.TB_AGENT_INFO; COUNT(*) ---------- 1751 SQL> select count( ...
- Oracle 表的连接方式(2)-----HASH JOIN的基本机制1
我们对hash join的常见误解,一般包括两个: 第一个误解:是我们经常以为hash join需要对两个做join的表都做全表扫描 第二个误解:是经常以为hash join会选择比较小的表做buil ...
- oracle 表连接 - hash join 哈希连接
一. hash 连接(哈希连接)原理 指的是两个表连接时, 先利用两表中记录较少的表在内存中建立 hash 表, 然后扫描记录较多的表并探測 hash 表, 找出与 hash 表相匹配的行来得到结果集 ...
- [20180713]关于hash join 测试中一个疑问.txt
[20180713]关于hash join 测试中一个疑问.txt --//上个星期做的测试,链接: http://blog.itpub.net/267265/viewspace-2157424/-- ...
- [20180705]关于hash join 2.txt
[20180705]关于hash join 2.txt --//昨天优化sql语句,执行计划hash join right sna,加入一个约束设置XX字段not null,逻辑读从上万下降到50.- ...
随机推荐
- C++（二十四） — 指向字符的指针为什么可以用字符串来初始化，而不是字符地址？
一.C语言中,为什么字符串可以赋值给字符指针变量? char *p: a='; p=&a; //显然是正确的, p="abcd"; //但为什么也可以这样赋值?? 问:一直 ...
- [洛谷U63006]导函数最小系数
U63006 导函数最小系数 题面 给出一个n次函数\(f(x)=a_{n}x^{n}+a_{n-1}x^{n-1}+...+a_{1}x+a_0\)的各项系数\(a_n,a_{n-1}...a_1, ...
- 1004: [HNOI2008]Cards burnside定理
https://www.lydsy.com/JudgeOnline/problem.php?id=1004 输入数据保证任意多次洗牌都可用这 m种洗牌法中的一种代替,且对每种洗牌法,都存在一种洗牌法使 ...
- uva 12086 线段树or树状数组练习
题目链接 https://vjudge.net/problem/34215/origin 这个题就是线段树裸题,有两种操作,实现单点更新和区间和的查找即可,这里第一次学习使用树状数组完成. 二者相 ...
- POJ 2891 中国剩余定理的非互质形式
中国剩余定理的非互质形式 任意n个表达式一对对处理,故只需处理两个表达式. x = a(mod m) x = b(mod n) km+a = b (mod n) km = (a-b)(mod n) 利 ...
- 移动设备web开发插件iScroll的使用详解
地址:http://blog.nnnv.cn/index.php/archives/65
- c# JScriptProvider包装
using System; using System.CodeDom.Compiler; using System.Reflection; using System.Web.UI; using Mic ...
- Getting command line access to PHP and MySQL running MAMP on OSX
建立自己profile路径应该在/Users/yourname/,最后要运行. ./.profile使文件生效,和windows中给添加环境变量是一个道理,还可以看出linux和UNIX默认运行路径为 ...
- PHP sessions that have already been started
In my PHP code, if a session has already started, and I try to start a new one, I get the following ...
- js跨域详解
跨域概念:Cross-origin resource sharing