ORACLE 执行计划分析
一、什么是执行计划
An explain plan is a representation of the access path that is taken when a query is executed within Oracle.
二、如何访问数据
At the physical level Oracle reads blocks of data. The smallest amount of data read is a single Oracle block, the largest is constrained by operating system limits (and multiblock i/o). Logically Oracle finds the data to read by using the following methods:
Full Table Scan (FTS) --全表扫描
Index Lookup (unique & non-unique) --索引扫描（唯一和非唯一）
Rowid --物理行id
三、执行计划层次关系
When looking at a plan, the rightmost (ie most inndented) uppermost operation is the first thing that is executed. --采用最右最上最先执行的原则看层次关系，在同一级如果某个动作没有子ID就最先执行
1.看一个简单的例子：
Query Plan
-----------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1234
**TABLE ACCESS FULL LARGE [:Q65001] [ANALYZED] --[:Q65001]表示是并行方式，[ANALYZED]表示该对象已经分析过了
优化模式是CHOOSE的情况下，看Cost参数是否有值来决定采用CBO还是RBO：
SELECT STATEMENT [CHOOSE] Cost=1234 --Cost有值，采用CBO
SELECT STATEMENT [CHOOSE] Cost= --Cost为空，采用RBO
2.层次的父子关系，看比较复杂的例子：
PARENT1
**FIRST CHILD
****FIRST GRANDCHILD
**SECOND CHILD
Here the same principles apply, the FIRST GRANDCHILD is the initial operation then the FIRST CHILD followed by the SECOND CHILD and finally the PARENT collates the output.
四、例子解说
Execution Plan
----------------------------------------------------------
0 **SELECT STATEMENT Optimizer=CHOOSE (Cost=3 Card=8 Bytes=248)
1 0 **HASH JOIN (Cost=3 Card=8 Bytes=248)
2 1 ****TABLE ACCESS (FULL) OF 'DEPT' (Cost=1 Card=3 Bytes=36)
3 1 ****TABLE ACCESS (FULL) OF 'EMP' (Cost=1 Card=16 Bytes=304)
左侧的两排数据，前面的是序列号ID，后面的是对应的PID（父ID）。
A shortened summary of this is:
Execution starts with ID=0: SELECT STATEMENT but this is dependand on it's child objects
So it executes its first child step: ID=1 PID=0 HASH JOIN but this is dependand on it's child objects
So it executes its first child step: ID=2 PID=1 TABLE ACCESS (FULL) OF 'DEPT'
Then the second child step: ID=3 PID=2 TABLE ACCESS (FULL) OF 'EMP'
Rows are returned to the parent step(s) until finished
五、表访问方式
1.Full Table Scan (FTS) 全表扫描
In a FTS operation, the whole table is read up to the high water mark (HWM). The HWM marks the last block in the table that has ever had data written to it. If you have deleted all the rows then you will still read up to the HWM. Truncate resets the HWM back to the start of the table. FTS uses multiblock i/o to read the blocks from disk. --全表扫描模式下会读数据到表的高水位线（HWM即表示表曾经扩展的最后一个数据块），读取速度依赖于Oracle初始化参数db_block_multiblock_read_count
Query Plan
------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
**INDEX UNIQUE SCAN EMP_I1 --如果索引里就找到了所要的数据，就不会再去访问表了
2.Index Lookup 索引扫描
There are 5 methods of index lookup:
index unique scan --索引唯一扫描
Method for looking up a single key value via a unique index. always returns a single value, You must supply AT LEAST the leading column of the index to access data via the index.
eg:
SQL> explain plan for select empno,ename from emp where empno=10;
index range scan --索引局部扫描
Index range scan is a method for accessing a range values of a particular column. AT LEAST the leading column of the index must be supplied to access data via the index. Can be used for range operations (e.g. > < <> >= <= between) .
eg:
SQL> explain plan for select mgr from emp where mgr = 5;
index full scan --索引全局扫描
Full index scans are only available in the CBO as otherwise we are unable to determine whether a full scan would be a good idea or not. We choose an index Full Scan when we have statistics that indicate that it is going to be more efficient than a Full table scan and a sort. For example we may do a Full index scan when we do an unbounded scan of an index and want the data to be ordered in the index order.
eg:
SQL> explain plan for select empno,ename from big_emp order by empno,ename;
index fast full scan --索引快速全局扫描，不带order by情况下常发生
Scans all the block in the index, Rows are not returned in sorted order, Introduced in 7.3 and requires V733_PLANS_ENABLED=TRUE and CBO, may be hinted using INDEX_FFS hint, uses multiblock i/o, can be executed in parallel, can be used to access second column of concatenated indexes. This is because we are selecting all of the index.
eg:
SQL> explain plan for select empno,ename from big_emp;
index skip scan --索引跳跃扫描，where条件列是非索引的前导列情况下常发生
Index skip scan finds rows even if the column is not the leading column of a concatenated index. It skips the first column(s) during the search.
eg:
SQL> create index i_emp on emp(empno, ename);
SQL> select /*+ index_ss(emp i_emp)*/ job from emp where ename='SMITH';
3.Rowid 物理ID扫描
This is the quickest access method available.Oracle retrieves the specified block and extracts the rows it is interested in. --Rowid扫描是最快的访问数据方式
六、表连接方式
有三种连接方式：
1.Sort Merge Join (SMJ) --由于sort是非常耗资源的，所以这种连接方式要避免
Rows are produced by Row Source 1 and are then sorted Rows from Row Source 2 are then produced and sorted by the same sort key as Row Source 1. Row Source 1 and 2 are NOT accessed concurrently.
SQL> explain plan for
select /*+ ordered */ e.deptno,d.deptno
from emp e,dept d
where e.deptno = d.deptno
order by e.deptno,d.deptno;
Query Plan
-------------------------------------
SELECT STATEMENT [CHOOSE] Cost=17
**MERGE JOIN
****SORT JOIN
******TABLE ACCESS FULL EMP [ANALYZED]
****SORT JOIN
******TABLE ACCESS FULL DEPT [ANALYZED]
Sorting is an expensive operation, especially with large tables. Because of this, SMJ is often not a particularly efficient join method.
2.Nested Loops (NL) --比较高效的一种连接方式
Fetches the first batch of rows from row source 1, Then we probe row source 2 once for each row returned from row source 1.
For nested loops to be efficient it is important that the first row source returns as few rows as possible as this directly controls the number of probes of the second row source. Also it helps if the access method for row source 2 is efficient as this operation is being repeated once for every row returned by row source 1.
SQL> explain plan for
select a.dname,b.sql
from dept a,emp b
where a.deptno = b.deptno;
Query Plan
-------------------------
SELECT STATEMENT [CHOOSE] Cost=5
**NESTED LOOPS
****TABLE ACCESS FULL DEPT [ANALYZED]
****TABLE ACCESS FULL EMP [ANALYZED]
3.Hash Join --最为高效的一种连接方式
New join type introduced in 7.3, More efficient in theory than NL & SMJ, Only accessible via the CBO. Smallest row source is chosen and used to build a hash table and a bitmap The second row source is hashed and checked against the hash table looking for joins. The bitmap is used as a quick lookup to check if rows are in the hash table and are especially useful when the hash table is too large to fit in memory.
SQL> explain plan for
select /*+ use_hash(emp) */ empno
from emp,dept
where emp.deptno = dept.deptno;
Query Plan
----------------------------
SELECT STATEMENT [CHOOSE] Cost=3
**HASH JOIN
****TABLE ACCESS FULL DEPT
****TABLE ACCESS FULL EMP
Hash joins are enabled by the parameter HASH_JOIN_ENABLED=TRUE in the init.ora or session. TRUE is the default in 7.3.
3.Cartesian Product --卡迪尔积，不算真正的连接方式，sql肯定写的有问题
A Cartesian Product is done where they are no join conditions between 2 row sources and there is no alternative method of accessing the data. Not really a join as such as there is no join! Typically this is caused by a coding mistake where a join has been left out.
It can be useful in some circumstances - Star joins uses cartesian products.Notice that there is no join between the 2 tables:
SQL> explain plan for
select emp.deptno,dept,deptno
from emp,dept
Query Plan
------------------------------
SLECT STATEMENT [CHOOSE] Cost=5
**MERGE JOIN CARTESIAN
****TABLE ACCESS FULL DEPT
****SORT JOIN
******TABLE ACCESS FULL EMP
The CARTESIAN keyword indicate that we are doing a cartesian product.
七、运算符
1.sort --排序，很消耗资源
There are a number of different operations that promote sorts:
order by clauses
group by
sort merge join
2.filter --过滤，如not in、min函数等容易产生
Has a number of different meanings, used to indicate partition elimination, may also indicate an actual filter step where one row source is filtering, another, functions such as min may introduce filter steps into query plans.
3.view --视图，大都由内联视图产生
When a view cannot be merged into the main query you will often see a projection view operation. This indicates that the 'view' will be selected from directly as opposed to being broken down into joins on the base tables. A number of constructs make a view non mergeable. Inline views are also non mergeable.
eg:
SQL> explain plan for
select ename,tot
from emp,(select empno,sum(empno) tot from big_emp group by empno) tmp
where emp.empno = tmp.empno;
Query Plan
------------------------
SELECT STATEMENT [CHOOSE]
**HASH JOIN
**TABLE ACCESS FULL EMP [ANALYZED]
**VIEW
****SORT GROUP BY
******INDEX FULL SCAN BE_IX
4.partition view --分区视图
Partition views are a legacy technology that were superceded by the partitioning option. This section of the article is provided as reference for such legacy systems.
假设LARGE_TABLE是一个较大的表，且username列上没有索引，则运行下面的语句：
|
SQL> SELECT * FROM LARGE_TABLE where USERNAME = ‘TEST’; Query Plan ----------------------------------------- SELECT STATEMENT Optimizer=CHOOSE (Cost=1234 Card=1 Bytes=14) TABLE ACCESS FULL LARGE_TABLE [:Q65001] [ANALYZED] |
在这个例子中，TABLE ACCESS FULL LARGE_TABLE是第一个操作，意思是在LARGE_TABLE表上做全表扫描。当这个操作完成之后，产生的row source中的数据被送往下一步骤进行处理，在此例中，SELECT STATEMENT操作是这个查询语句的最后一步。
Optimizer=CHOOSE 指明这个查询的optimizer_mode，即optimizer_mode初始化参数指定的值，它并不是指语句执行时真的使用了该优化器。决定该语句使用何种优化器的唯一方法是看后面的cost部分。例如，如果给出的是下面的形式，则表明使用的是CBO优化器，此处的cost表示优化器认为该执行计划的代价:
SELECT STATEMENT ptimizer=CHOOSE (Cost=1234 Card=1 Bytes=14)
然而假如执行计划中给出的是类似下面的信息，则表明是使用RBO优化器，因为cost部分的值为空，或者压根就没有cost部分。
SELECT STATEMENT ptimizer=CHOOSE Cost=
SELECT STATEMENT ptimizer=CHOOSE
这样我们从Optimizer后面的信息中可以得出执行该语句时到底用了什么样的优化器。特别的，如果Optimizer=ALL_ROWS| FIRST_ROWS| FIRST_ROWS_n，则使用的是CBO优化器;如果Optimizer=RULE，则使用的是RBO优化器。
cost属性的值是一个在oracle内部用来比较各个执行计划所耗费的代价的值，从而使优化器可以选择最好的执行计划。不同语句的cost值不具有可比性，只能对同一个语句的不同执行计划的cost值进行比较。
[:Q65001] 表明该部分查询是以并行方式运行的。里面的数据表示这个操作是由并行查询的一个slave进程处理的，以便该操作可以区别于串行执行的操作。
[ANALYZED] 表明操作中引用的对象被分析过了，在数据字典中有该对象的统计信息可以供CBO使用。
例2：
假定A、B、C都是不是小表，且在A表上一个组合索引：A(a.col1,a.col2) ，注意a.col1列为索引的引导列。
考虑下面的查询：
|
select A.col4 from A , B , C where B.col3 = 10 and A.col1 = B.col1 and A.col2 = C.col2 and C.col3 = 5 Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=CHOOSE 1 0 MERGE JOIN 2 1 SORT (JOIN) 3 2 NESTED LOOPS 4 3 TABLE ACCESS (FULL) OF 'B' 5 3 TABLE ACCESS (BY INDEX ROWID) OF 'A' 6 5 INDEX (RANGE SCAN) OF 'INX_COL12A' (NON-UNIQUE) 7 1 SORT (JOIN) 8 7 TABLE ACCESS (FULL) OF 'C'
Statistics |
在表做连接时，只能2个表先做连接，然后将连接后的结果作为一个row source，与剩下的表做连接，在上面的例子中，连接顺序为B与A先连接，然后再与C连接：
B <---> A <---> C
col3=10 col3=5
如果没有执行计划，分析一下，上面的3个表应该拿哪一个作为第一个驱动表?从SQL语句看来，只有B表与C表上有限制条件，所以第一个驱动表应该为这2个表中的一个，到底是哪一个呢?
B表有谓词B.col3 = 10，这样在对B表做全表扫描的时候就将where子句中的限制条件(B.col3 = 10)用上，从而得到一个较小的row source, 所以B表应该作为第一个驱动表。而且这样的话，如果再与A表做关联，可以有效利用A表的索引(因为A表的col1列为leading column)。
当然上面的查询中C表上也有谓词(C.col3 = 5)，有人可能认为C表作为第一个驱动表也能获得较好的性能。让我们再来分析一下：如果C表作为第一个驱动表，则能保证驱动表生成很小的row source，但是看看连接条件A.col2 = C.col2，此时就没有机会利用A表的索引，因为A表的col2列不为leading column，这样nested loop的效率很差，从而导致查询的效率很差。所以对于NL连接选择正确的驱动表很重要。
因此上面查询比较好的连接顺序为(B - - > A) - - > C。如果数据库是基于代价的优化器，它会利用计算出的代价来决定合适的驱动表与合适的连接顺序。一般来说，CBO都会选择正确的连接顺序，如果CBO选择了比较差的连接顺序，我们还可以使用ORACLE提供的hints来让CBO采用正确的连接顺序。如下所示：
|
select /*+ ordered */ A.col4 from B,A,C where B.col3 = 10 and A.col1 = B.col1 and A.col2 = C.col2 and C.col3 = 5 |
既然选择正确的驱动表这么重要，那么让我们来看一下执行计划，到底各个表之间是如何关联的，从而得到执行计划中哪个表应该为驱动表：
在执行计划中，需要知道哪个操作是先执行的，哪个操作是后执行的，这对于判断哪个表为驱动表有用处。判断之前，如果对表的访问是通过rowid，且该rowid的值是从索引扫描中得来得，则将该索引扫描先从执行计划中暂时去掉。然后在执行计划剩下的部分中，判断执行顺序的指导原则就是：最右、最上的操作先执行。具体解释如下：
得到去除妨碍判断的索引扫描后的执行计划：
|
Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=CHOOSE 1 0 MERGE JOIN 2 1 SORT (JOIN) 3 2 NESTED LOOPS 4 3 TABLE ACCESS (FULL) OF 'B' 5 3 TABLE ACCESS (BY INDEX ROWID) OF 'A' 7 1 SORT (JOIN) 8 7 TABLE ACCESS (FULL) OF 'C' |
看执行计划的第3列，即字母部分，每列值的左面有空格作为缩进字符。在该列值左边的空格越多，说明该列值的缩进越多，该列值也越靠右。如上面的执行计划所示：第一列值为6的行的缩进最多，即该行最靠右;第一列值为4、5的行的缩进一样，其靠右的程度也一样，但是第一列值为4的行比第一列值为5的行靠上;谈论上下关系时，只对连续的、缩进一致的行有效。
从这个图中我们可以看到，对于NESTED LOOPS部分，最右、最上的操作是TABLE ACCESS (FULL) OF 'B'，所以这一操作先执行，所以该操作对应的B表为第一个驱动表(外部表)，自然，A表就为内部表了。从图中还可以看出，B与A表做嵌套循环后生成了新的row source ，对该row source进行来排序后，与C表对应的排序了的row source(应用了C.col3 = 5限制条件)进行MSJ连接操作。所以从上面可以得出如下事实：B表先与A表做嵌套循环，然后将生成的row source与C表做排序—合并连接。
通过分析上面的执行计划，我们不能说C表一定在B、A表之后才被读取，事实上，B表有可能与C表同时被读入内存，因为将表中的数据读入内存的操作可能为并行的。事实上许多操作可能为交叉进行的，因为ORACLE读取数据时，如果就是需要一行数据也是将该行所在的整个数据块读入内存，而且还有可能为多块读。
看执行计划时，我们的关键不是看哪个操作先执行，哪个操作后执行，而是关键看表之间连接的顺序(如得知哪个为驱动表，这需要从操作的顺序进行判断)、使用了何种类型的关联及具体的存取路径(如判断是否利用了索引)
在从执行计划中判断出哪个表为驱动表后，根据我们的知识判断该表作为驱动表(就像上面判断ABC表那样)是否合适，如果不合适，对SQL语句进行更改，使优化器可以选择正确的驱动表。
对于RBO优化器：
在ORACLE文档上说：对于RBO来说，以from 子句中从右到左的顺序选择驱动表，即最右边的表为第一个驱动表，这是其英文原文：All things being equal RBO chooses the driving order by taking the tables in the FROM clause RIGHT to LEFT。不过，在我做的测试中，从来也没有验证过这种说法是正确的。我认为，即使在RBO中，也是有一套规则来决定使用哪种连接类型和哪个表作为驱动表，在选择时肯定会考虑当前索引的情况，还可能会考虑where 中的限制条件，但是肯定是与where中限制条件的位置无关。
测试：
如果我创建3个表：
create table A(col1 number(4,0),col2 number(4,0), col4 char(30));
create table B(col1 number(4,0),col3 number(4,0), name_b char(30));
create table C(col2 number(4,0),col3 number(4,0), name_c char(30));
create index inx_col12A on a(col1,col2);
执行查询：
|
select A.col4 from B, A, C where B.col3 = 10 and A.col1 = B.col1 and A.col2 = C.col2 and C.col3 = 5; Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=RULE 1 0 MERGE JOIN 2 1 SORT (JOIN) 3 2 NESTED LOOPS 4 3 TABLE ACCESS (FULL) OF 'B' 5 3 TABLE ACCESS (BY INDEX ROWID) OF 'A' 6 5 INDEX (RANGE SCAN) OF 'INX_COL12A' (NON-UNIQUE) 7 1 SORT (JOIN) 8 7 TABLE ACCESS (FULL) OF 'C'
select A.col4 |
将A表上的索引inx_col12A删除后：
|
select A.col4 from B, A, C where A.col1 = B.col1 and A.col2 = C.col2; Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=RULE 1 0 MERGE JOIN 2 1 SORT (JOIN) 3 2 MERGE JOIN 4 3 SORT (JOIN) 5 4 TABLE ACCESS (FULL) OF 'C' 6 3 SORT (JOIN) 7 6 TABLE ACCESS (FULL) OF 'A' 8 1 SORT (JOIN) 9 8 TABLE ACCESS (FULL) OF 'B' |
通过上面的这些例子，使我对oracle文档上的” All things being equal RBO chooses the driving order by taking the tables in the FROM clause RIGHT to LEFT”这句话持怀疑态度。此时，我也不能使用hints来强制优化器使用nested loop，如果使用了hints，这样就自动使用CBO优化器，而不是RBO优化器了。
对于CBO优化器：
CBO根据统计信息选择驱动表，假如没有统计信息，则在from 子句中从左到右的顺序选择驱动表。这与RBO选择的顺序正好相反。这是英文原文(CBO determines join order from costs derived from gathered statistics. If there are no stats then CBO chooses the driving order of tables from LEFT to RIGHT in the FROM clause. This is OPPOSITE to the RBO) 。我还是没法证实这句话的正确性。不过经过验证：“如果用ordered 提示(此时肯定用CBO)，则以from 子句中按从左到右的顺序选择驱动表”这句话是正确的。实际上在CBO中，如果有统计数据(即对表与索引进行了分析)，则优化器会自动根据cost值决定采用哪种连接类型，并选择合适的驱动表，这与where子句中各个限制条件的位置没有任何关系。如果我们要改变优化器选择的连接类型或驱动表，则就需要使用hints了，具体hints的用法在后面会给予介绍。
测试：
如果我创建的3个表：
|
create table A(col1 number(4,0),col2 number(4,0), col4 char(30)); create table B(col1 number(4,0),col3 number(4,0), name_b char(30)); create table C(col2 number(4,0),col3 number(4,0), name_c char(30)); create index inx_col12A on a(col1,col2); |
执行查询：
|
select A.col4 from B, A, C where B.col3 = 10 and A.col1 = B.col1 and A.col2 = C.col2 and C.col3 = 5; Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=ALL_ROWS (Cost=3 Card=1 Bytes=110) 1 0 NESTED LOOPS (Cost=3 Card=1 Bytes=110) 2 1 MERGE JOIN (CARTESIAN) (Cost=2 Card=1 Bytes=52) 3 2 TABLE ACCESS (FULL) OF 'B' (Cost=1 Card=1 Bytes=26) 4 2 SORT (JOIN) (Cost=1 Card=1 Bytes=26) 5 4 TABLE ACCESS (FULL) OF 'C' (Cost=1 Card=1 Bytes=26) 6 1 TABLE ACCESS (FULL) OF 'A' (Cost=1 Card=82 Bytes=4756)
select A.col4 |
将A表上的索引inx_col12A删除后：
|
select A.col4 from B, A, C where A.col1 = B.col1 and A.col2 = C.col2; Execution Plan ---------------------------------------------------------- 0 SELECT STATEMENT ptimizer=ALL_ROWS (Cost=5 Card=55 Bytes=4620) 1 0 HASH JOIN (Cost=5 Card=55 Bytes=4620) 2 1 HASH JOIN (Cost=3 Card=67 Bytes=4757) 3 2 TABLE ACCESS (FULL) OF 'B' (Cost=1 Card=82 Bytes=1066) 4 2 TABLE ACCESS (FULL) OF 'A' (Cost=1 Card=82 Bytes=4756) 5 1 TABLE ACCESS (FULL) OF 'C' (Cost=1 Card=82 Bytes=1066)
select /*+ ORDERED */A.col4 |
这个查询验证了通过ORDERED提示可以正确的提示优化器选择哪个表作为优化器。