ExplainPlan
set autotrace off
set autotrace on
set autotrace traceonly
set autotrace on explain
set autotrace on statistics
set autotrace on explain statistics
set autotrace traceonly explain
set autotrace traceonly statistics
set autotrace traceonly explain statistics
set autotrace off explain
set autotrace off statistics
set autotrace off explain statistics
Setting autotrace allows to display some statistics and/or an query execution plan for DML statements.
set autotrace on: Shows the execution plan as well as statistics of the statement.
set autotrace on explain: Displays the execution plan only.
set autotrace on statistics: Displays the statistics only.
set autotrace traceonly: Displays the execution plan and the statistics (as set autotrace on does), but doesn't print a query's result.
set autotrace off: Disables all autotrace
Interpreting Explain Plan
=========================
Note: This article was developed under the Oracle 8i time frame but much of the information contained herein is still relevant on
more recent versions.
What's an explain plan?
An explain plan is a representation of the access path that is taken when
a query is executed within Oracle.
Query processing can be divided into 7 phases:
[1] Syntactic - checks the syntax of the query
[2] Semantic - checks that all objects exist and are accessible
[3] View Merging - rewrites query as join on base tables as
opposed to using views
[4] Statement Transformation - rewrites query transforming some complex
constructs into simpler ones where
appropriate (e.g. subquery unnesting, in/or
transformation)
[5] Optimization - determines the optimal access path for the
query to take. With the Rule Based
Optimizer (RBO) it uses a set of heuristics
to determine access path. With the Cost
Based Optimizer (CBO) we use statistics
to analyze the relative costs of accessing
objects.
[6] QEP Generation
[7] QEP Execution
(QEP = Query Evaluation Plan)
Steps [1]-[6] are sometimes grouped under the term 'Parsing'
Step [7] is the execution of the statement.
The explain plan a representation of the access path produced in step 6.
Once the access path has been decided upon it is stored in the library cache
together with the statement itself. Queries are stored in the library cache
based upon a hashed representation of the query. When looking for a statement
in the library cache, we first apply a hashing algorithm to the current
statement and then look for this hash value in the library cache.
This access path will be used until the query is re-parsed.
Terminology
Row Source - A row source is a software function that implements specific
operations (such as a table scan or a hash join)
and returns a set of rows.
Predicate - where clause of a query
Tuples - rows
Driving Table - This is the row source that we use to seed the query.
If this returns a lot of rows then this can have a negative
affect on all subsequent operations
Probed Table - This is the object we look-up data in after we have retrieved
relevant key data from the driving table.
How does Oracle access data?
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 multi-block i/o).
Logically Oracle finds the data to read by using the following methods:
Full Table Scan (FTS)
Index Look-up (unique & non-unique)
Rowid
Explain plan Hierarchy
Simple explain plan:
Query Plan
-----------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1234
TABLE ACCESS FULL LARGE [:Q65001] [ANALYZED]
When looking at a plan, the rightmost (ie most inndented) uppermost operation
is the first thing that is executed.
A more detailed explanation of this much simplified desctiption can be found below.
In this case TABLE ACCESS FULL LARGE is the first operation.
This statement means we are doing a full table scan of table LARGE.
When this operation completes then the resultant row source is passed up to the
next level of the query for processing. In this case it is the SELECT STATEMENT
which is the top of the query.
[CHOOSE] is an indication of the optimizer_goal for the query. This DOES NOT
necessarily indicate that plan has actually used this goal. The only way to
confirm this is to check the cost= part of the explain plan as well.
For example the following query indicates that the CBO has
been used because there is a cost in the cost field:
SELECT STATEMENT [CHOOSE] Cost=1234
However the explain plan below indicates the use of the RBO because the cost
field is blank:
SELECT STATEMENT [CHOOSE] Cost=
The cost field is a comparative cost that is used internally to determine the
best cost for particular plans. The costs of different statements are not
really directly comparable.
[:Q65001] indicates that this particular part of the query is being executed
in parallel. This number indicates that the operation will be processed by a
parallel query slave as opposed to being executed serially.
[ANALYZED] indicates that the object in question has been analyzed and there
are currently statistics available for the CBO to use. There is no indication
of the 'level' of analysis done.
More about Plans and the order of execution
===========================================
To understand plans and the order of execution, it is necessary to
understand the PARENT -- CHILD relationships involved:
PARENT
FIRST CHILD
SECOND CHILD
In this example, the FIRST CHILD is executed first followed by the
SECOND CHILD, then the PARENT collates the output in some way.
A more complex case is:
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.
These priniples can be applied to real operations as in the examples below.
Consider the following query:
EXAMPLE 1
=========
set autotrace traceonly explain
select ename,dname
from emp, dept
where emp.deptno=dept.deptno
and dept.dname in ('ACCOUNTING','RESEARCH','SALES','OPERATIONS');
15 rows selected.
This produces the following explain plan:
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)
In the Execution Plan above, there are 2 columns of numbers that precede
the plan text of each line. The first number is the Statement ID
(usually known simply as "ID"), the second number is the Parent ID
of the line (The first line of the plan shows no parent because,
as the first line, it has no parent). The ID and PARENT ID are what
the plan generator uses to construct the explain plan.
The steps in the explain plan are indented to indicate the hierarchy of
operations and which steps are dependent on which other steps.
What follows is a walk-through of the plan above:
Execution starts with: ID=0:
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=3 Card=8 Bytes=248)
ID=0 has no parent but it has 1 child.
ID=0 is the parent of ID=1 and is dependent upon it for rows.
So ID=1 must be executed prior to ID=0
Moving on to ID=1:
1 0 HASH JOIN (Cost=3 Card=8 Bytes=248)
ID=1 is the child of ID=0
ID=1 is the parent of ID=2 and ID=3 and is dependent upon them for rows.
So ID=2 and ID=3 must be executed prior to ID=1
Moving on to ID=2:
2 1 TABLE ACCESS (FULL) OF 'DEPT' (Cost=1 Card=3 Bytes=36)
ID=2 is the first child of ID=1
ID=2 is not the parent of any other step.
This is the first step that is executed by the query.
Rows are provided to ID=1 from this step
ID=1 is also dependent on ID=3:
3 1 TABLE ACCESS (FULL) OF 'EMP' (Cost=1 Card=16 Bytes=304)
ID=3 is the second child of ID=1
ID=3 is not the parent of any other step.
This is the second step that is executed by the query.
Rows are provided to ID=1 from this step
ID=1 processes the rows it receives from its dependent steps (ID=2 & ID=3)
and returns them to its parent ID=0.
ID=0 returns the rows to the user.
A shortened summary of this is:
Execution starts with ID=0: SELECT STATEMENT but this is dependent on it's child objects
So it executes its first child step: ID=1 PID=0 HASH JOIN but this is dependent 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
EXAMPLE 2
=========
select /*+ RULE */ ename,dname
from dept , emp
where emp.deptno=dept.deptno
and dept.dname in ('ACCOUNTING','RESEARCH','SALES','OPERATIONS');
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=HINT: RULE
1 0 NESTED LOOPS
2 1 TABLE ACCESS (FULL) OF 'EMP'
3 1 TABLE ACCESS (BY INDEX ROWID) OF 'DEPT'
4 3 INDEX (UNIQUE SCAN) OF 'PK_DEPT' (UNIQUE)
Execution starts with ID=0: SELECT STATEMENT but this is dependent on it's child objects
So it executes it's first child step: ID=1 PID=0 NESTED LOOPS but this is dependent on it's child objects
So it executes it's first child step: ID=2 PID=1 TABLE ACCESS (FULL) OF 'EMP'
Rows are fetched in ID=2 These rows are returned to the parent
to drive the join in the parent: ID=1 PID=0 NESTED LOOPS
ID=1 uses the rows to execute ID=3: ID=3 PID=1 TABLE ACCESS (BY INDEX ROWID) OF 'DEPT' but this is dependent on it's child objects
So it executes it's first child step: ID=4 PID=3 INDEX (UNIQUE SCAN) OF 'PK_DEPT'
Rows are returned to the parent ID=1
ID=1 returns rows to it's parent ID=0
Once the rows fetched from ID=2 are exhausted another fetch is done to
get the next set of rows.
This process repeats until all the rows retched from ID=2 are exhausted
When starting to looks at explain plans, it is useful to use the ID/PARENT ID
but with practice the indentation, ordering and operations alone will become
enough to determine the access path used.
There are many ways of describing how to determine the operation in a plan,
once familiar with a particular method it becomes second nature. One such
description was to say that the rightmost-uppermost operation of an explain
plan is executed first, but although this proved an intuitive description
after some practice it is confusing to some readers. If in doubt
consult the id and parent id hierarchy.
Access Methods in detail
========================
Full Table Scan (FTS)
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 multi-block i/o to read the blocks from disk.
Multi-block i/o is controlled by the parameter <>
This defaults to:
db_block_buffers / ( (PROCESSES+3) / 4 )
The Maximum values are OS dependent
Buffers from FTS operations are placed on the Least Recently Used (LRU) end of
the buffer cache so will be quickly aged out.
FTS is not recommended for large tables unless you are reading >5-10% of it
(or so) or you intend to run in parallel.
Example FTS explain plan:
SQL> explain plan for select * from dual;
Query Plan
-----------------------------------------
SELECT STATEMENT [CHOOSE] Cost=
TABLE ACCESS FULL DUAL
Index lookup
Data is accessed by looking up key values in an index and returning rowids.
A rowid uniquely identifies an individual row in a particular data block.
This block is read via single block i/o.
In this example an index is used to find the relevant row(s) and then the
table is accessed to lookup the ename column (which is not included in the
index):
SQL> explain plan for select empno,ename from emp where empno=10;
Query Plan
------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
TABLE ACCESS BY ROWID EMP [ANALYZED]
INDEX UNIQUE SCAN EMP_I1
Notice the 'TABLE ACCESS BY ROWID' section. This indicates that the table data
is not being accessed via a FTS operation but rather by a rowid lookup. In this
case the rowid has been produced by looking up values in the index first.
The index is being accessed by an 'INDEX UNIQUE SCAN' operation. This is
explained below. The index name in this case is EMP_I1.
If all the required data resides in the index then a table lookup may be
unnecessary and all you will see is an index access with no table access.
In the following example all the columns (empno) are in the index. Notice that
no table access takes place:
SQL> explain plan for select empno from emp where empno=10;
Query Plan
------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
INDEX UNIQUE SCAN EMP_I1
Indexes are presorted so sorting may be unecessary if the sort order required
is the same as the index.
e.g.
SQL> explain plan for
select empno,ename
from emp
where empno > 7876 order by empno;
Query Plan
--------------------------------------------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
TABLE ACCESS BY ROWID EMP [ANALYZED]
INDEX RANGE SCAN EMP_I1 [ANALYZED]
In this case the index is sorted so ther rows will be returned in the order of
the index hence a sort is unecessary.
explain plan for
select /*+ Full(emp) */ empno,ename
from emp
where empno> 7876 order by empno;
Query Plan
--------------------------------------------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=9
SORT ORDER BY
TABLE ACCESS FULL EMP [ANALYZED] Cost=1 Card=2 Bytes=66
Because we have forced a FTS the data is unsorted and so we must sort the data
after it has been retrieved.
There are 5 methods of index lookup:
index unique scan
index range scan
index full scan
index fast full scan
index skip scan
Index unique scan
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, However this may return > 1 row as the uniqueness will not be
guaranteed.
example explain plan:
SQL> explain plan for
select empno,ename from emp where empno=10;
Query Plan
------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
TABLE ACCESS BY ROWID EMP [ANALYZED]
INDEX UNIQUE SCAN EMP_I1
Index range scan
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 )
e.g.
SQL> explain plan for
select empno,ename from emp
where empno > 7876 order by empno;
Query Plan
--------------------------------------------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
TABLE ACCESS BY ROWID EMP [ANALYZED]
INDEX RANGE SCAN EMP_I1 [ANALYZED]
A non-unique index may return multiple values for the predicate
col1 = 5 and will use an index range scan
SQL> explain plan for select mgr from emp where mgr = 5
Query plan
--------------------
SELECT STATEMENT [CHOOSE] Cost=1
INDEX RANGE SCAN EMP_I2 [ANALYZED]
Index Full Scan
In certain circumstances it is possible for the whole index to be scanned as
opposed to a range scan (i.e. where no constraining predicates are provided for
a table).
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.
The optimizer may decide that selecting all the information from the index
and not sorting is more efficient than doing a FTS or a Fast Full Index Scan
and then sorting.
An Index full scan will perform single block i/o's and so it may prove to be
inefficient.
e.g.
Index BE_IX is a concatenated index on big_emp (empno,ename)
SQL> explain plan for
select empno,ename from big_emp order by empno,ename;
Query Plan
--------------------------------------------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=26
INDEX FULL SCAN BE_IX [ANALYZED]
Index Fast Full Scan
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.
Note that INDEX FAST FULL SCAN is the mechinism behind fast index create
and recreate.
e.g.
Index BE_IX is a concatenated index on big_emp (empno,ename)
SQL> explain plan for select empno,ename from big_emp;
Query Plan
------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
INDEX FAST FULL SCAN BE_IX [ANALYZED]
Selecting the 2nd column of concatenated index:
SQL> explain plan for select ename from big_emp;
Query Plan
------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
INDEX FAST FULL SCAN BE_IX [ANALYZED]
Index skip scan
---------------
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.
The next example checks ename='SMITH' for each index key
even though ename is not the leading column of the index. The leading column
(empno) is skipped.
create index i_emp on emp(empno, ename);
select /*+ index_ss(emp i_emp)*/ job from emp where ename='SMITH';
PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost |
---------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 12 | 61 |
| 1 | TABLE ACCESS BY INDEX ROWID| EMP | 1 | 12 | 61 |
|* 2 | INDEX SKIP SCAN | I_EMP | 1 | | 11 |
---------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("EMP"."ENAME"='SMITH')
filter("EMP"."ENAME"='SMITH')
Rowid
This is the quickest access method available.
Oracle retrieves the specified block and extracts the rows it is
interested in.
Most frequently seen in explain plans as Table access by Rowid
Access by rowid :
SQL> explain plan for select * from dept where rowid = ':x';
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=1 Card=1 Bytes=18)
1 0 TABLE ACCESS (BY USER ROWID) OF 'DEPT' (Cost=1 Card=1 Bytes=18)
(BY USER ROWID) shows that the user provided the rowid.
Table is accessed by rowid following index lookup:
SQL> explain plan for
select empno,ename from emp where empno=10;
Execution Plan
----------------------------------------------------------
0 SELECT STATEMENT Optimizer=CHOOSE (Cost=1 Card=1 Bytes=19)
1 0 TABLE ACCESS (BY INDEX ROWID) OF 'EMP' (Cost=1 Card=1 Bytes=19)
2 1 INDEX (RANGE SCAN) OF 'EMP_PK' (UNIQUE) (Cost=1 Card=1)
(BY INDEX ROWID) shows that the rowid comes from the index access in the
step with ID = 2.
Joins
=====
A Join is a predicate that attempts to combine 2 row sources
We only ever join 2 row sources together:
At a given time only one Join Step is performed even though underlying
row sources may have been accessed in parallel. Once 2 row sources have been
the resultant row source can start to be joined it to other row sources.
Note that some join methods (such as nested loops) allow a batch of
fetched rows to be returned to the higher levels before fetching the next batch.
Join order - order in which joins are performed
The join order makes a significant difference to the way in which the
query is executed. By accessing particular row sources first, certain
predicates may be satisfied that are not satisfied by with other join orders.
This may prevent certain access paths from being taken.
e.g. Suppose there is a concatenated index on A(a.col1,a.col2)
Note that a.col1 is the leading column.
Consider the following query:
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
We could represent the joins present in the query using the following scehmatic:
B <---> A <---> C
col3=10 col3=5
There are really only 2 ways we can drive the query: via B.col3 or C.col3.
We would have to do a Full scan of A to be able to drive off it. This is
unlikely to be efficient with large tables;
If we drive off table B, using predicate B.col3=10 (as a filter or lookup key)
then we will retrieve the value for B.col1 and join to A.col1. Because we have
now filled the leading column of the concatenated index on table A we can use
this index to give us values for A.col2 and join to A.
However if we drive of table c, then we only get a value for a.col2 and since
this is a trailing column of a concatenated index and the leading column has
not been supplied at this point, we cannot use the index on a to lookup the
data.
So it is likely that the best join order will be B A C.
The CBO will obviously use costs to establish whether the individual access
paths are a good idea or not.
If the CBO does not choose this join order then we can hint it by changing the
from clause to read:
from B,A,C
and using the /*+ ordered */ hint. The resultant query would be:
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
Join Types
Sort Merge Join (SMJ)
Nested Loops (NL)
Hash Join
Sort Merge Join
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
Sorted rows from both sides are then merged together (joined)
MERGE
/ \
SORT SORT
| |
Row Source 1 Row Source 2
If the row sources are already (known to be) sorted then the sort operation is
unecessary as long as both 'sides' are sorted using the same key.
Presorted row sources include indexed columns and row sources that have already
been sorted in earlier steps.
Although the merge of the 2 row sources is handled serially, the row sources
could be accessed in parallel.
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.
Nested Loops
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
Row source 1
Row 1 -------------- -- Probe -> Row source 2
Row 2 -------------- -- Probe -> Row source 2
Row 3 -------------- -- Probe -> Row source 2
Row source 1 is known as the outer table
Row source 2 is known as the inner table
Accessing row source 2 is known a probing the inner table
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]
Hash Join
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
Cartesian Product
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.
Operations
==========
Operations that show up in explain plans
sort
filter
view
Sorts
There are a number of different operations that promote sorts
order by clauses
group by
sort merge join
Note that if the row source is already appropriately sorted then no sorting is
required. This is now indicated in 7.3:
SORT GROUP BY NOSORT
INDEX FULL SCAN .....
In this case the group by operation simply groups the rows it does not do the
sort operation as this has already been completed.
Sorts are expensive operations especially on large tables where the rows do
not fit in memory and spill to disk. By default sort blocks are placed into the
buffer cache. This may result in aging out of other blocks that may be reread
by other processes. To avoid this you can use the parameter
<Parameter:SORT_DIRECT_WRITES> which does not place sort blocks into the buffer
cache.
Filter
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
In this example there are 2 filter steps. The first is effectively like a NL
except that it stops when it gets something that it doesn't like
(i.e. a bounded NL). This is there because of the not in.
The second is filtering out the min value:
SQL> explain plan for
select *
from emp
where empno not in (select min(empno) from big_emp group by empno);
Query Plan
------------------
SELECT STATEMENT [CHOOSE] Cost=1
FILTER This is like a bounded nested loops
TABLE ACCESS FULL EMP [ANALYZED]
FILTER This filter is introduced by the min
SORT GROUP BY NOSORT
INDEX FULL SCAN BE_IX
This example is also interesting in that it has a NOSORT function. The group
by does not need to sort because the index row source is already pre sorted.
Views
=====
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.
In the following example the select contains an inline view which cannot be
merged:
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
In this case the inline view tmp which contains an aggregate function cannot be
merged into the main query. The explain plan shows this as a view step.
Sometimes genreated VIEWs can be seen in the execution plan:
VIEW in the FROM clause (INLINE VIEW)
-------------------------------------
If a view cannot be merged into the main query then a VIEW keyword will
be shown in the plan to represent this.
explain plan for
select ename,tot
from emp,
(select empno x, sum(empno) tot from big_emp group by empno)
where emp.empno = x;
----------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 16 | 720 | | 46 |
|* 1 | HASH JOIN | | 16 | 720 | | 46 |
| 2 | TABLE ACCESS FULL | EMP | 16 | 304 | | 1 |
| 3 | VIEW | | 10000 | 253K| | 44 |
| 4 | SORT GROUP BY | | 10000 | 30000 | 248K| 44 |
| 5 | TABLE ACCESS FULL| BIG_EMP | 10000 | 30000 | | 24 |
----------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - access("EMP"."EMPNO"="from$_subquery$_002"."X")
"from$_subquery$_002" is the generated name of the inline view that is
produced in id=3 line. If the inline view was aliased in the from clause then
the system generated name ("from$_subquery$_002") would be replaced by this
alias.
Subquery isunnested but resultant view is not merged:
-----------------------------------------------------
In preoptimization there is a stage where subqueries can be unnested (combined
with base query). There are numerous and complex ruled governing this activity
which are beyond the scope of this article. However, if successful, a subquery
may be unnested to produce an inline view which represents the subquery. This
view is subject to view merging. If the view is non-mergeable then a VIEW keyword
will appear in the plan.
explain plan for
select ename
from emp
where emp.empno in (select empno x from big_emp group by empno);
----------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost |
----------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 16 | 512 | | 46 |
|* 1 | HASH JOIN SEMI | | 16 | 512 | | 46 |
| 2 | TABLE ACCESS FULL | EMP | 16 | 304 | | 1 |
| 3 | VIEW | VW_NSO_1 | 10000 | 126K| | 44 |
| 4 | SORT GROUP BY | | 10000 | 30000 | 248K| 44 |
| 5 | TABLE ACCESS FULL| BIG_EMP | 10000 | 30000 | | 24 |
----------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - access("EMP"."EMPNO"="VW_NSO_1"."X")
In this example, the subquery is unnested but the resultant inline view
cannot be merged due to the 'group by'. This unmergeable view is given the
system generated name of "VW_NSO_1".
Note: In later versions, as with any code, certain inline views may be mergeable following
code improvements or as new features are introduced.
Partition Views
===============
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.
Allows a large table to be broken up into a number of smaller partitions
which can be queried much more quickly than the table as a whole
a union all view is built over the top to provide the original functionality
Check constraints or where clauses provide partition elimination capabilities
SQL> explain plan for
select /*+ use_nl(p1,kbwyv1) ordered */ sum(prc_pd)
from parent1 p1, kbwyv1
where p1.class = 22
and kbwyv1.bitm_numb = p1.bitm_numb
and kbwyv1.year = 1997
and kbwyv1.week between 32 and 33 ;
Query Plan
-----------------------------------------
SELECT STATEMENT [FIRST_ROWS] Cost=1780
SORT AGGREGATE
NESTED LOOPS [:Q65001] Ct=1780 Cd=40 Bt=3120
TABLE ACCESS FULL PARENT1 [:Q65000] [AN] Ct=20 Cd=40 Bt=1040
VIEW KBWYV1 [:Q65001]
UNION-ALL PARTITION [:Q65001]
FILTER [:Q64000]
TABLE ACCESS FULL KBWYT1 [:Q64000] [AN] Ct=11 Cd=2000 Bt=104000
TABLE ACCESS FULL KBWYT2 [:Q63000] [AN] Ct=11 Cd=2000 Bt=104000
TABLE ACCESS FULL KBWYT3 [:Q62000] [AN] Ct=11 Cd=2000 Bt=104000
FILTER [:Q61000]
TABLE ACCESS FULL KBWYT4 [:Q61000] [AN] Ct=11 Cd=2000 Bt=104000
KBWYV1 is a view on 4 tables KBWYT1-4.
KBWYT1-4 contain rows for week 31-34 respectively and are maintained by check
constraints.
This query should only return rows from partions 2 & 3. The filter operation
indicates this. Partitions 1 & 4 are eliminated at execution time.
The view line indicates that the view is not merged. The union-all partion
information indicates that we have recognised this as a partition view.
Note that the tables can be accessed in parallel.
Partition handling
==================
The examples are based on a simple range partitioned table:
create table range_part (col1 number(9))
partition by range (col1)
(partition p1 values less than (10) ,
partition p2 values less than (20) ,
partition p3 values less than(30) ,
partition p4 values less than (MAXVALUE)) ;
Eliminated partitions can be determined at parse time
=====================================================
As optimizer can check the partition ranges in the dictionary, in some
cases it knows which partitions should be used and the others are
"elliminated" when the select is optimized.
This activity is reflected in the following plan_table columns:
PARTITION_START
PARTITION_STOP
Example:
explain plan for select * from range_part where col1=15;
start $ORACLE_HOME/rdbms/admin/utlxpls.sql
------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost | Pstart| Pstop |
------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 13 | 1 | | |
|* 1 | TABLE ACCESS FULL | RANGE_PART | 1 | 13 | 1 | 2 | 2 |
------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("RANGE_PART"."COL1"=15)
The Pstart and Pstop columns show that the optimizer believes that
only partition 2 is required to satisfy this query.
Eliminated partitions can only be determined at run time
========================================================
In this example, since the partition key references a bind variable,
the actual partition numbers cannot be determined during optimization.
The optimizer can only determine that there is a possibility that
partitions can be eliminated at runtime. The
PARTITION_START and PARTITION_START columns in the plan table will contain
either KEY or ROW LOCATION depending on the nature of the query and the partition
itself.
Example:
explain plan for select * from range_part where col1= :x;
start $ORACLE_HOME/rdbms/admin/utlxpls.sql
--------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost | Pstart| Pstop |
--------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 13 | 1 | | |
| 1 | PARTITION RANGE SINGLE| | | | | KEY | KEY |
|* 2 | TABLE ACCESS FULL | RANGE_PART | 1 | 13 | 1 | KEY | KEY |
--------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - filter("RANGE_PART"."COL1"=TO_NUMBER(:Z))
You can see more examples in
Remote Queries
==============
Only shows remote in the OPERATION column
OTHER column shows query executed on remote node
OTHER_NODE shows where it is executed
Different operational characteristics for RBO & CBO
RBO - Drags everything across the link and joins locally
CBO - Uses cost estimates to determine whether to execute remotely or locally
SQL> explain plan for
select *
from dept@loop_link;
Query Plan
-------------------------------------------------------
SELECT STATEMENT REMOTE [CHOOSE] Cost=1
TABLE ACCESS FULL DEPT [SJD.WORLD] [ANALYZED]
In this case the whole query has been sent to the remote site. The other column
shows nothing.
SQL> explain plan for
select a.dname,avg(b.sal),max(b.sal)
from dept@loop_link a, emp b
where a.deptno=b.deptno
group by a.dname
order by max(b.sal),avg(b.sal) desc;
Query Plan
-----------------------------------------------------
SELECT STATEMENT [CHOOSE] Cost=20
SORT ORDER BY [:Q137003] [PARALLEL_TO_SERIAL]
SORT GROUP BY [:Q137002] [PARALLEL_TO_PARALLEL]
NESTED LOOPS [:Q137001] [PARALLEL_TO_PARALLEL]
REMOTE [:Q137000] [PARALLEL_FROM_SERIAL]
TABLE ACCESS FULL EMP [:Q137001] [ANALYZED]
[PARALLEL_COMBINED_WITH_PARENT]
OTHER (for REMOTE)
----------------------------------------------------------------
SELECT "DEPTNO","DNAME" FROM "DEPT" A
Bind Variables
==============
Bind variables are recommended in most cases because they promote sharing of
sql code
During optimization, the optimizer has NO IDEA what the bind variable contains.
With RBO this makes no difference but with CBO, which relies on accurate
statistics to produce plans, this can be a problem.
To check the execution plan of a query using bind variables in sqlplus, the
following can be used:
variable x varchar2(18);
begin :x := 'hello';
end;
/
SQL> explain plan for
select * from dept where rowid = ':x';
Query Plan
------------------------------------
SELECT STATEMENT [CHOOSE] Cost=1
TABLE ACCESS BY ROWID DEPT [ANALYZED]
Parallel Query
==============
Main indicators that a query is using PQO:
o [:Q1000004] entries in the explain plan
o Checkout the other column for details of what the slaves are executing
o v$pq_slave will show any parallel activity
Columns to look in for information
other - contains the query passed to the slaves
other_tag - describes the contents of other
object_node - indicates order of pqo slaves
Parallel Query operates on a producer/consumer basis.
When you specify parallel degree 4 oracle tries to allocate 4 producer slaves
and 4 consumer slaves. The producers can feed any of the consumers.
If there are only 2 slaves available then we use these.
If there is only 1 slave available then we go serial
If there are none available then we use serial.
If parallel_min_percent is set then we error ora 12827 instead of using a lower
number of slaves or going serial
Consumer processes typically perform a sorting function. If there is no
requirement for the data to be sorted then the consumer slaves are not produced
and we end up with the number of slaves used matching the degree of parallelism
as opposed to being 2x the degree.
Parallel Terms
PARALLEL_FROM_SERIAL This means that source of the data is serial
but it is passed to a parallel consumer
PARALLEL_TO_PARALLEL Both the consumer and the producer are
parallel
PARALLEL_COMBINED_WITH_PARENT This operation has been combined with the
parent operator. For example in a sort merge
join the sort operations would be shown
as PARALLEL_COMBINED_WITH_PARENT because the
sort and the merge are handled as 1 operation.
PARALELL_TO_SERIAL The source of the data is parallel but it is
passed to a serial consumer.
This typically will happen at the top of the
explain plan but could occur anywhere.
EXAMPLES OF PARALLEL QUERIES
Assumptions
OPTIMIZER_MODE = CHOOSE
DEPT is small compared to EMP
DEPT has an index (DEPT_INDX) on deptno column
Three examples are presented
Query #1 Serial
Query #2 Parallel
Query #3 Parallel, with forced optimization to
RULE and forced usage of DEPT_INDX
Sample Query #1 (Serial)
========================
select A.dname, avg(B.sal), max(B.sal)
from dept A, emp B
where A.deptno = B.deptno
group by A.dname
order by max(B.sal), avg(B.sal) desc;
Execution Plan #1 (Serial)
OBJECT_NAME OBJECT_NODE OTHER
------------------------------- ----------- -------
SELECT STATEMENT
SORT ORDER BY
SORT GROUP BY
MERGE JOIN
SORT JOIN
TABLE ACCESS FULL emp
SORT JOIN
TABLE ACCESS FULL dept
Notice that the object_node and other columns are empty
Sample Query #2 (Query #1 with parallel hints)
==============================================
select /*+ parallel(B,4) parallel(A,4) */
A.dname, avg(B.sal), max(B.sal)
from dept A, emp B
where A.deptno = B.deptno
group by A.dname
order by max(B.sal), avg(B.sal) desc;
Execution Plan #2 (Parallel)
OBJECT_NAME OBJECT_NODE OTHER
------------------------------- ----------- -------
SELECT STATEMENT Cost = ??
SORT ORDER BY :Q55004 [7]
SORT GROUP BY :Q55003 [6]
MERGE JOIN :Q55002 [5]
SORT JOIN :Q55002 [4]
TABLE ACCESS FULL emp :Q55001 [2]
SORT JOIN :Q55002 [3]
TABLE ACCESS FULL dept :Q55000 [1]
Execution Plan #2 -- OTHER column
[1] (:Q55000) "PARALLEL_FROM_SERIAL"
Serial execution of
SELECT DEPTNO, DNAME FROM DEPT
[2] (:Q55001) "PARALLEL_TO_PARALLEL"
SELECT /*+ ROWID(A1)*/
A1."DEPTNO" C0, A1."SAL" C1
FROM "EMP" A1
WHERE ROWID BETWEEN :1 AND :2
[3] (:Q55002) "PARALLEL_COMBINED_WITH_PARENT"
[4] (:Q55002) "PARALLEL_COMBINED_WITH_PARENT"
[5] (:Q55002) "PARALLEL_TO_PARALLEL"
SELECT /*+ ORDERED USE_MERGE(A2)*/
A2.C1 C0, A1.C1 C1
FROM :Q55001 A1,:Q55000 A2
WHERE A1.C0=A2.C0
[6] (:Q55003) "PARALLEL_TO_PARALLEL"
SELECT MAX(A1.C1) C0, AVG(A1.C1) C1, A1.C0 C2
FROM :Q55002 A1
GROUP BY A1.C0
[7] (:Q55004) "PARALLEL_FROM_SERIAL"
SELECT A1.C0 C0, A1.C1 C1, A1.C2 C2
FROM :Q55003 A1
ORDER BY A1.CO, A1.C1 DESC
Sample Query #3 (Query #2 with fudged hints)
============================================
select /*+ index(A dept_indx) parallel(B,4) parallel(A,4) */
A.dname, avg(B.sal), max(B.sal)
from dept A, emp B
where A.deptno = B.deptno
group by A.dname
order by max(B.sal), avg(B.sal) desc;
Execution Plan #3 (Parallel)
OBJECT_NAME OBJECT_NODE OTHER
----------------------------------- ----------- -------
SELECT STATEMENT Cost = ??
SORT ORDER BY :Q58002 [6]
SORT GROUP BY :Q58001 [5]
NESTED LOOPS JOIN :Q58000 [4]
TABLE ACCESS FULL emp :Q58000 [3]
TABLE ACCESS BY ROWID dept :Q58000 [2]
INDEX RANGE SCAN dept_indx :Q58000 [1]
Execution Plan #3 -- OTHER column
[1] (:Q58000) "PARALLEL_COMBINED_WITH_PARENT"
[2] (:Q58000) "PARALLEL_COMBINED_WITH_PARENT"
[3] (:Q58000) "PARALLEL_COMBINED_WITH_PARENT"
[4] (:Q58000) "PARALLEL_TO_PARALLEL"
SELECT /*+ ORDERED USE_NL(A2) INDEX(A2) */
A2."DNAME" C0, A1.C0 C1
FROM
(SELECT /*+ ROWID(A3) */
A3."SAL" CO, A3."DEPTNO" C1
FROM "EMP" A3
WHERE ROWID BETWEEN :1 AND :2) A1,
"DEPT" A2
WHERE A2."DEPTNO" = A1.C1
[5] (:Q58001) "PARALLEL_TO_PARALLEL"
SELECT MAX(A1.C1) C0, AVG(A1.C1) C1, A1.C0 C2
FROM :Q58000 A1
GROUP BY A1.C0
[6] (:Q58002) "PARALLEL_TO_SERIAL"
SELECT A1.C0 C0, A1.C1 C1, A1.C2 C2
FROM :Q58001 A1
ORDER BY A1.C0, A1.C1 DESC
How to obtain explain plans
===========================
Explain plan for
Main advantage is that it does not actually run the query - just parses the
sql. This means that it executes quickly.
In the early stages of tuning explain plan gives you an idea of
the potential performance of your query without actually running it.
You can then make a judgement as to any modifications you may choose to make.
Autotrace
Autotrace can be configured to run the sql & gives a plan and statistics
afterwards or just give you an explain plan without executing the query.
Tkprof
analyzes trace files
The exact methods are explained in:
Note:235530.1 Recommended Method for Obtaining a Formatted Explain Plan
Note:31101.1 SCRIPT: DIAG - Obtaining Formatted Explain Plan Output
Note.41634.1 TKPROF and Problem Solving
REFERENCES:
==============
Note:33838.1 Determining the execution plan for a distributed query
Note:236026.1 Display Execution Plans with dbms_xplan.display
Note:235530.1 Recommended Method for Obtaining a Formatted Explain Plan
Note:31101.1 SCRIPT: DIAG - Obtaining Formatted Explain Plan Output
Note.41634.1 TKPROF and Problem Solving
Note.209197.1 Using Statspack to Record Explain Plan Details
Note.237287.1 How To Verify Parallel Execution is running
Note:179518.1 Partition Pruning and Joins
No comments:
Post a Comment