为什么需要SQL Profile

为什么需要SQL Profile

Why oracle need SQL Profiles,how it work and what are SQL Profiles...

使用DBMS_XPLAN.DISPLAY分析SQL执行计划，通常会看到Note中有类似下面这样的提示；

Note
-----
 
   - SQL profile "SYS_SQLPROF_0158283a9b920000" used for this statement

SQL profile由人为手工创建或在Automatic SQL Tunning阶段由SQL tuning advisor创建，它看起来有如下的意思:

1. 在优化器评估SQL时使用了额外的对象帮助完成评估；
2. 对象改变了优化器原先的评估计划；

当看到这些信息，比较关心的是这个对象(SLQ profile)是什么?它做了什么?是否真的需要它?带着这些疑问学习和探索，最终决解了遇到的问题。

SQL> @i

USERNAME             INST_NAME            HOST_NAME                 SID   SERIAL#  VERSION    STARTED  SPID       OPID  CPID            SADDR            PADDR
-------------------- -------------------- ------------------------- ----- -------- ---------- -------- ---------- ----- --------------- ---------------- ----------------
OPS$SYWU             sydb                 sywu.com                  288   22197    11.2.0.4.0 20160421 13736      46    3392:1312       0000000071FE0DA0 0000000072149F40

遇到的问题

假设有这样一张类似订单的表orders；

create table orders(order_no,order_date)
as
select
    level,cast(sysdate-level/24 as date) 
from 
    dual
connect by level<=5E5;

SQL> @desc orders
           Name                            Null?    Type
           ------------------------------- -------- ----------------------------
    1      ORDER_NO                                 NUMBER
    2      ORDER_DATE                               DATE

保存订单信息，order_date上创建了索引。

create index idx_orders_dt on orders(order_date);

在交易中可能经常遇到某些原因导致交易延期的情况，为了测试这个问题，开发人员添加了未来某一天这样的日期值测试；这里用一个清晰的时间来代替未来的日期；

INSERT INTO ORDERS VALUES (-1, DATE '9999-01-01');

和正常使用的一样，该表定期收集了统计信息；

exec dbms_stats.gather_table_stats(user,'orders', cascade => true);

当系统查询当天的交易记录时发现优化器使用全表扫描，并非索引扫描；

-------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation         | Name   | Starts | E-Rows |E-Bytes| Cost (%CPU)| E-Time   | A-Rows |   A-Time   | Buffers | Reads  |
-------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |        |      1 |        |       |   130 (100)|          |     10 |00:00:00.23 |     329 |    323 |
|*  1 |  TABLE ACCESS FULL| ORDERS |      1 |    496K|  6302K|   130  (26)| 00:00:02 |     10 |00:00:00.23 |     329 |    323 |
-------------------------------------------------------------------------------------------------------------------------------

Query Block Name / Object Alias (identified by operation id):
-------------------------------------------------------------

   1 - SEL$1 / ORDERS@SEL$1

Outline Data
-------------

  /*+
      BEGIN_OUTLINE_DATA
      IGNORE_OPTIM_EMBEDDED_HINTS
      OPTIMIZER_FEATURES_ENABLE('11.2.0.4')
      DB_VERSION('11.2.0.4')
      ALL_ROWS
      OUTLINE_LEAF(@"SEL$1")
      FULL(@"SEL$1" "ORDERS"@"SEL$1")
      END_OUTLINE_DATA
  */

Predicate Information (identified by operation id):
---------------------------------------------------

   1 - filter("ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))

Column Projection Information (identified by operation id):
-----------------------------------------------------------

   1 - "ORDERS"."ORDER_NO"[NUMBER,22], "ORDER_DATE"[DATE,7]

显然对于这样一张交易记录表，实际当天的记录数据只占全表数据量的4.1%左右，使用索引扫描的方式开销小于全表扫描，但优化器对范围评估错误。接着使用DBMS_SQLTUNE分析SQL；

var task_name varchar2(30)

BEGIN
    :task_name := DBMS_SQLTUNE.CREATE_TUNING_TASK(
         sql_text    => 'select * from orders where order_date>=trunc(sysdate,''DD'')',
         user_name   => user,
         scope       => 'COMPREHENSIVE',
         time_limit  => 60,
         task_name   => 'orders_tuning_task');
END;
/

执行分析；

alter session set events '10046 trace name context forever,level 12';
exec DBMS_SQLTUNE.execute_tuning_task(task_name => :task_name);

分析结果；

col REPORT_TUNING format a200
select
    dbms_sqltune.report_tuning_task(:task_name) REPORT_TUNING
from
    dual;

REPORT_TUNING
----------------------------------------------------------------------------------------------------------------------
GENERAL INFORMATION SECTION
-------------------------------------------------------------------------------
Tuning Task Name   : orders_tuning_task
Tuning Task Owner  : OPS$SYWU
Workload Type      : Single SQL Statement
Scope              : COMPREHENSIVE
Time Limit(seconds): 60
Completion Status  : COMPLETED
Started at         : 11/07/2016 21:43:25
Completed at       : 11/07/2016 21:43:27

-------------------------------------------------------------------------------
Schema Name: OPS$SYWU
SQL ID     : 9ybj4xdc5hsrb
SQL Text   : select * from orders where order_date>=trunc(sysdate,'DD')

-------------------------------------------------------------------------------
FINDINGS SECTION (1 finding)
-------------------------------------------------------------------------------

1- SQL Profile Finding (see explain plans section below)
--------------------------------------------------------
  A potentially better execution plan was found for this statement.

  Recommendation (estimated benefit: 98.78%)
  ------------------------------------------
  - Consider accepting the recommended SQL profile.
    execute dbms_sqltune.accept_sql_profile(task_name =>
            'orders_tuning_task', task_owner => 'OPS$SYWU', replace => TRUE);

  Validation results
  ------------------
  The SQL profile was tested by executing both its plan and the original plan
  and measuring their respective execution statistics. A plan may have been
  only partially executed if the other could be run to completion in less time.

                           Original Plan  With SQL Profile  % Improved
                           -------------  ----------------  ----------
  Completion Status:            COMPLETE          COMPLETE
  Elapsed Time (s):             .041546           .000132      99.68 %
  CPU Time (s):                 .029895             .0001      99.66 %
  User I/O Time (s):            .015204           .000032      99.78 %
  Buffer Gets:                      328                 4      98.78 %
  Physical Read Requests:            45                 0        100 %
  Physical Write Requests:            0                 0
  Physical Read Bytes:         10682368              9830       99.9 %
  Physical Write Bytes:               0                 0
  Rows Processed:                    10                10
  Fetches:                           10                10
  Executions:                         1                 1

  Notes
  -----
  1. Statistics for the original plan were averaged over 10 executions.
  2. Statistics for the SQL profile plan were averaged over 10 executions.

-------------------------------------------------------------------------------
EXPLAIN PLANS SECTION
-------------------------------------------------------------------------------

1- Original With Adjusted Cost
------------------------------
Plan hash value: 1275100350

----------------------------------------------------------------------------
| Id  | Operation         | Name   | Rows  | Bytes | Cost (%CPU)| Time     |
----------------------------------------------------------------------------
|   0 | SELECT STATEMENT  |        |    18 |   234 |   130  (26)| 00:00:02 |
|*  1 |  TABLE ACCESS FULL| ORDERS |    18 |   234 |   130  (26)| 00:00:02 |
----------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   1 - filter("ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))

2- Using SQL Profile
--------------------
Plan hash value: 3364688013

---------------------------------------------------------------------------------------------
| Id  | Operation                   | Name          | Rows  | Bytes | Cost (%CPU)| Time     |
---------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT            |               |    18 |   234 |     3   (0)| 00:00:01 |
|   1 |  TABLE ACCESS BY INDEX ROWID| ORDERS        |    18 |   234 |     3   (0)| 00:00:01 |
|*  2 |   INDEX RANGE SCAN          | IDX_ORDERS_DT |    10 |       |     2   (0)| 00:00:01 |
---------------------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))

-------------------------------------------------------------------------------

oracle通过分析发现了问题，产生了新的执行计划，并对比两个执行计划，新的执行计划改善90%+的性能，并且改善性能问题只需要同意使用SQL Profile即可；然后允许数据库使用SQL Profile。

exec dbms_sqltune.ACCEPT_SQL_PROFILE(TASK_NAME=>:task_name);

再次执行SQL时，优化器使用了SQL Profile和新的执行计划。

select * from orders where order_date>=trunc(sysdate,'DD');

SQL_ID  3zcvw1pxfcypm, child number 0
-------------------------------------
select * from orders where order_date>=trunc(sysdate,'DD')

Plan hash value: 3364688013

---------------------------------------------------------------------------------------------------------------------------------------
| Id  | Operation                   | Name          | Starts | E-Rows |E-Bytes| Cost (%CPU)| E-Time   | A-Rows |   A-Time   | Buffers |
---------------------------------------------------------------------------------------------------------------------------------------
|   0 | SELECT STATEMENT            |               |      1 |        |       |     3 (100)|          |     10 |00:00:00.01 |       6 |
|   1 |  TABLE ACCESS BY INDEX ROWID| ORDERS        |      1 |     18 |   234 |     3   (0)| 00:00:01 |     10 |00:00:00.01 |       6 |
|*  2 |   INDEX RANGE SCAN          | IDX_ORDERS_DT |      1 |     10 |       |     2   (0)| 00:00:01 |     10 |00:00:00.01 |       3 |
---------------------------------------------------------------------------------------------------------------------------------------

Query Block Name / Object Alias (identified by operation id):
-------------------------------------------------------------

   1 - SEL$1 / ORDERS@SEL$1
   2 - SEL$1 / ORDERS@SEL$1

Outline Data
-------------

  /*+
      BEGIN_OUTLINE_DATA
      IGNORE_OPTIM_EMBEDDED_HINTS
      OPTIMIZER_FEATURES_ENABLE('11.2.0.4')
      DB_VERSION('11.2.0.4')
      ALL_ROWS
      OUTLINE_LEAF(@"SEL$1")
      INDEX_RS_ASC(@"SEL$1" "ORDERS"@"SEL$1" ("ORDERS"."ORDER_DATE"))
      END_OUTLINE_DATA
  */

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access("ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))

Column Projection Information (identified by operation id):
-----------------------------------------------------------

   1 - "ORDERS"."ORDER_NO"[NUMBER,22], "ORDER_DATE"[DATE,7]
   2 - "ORDERS".ROWID[ROWID,10], "ORDER_DATE"[DATE,7]

Note
-----
 
   - SQL profile SYS_SQLPROF_01582d15092f0001 used for this statement


SQL>  @sql 2061925043
Show SQL text, child cursors and execution stats for SQL hash value 2061925043 child OPS$SYWU@sydb_111 report

HASH_VALUE   CH#  PLAN_HASH SQL_TEXT                                                                                                       SQL_PROFILE
---------- ----- ---------- -------------------------------------------------------------------------------------------------------------- ------------------------------
2061925043     0 3364688013 select * from orders where order_date>=trunc(sysdate,'DD')                                               SYS_SQLPROF_01582d15092f0001 


  CH# PARENT_HANDLE    OBJECT_HANDLE        PARSES   H_PARSES EXECUTIONS    FETCHES ROWS_PROCESSED       LIOS       PIOS      SORTS     CPU_MS     ELA_MS USERS_EXECUTING
----- ---------------- ---------------- ---------- ---------- ---------- ---------- -------------- ---------- ---------- ---------- ---------- ---------- ---------------
    0 0000000064618858 0000000063A03108          1          3          1          2             10         16          6          0      7.999      8.621               0

Oracle 分析背后做了什么

很惊奇，为什么分析后优化器就能找出问题所在，此时焦点都集中在trace文件了；分析trace文件，发现如下信息；

SELECT /* DS_SVC */ /*+ dynamic_sampling(0) no_sql_tune no_monitoring
  optimizer_features_enable(default) opt_param('parallel_execution_enabled',
  'false')  result_cache */ COUNT(C1)
FROM
 (SELECT /*+ qb_name("innerQuery") NO_INDEX_FFS( "ORDERS")  */ 1 AS C1 FROM
  "ORDERS" SAMPLE BLOCK(56.8182, 2) SEED(1)  "ORDERS") innerQuery

SELECT /* DS_SVC */ /*+ dynamic_sampling(0) no_sql_tune no_monitoring
  optimizer_features_enable(default) opt_param('parallel_execution_enabled',
  'false')  result_cache */ COUNT(C1)
FROM
 (SELECT /*+ qb_name("innerQuery") INDEX_FFS( "ORDERS" "IDX_ORDERS_DT")  */ 1
  AS C1 FROM "ORDERS" SAMPLE BLOCK(59.5238, 2) SEED(1)  "ORDERS" WHERE (
  "ORDERS".ORDER_DATE IS NOT NULL)) innerQuery

SELECT /* DS_SVC */ /*+ dynamic_sampling(0) no_sql_tune no_monitoring
  optimizer_features_enable(default) opt_param('parallel_execution_enabled',
  'false')  result_cache */ COUNT(C1)
FROM
 (SELECT /*+ qb_name("innerQuery") NO_INDEX_FFS( "ORDERS")  */ 1 AS C1 FROM
  "ORDERS" SAMPLE BLOCK(56.8182, 2) SEED(1)  "ORDERS" WHERE
  ("ORDERS"."ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))) innerQuery

SELECT /* DS_SVC */ /*+ dynamic_sampling(0) no_sql_tune no_monitoring
  optimizer_features_enable(default) opt_param('parallel_execution_enabled',
  'false')  result_cache OPT_ESTIMATE(@"innerQuery", TABLE, "ORDERS",
  SCALE_ROWS=3.545138895e-05) */ C1, C2, C3
FROM
 (SELECT /*+ qb_name("innerQuery") INDEX( "ORDERS" "IDX_ORDERS_DT")  */
  COUNT(*) AS C1, 4294967295 AS C2, COUNT(*) AS C3  FROM "ORDERS" "ORDERS"
  WHERE ("ORDERS"."ORDER_DATE">=TRUNC(SYSDATE@!,'fmdd'))) innerQuery

/* SQL Analyze(156,0) */ select * from orders where order_date>=trunc(sysdate,'DD')

从trace文件分析得出，oracle做了如下的操作:

1. 动态采样分析全表数据(无索引)；
2. 动态采样分析表中非空数据(无索引)；
3. 用全表扫描的方式动态采样分析相关的数据；
4. 用索引扫描的方式动态采样分析相关的数据；
5. 对比旧的和新的执行计划；

Oracle对比旧的和新的执行计划后，将消耗小的执行计划信息保存到SQL profile中。

优化器的问题

通常优化器相信:

1. 数据值分布均匀(比如假设：表列数据中，数值2比数值5一样使用频繁)；
2. 数据行分布均匀(比如假设：没有物理cluster或者数据排序)；
3. 对于范围数据是连续的，没有漏缺。

优化器信任收集的统计信息，这些信息包括表行记录数,distinct value,max/min value,直方图信息；换一种通俗的说法，统计信息捕获了表整体形状数据，但有些低级别的信息丢失了。这种分析对于大多数数据来说往往工作得很好，但实际情况中，不可避免的有违反规则的例外，比如对于一张大表，98%的数据可能以随机的方式分配在整个段中(segment)，剩余的2%的数据可能只集中在几个数据块中；不幸的是收集统计信息时没有记录这些细节；这就引发一个问题，已经有的统计信息不能完全有效的帮助优化器生成正确的执行计划，所以到了这里问题转变为什么可以弥补或纠正这些信息，让优化器面对这种特定的SQL时可以评估正确，生成好的执行计划。

什么是SQL Profile

通过上面的实验大体将SQL Profile定义为：

1. 为特定SQL创建和保存执行计划信息；
2. 使用实际运行时的数据来帮助优化器为特定的SQL评估和生成更好的执行计划；

首先通过dbms_sqltune.create_tunning_task创建任务告诉数据库存在问题的可以改善的SQL，这个操作在11G或以后的版本中可以通过Automatic SQL Tuning在对"most active" SQL 分析时创建；然后运行dbms_sqltune.execute_tuning_task评估，这个过程包括三个主要步骤:

1. 动态采样分析表数据，获取到真实的实数信息(最重要的比如，Cardinality)
2. 提供这些真实的实数信息给优化器，让优化器重新评估；
3. 如果优化器评估出新的执行计划，重复多次运行旧的执行计划和新的执行计划，最后对比性能；

如果优化器试运行得出的结果为:

1. 优化器评估后产生了新的执行计划；
2. 新的执行计划比旧的执行计划性能消耗更小，大幅提升性能；

则表明优化器证明旧的评估对于特定的SQL是错误的，一些低级的数据被分析出，进一步，oracle会将这些信息保存供以后使用。但是如何保存这些信息呢?不可能通过定期性的更新统计信息，因为统计信息不包括这些信息。所以，数据库使用一个独立的对象(SQL Profile)保存SQL和这些(cardinality)信息。oracle 以opt_estimate hints的格式保存cardinality信息；
/+ opt_estimate(table, orders, scale_rows=10) */
或者
/+ opt_estimate(index_scan, orders, IDX_ORDERS_DT, scale_rows=0.001) */
所以如果使用了SQL Profile，评估时默认的cardinality将乘以这些数字，优化器会更真实的查看到表中的数据信息，然后做出评估。