Oracle标量子查询算法如下：
主表返回多少行，子表就要被扫描COUNT(DISTINCT NVL(主表JOIN列,0))这么多次，也就是说Oracle标量子查询带有CACHE功能
大多数国产数据库标量子查询还未实现CACHE功能，算法还停留在主表返回多少行，子表被扫描多少次
标量子查询CACHE功能在一些极致性能优化场景中非常有用，笔者曾利用该特性优化过几百条SQL
今年夏天(2025年)给某头部券商做SQL优化的时候遇到一条SQL跑3.7秒，逻辑读70W，GROUP BY后最终返回4294行数据，要求优化到1秒内
SQL大致如下(由于是证券行业，受限于截图有水印以及保密原则，无法给出完整SQL和执行计划，请谅解)
sql
select ...,b.一个列,d.一个列
from (select c_pa_code,
n_hldcst,
n_hldcst_locl,
c_ml_attr,
c_cury_code,
n_valrate,
c_sec_code,
c_port_code,
d_hold
from gzdb.vh_repurchase
where c_pa_code in
('MRFSJRZC', 'MCHGJRZC', 'YSLX_ZQ', 'YFLX_ZQ', 'JZZB')) a
left join gzdb.vb_port_baseinfo b ---b.c_port_code是唯一的
on a.c_port_code = b.c_port_code
left join gzdb.vb_security c
on c.c_sec_code = a.c_sec_code
and c.c_sec_var in ('CJ', 'HG')
left join (select c_port_code,
d_biz,
n_hldmkv_locl,
row_number() over(partition by c_port_code, d_biz order by c_update_time desc) as r
from gzdb.vn_port_index
where c_idx_code = 'ZCJZ'
and c_port_class = 'NA') d
on a.c_port_code = d.c_port_code
and a.d_hold = d.d_biz
and d.r = 1
where a.d_hold = to_date(:endt, 'yyyy-mm-dd')) a
group by ....a过滤完之后返回31793行数据，a通过join列谓词推入到d，d被扫描了31793次，a nl d之后累计耗时3.5秒，注意SQL总耗时3.7秒
a.c_port_code,a.d_hold有大量重复值，b和d只返回一个列，b.c_port_code是唯一的，d对关联列partition by后又限制只取1行
优化手段就是把b和d改写到标量子查询中(b不是引起性能问题的重点，可以改到标量子查询中，也可以不改，重点是d)
利用标量子查询CACHE功能减少d扫描次数，减少b扫描次数
具体改写后的SQL就不贴了，改写完之后逻辑读从70W降低到20W，SQL执行时间降低到1秒内
注意：
1.如果a.c_port_code,a.d_hold没有大量重复值，改写为标量子查询无法提升性能
2.如果Oracle标量子查询没有CACHE功能，那就只能把d和b封装到函数中，再对函数加上RESULT CACHE功能来减少扫描次数，这就太麻烦了
为了照顾SQL优化基础较弱的同学，这里举个例子演示一下Oracle(11.2.0.4)标量子查询CACHE功能
sql
SQL> select *
from (select owner, object_id
from test02
where owner = 'SYS'
order by object_id)
where rownum <= 3; 2 3 4 5 6
OWNER OBJECT_ID
------------------------------ ----------
SYS 2
SYS 3
SYS 4
SQL> select *
from (select owner, object_id
from test02
where owner = 'PUBLIC'
order by object_id)
where rownum <= 3; 2 3 4 5 6
OWNER OBJECT_ID
------------------------------ ----------
PUBLIC 117
PUBLIC 280
PUBLIC 367
SQL> alter session set statistics_level=all;
Session altered.
SQL> select object_id, (select count(*) from test01 where owner = t2.owner) cnt
from test02 t2
where object_id in (2, 3, 4, 117, 280, 367); 2 3
OBJECT_ID CNT
---------- ----------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
SQL> select * from table(dbms_xplan.display_cursor(null,null,'allstats last'));
PLAN_TABLE_OUTPUT
--------------------------------------------------------------------------------------------------------
SQL_ID dcm994g6n6tq0, child number 3
-------------------------------------
select object_id, (select count(*) from test01 where owner = t2.owner)
cnt from test02 t2 where object_id in (2, 3, 4, 117, 280, 367)
Plan hash value: 384367355
--------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers |
--------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 6 |00:00:00.01 | 15 |
| 1 | SORT AGGREGATE | | 2 | 1 | 2 |00:00:05.60 | 1271K|
|* 2 | TABLE ACCESS FULL | TEST01 | 2 | 421K| 36M|00:00:04.52 | 1271K|
| 3 | INLIST ITERATOR | | 1 | | 6 |00:00:00.01 | 15 |
| 4 | TABLE ACCESS BY INDEX ROWID| TEST02 | 6 | 6 | 6 |00:00:00.01 | 15 |
|* 5 | INDEX RANGE SCAN | IDX_TEST02_ID | 6 | 6 | 6 |00:00:00.01 | 9 |
--------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - filter("OWNER"=:B1)
5 - access(("OBJECT_ID"=2 OR "OBJECT_ID"=3 OR "OBJECT_ID"=4 OR "OBJECT_ID"=117 OR
"OBJECT_ID"=280 OR "OBJECT_ID"=367))主表TEST02过滤后返回6行数据，通过连接列OWNER传值给标量子查询的子表TEST01
TEST02过滤后OWNER列只有2个DISTINCT值：SYS和PUBLIC
执行计划中Starts表示扫描次数，ID=2的STARTS是2，说明Oracle标量子查询有CACHE功能，如果没有CACHE功能，TEST01应该扫描6次
当然了，表TEST01的OWNER列需要创建索引，地球人都知道，我们这里不创建
现在数据库切换到崖山23.5.1
sql
SQL> alter session set statistics_level=all;
Succeed.
SQL> set autot trace
SQL> select object_id, (select count(*) from test01 where owner = t2.owner) cnt
from test02 t2
where object_id in (2, 3, 4, 117, 280, 367); 2 3
Execution Plan
----------------------------------------------------------------
SQL hash value: 2923445856
Optimizer: ADOPT_C
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| Id | Operation type | Name | Owner | E - Rows | A - Rows | Cost(%CPU) | A - Time | Loops | Memory | Disk | Partition info |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| 0 | SELECT STATEMENT | | | | 6| | 90| 7| 0| 0| |
| 1 | SUBQUERY | QUERY[1] | | | 2| | 7889087| 4| 0| 0| |
| 2 | AGGREGATE | | | 1| 2| 1151005( 0)| 7889079| 4| 0| 0| |
|* 3 | TABLE ACCESS FULL | TEST01 | SCOTT | 446| 36757504| 1151005( 0)| 6349946| 36757506| 0| 0| |
| 4 | TABLE ACCESS BY INDEX ROWID | TEST02 | SCOTT | 6| | 1( 0)| | | | | |
|* 5 | INDEX RANGE SCAN | IDX_TEST02_ID | SCOTT | 6| 6| 1( 0)| 88| 7| 0| 0| |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
Operation Information (identified by operation id):
---------------------------------------------------
1 - Subquery NDV info - NDV percentage: 0.500000, NDV Expression: ("T2"."OWNER"[OPTMZ-3])
3 - Predicate : filter("TEST01"."OWNER"[OPTMZ-1] = "T2"."OWNER"[OPTMZ-3][OPTMZ-2])
5 - Predicate : access("T2"."OBJECT_ID" IN (2[OPTMZ-0], 3[OPTMZ-0], 4[OPTMZ-0], 117[OPTMZ-0], 280[OPTMZ-0], 367[OPTMZ-0]))
Statistics
----------------------------------------------------------------------------------------------------
0 physical reads
1299440 db block gets
0 consistent gets
0 redo size
0 recursive calls
0 bytes sent via SQL*Net to client
0 bytes received via SQL*Net from client
0 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
6 rows processed
0 bytes sent via PX
0 block received
33 rows fetched.
Elapsed: 00:00:07.897崖山的A-TIME看起来没有转换为Oracle A-TIME的时分秒样式，希望改进一下
ID=3 Loops=36757506，这是啥情况(如果有CACHE功能，ID=3的Loops应该显示为2，如果没有CACHE功能Loops应该显示为6)，希望改进和Oracle一模一样
Loops不准，那怎么验证崖山标量子查询是否有CACHE功能呢？
可以根据SQL执行时间来判断
1.创建函数(注意：因为Loops不准我才搞的函数来验证，没其他意思)
sql
CREATE OR REPLACE FUNCTION f_get_cnt_by_owner(p_owner varchar) RETURN int AS
v_cnt int;
BEGIN
select count(*) into v_cnt from test01 where owner = p_owner;
RETURN v_cnt;
END;
/2.设置statistics_level=TYPICAL，之前设置的是ALL，还原回来
sql
alter session set statistics_level=TYPICAL; 扫描函数一次，耗时2.7秒
sql
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2);
OBJECT_ID F_GET_CNT_BY_OWNER(OWNER)
----------- -------------------------
2 19348992
1 row fetched.
Elapsed: 00:00:02.707扫描函数6次，耗时5.7秒，唉？什么情况？这里应该耗时2.7*6=16.2秒啊，怎么等于5.7秒，5.7秒那就是只扫描了2次
sql
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID F_GET_CNT_BY_OWNER(OWNER)
----------- -------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows fetched.
Elapsed: 00:00:05.731看一下执行计划
sql
SQL> set autot trace
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2,3,4,117,280,367);
Execution Plan
----------------------------------------------------------------
SQL hash value: 1080230727
Optimizer: ADOPT_C
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| Id | Operation type | Name | Owner | E - Rows | A - Rows | Cost(%CPU) | A - Time | Loops | Memory | Disk | Partition info |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| 0 | SELECT STATEMENT | | | | | | | | | | |
| 1 | SUBQUERY | QUERY[1] | | | | | | | | | |
| 2 | AGGREGATE | | | 1| | 1151005( 0)| | | | | |
|* 3 | TABLE ACCESS FULL | TEST01 | SCOTT | 446| | 1151005( 0)| | | | | |
| 4 | TABLE ACCESS BY INDEX ROWID | TEST02 | SCOTT | 6| | 1( 0)| | | | | |
|* 5 | INDEX RANGE SCAN | IDX_TEST02_ID | SCOTT | 6| | 1( 0)| | | | | |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
Operation Information (identified by operation id):
---------------------------------------------------
1 - Subquery NDV info - NDV percentage: 0.500000, NDV Expression: ("TEST02"."OWNER")
3 - Predicate : filter("TEST01"."OWNER" = "TEST02"."OWNER")
5 - Predicate : access("TEST02"."OBJECT_ID" IN (2, 3, 4, 117, 280, 367))
Statistics
----------------------------------------------------------------------------------------------------
20 rows fetched.
Elapsed: 00:00:05.612原来崖山把SELECT调用自定义函数直接转成标量子查询了，同时也证明了崖山标量子查询有CACHE功能
在Oracle中，SELECT后调用自定义函数，要改写为SELECT (SELECT 自定义函数 FROM DUAL) FROM ...
或者是对函数开启RESULT CACHE功能才能减少函数被调用次数
sql
SQL> select object_id,(select f_get_cnt_by_owner(owner) from dual) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID (SELECTF_GET_CNT_BY_OWNER(OWNER)FROMDUAL)
----------- -----------------------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows fetched.
Elapsed: 00:00:05.803现在把函数稍作修改
sql
CREATE OR REPLACE FUNCTION f_get_cnt_by_owner(p_owner varchar) RETURN int AS
v_cnt int;
BEGIN
IF 1=1 THEN ---加了这里
select count(*) into v_cnt from test01 where owner = p_owner;
END IF; ---加了这里
RETURN v_cnt;
END;
/
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID F_GET_CNT_BY_OWNER(OWNER)
----------- -------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows fetched.
Elapsed: 00:00:15.228SQL耗时15.2秒，2.7*6=16.2，和15.2秒接近，说明函数被调用了6次
sql
SQL> select object_id,(select f_get_cnt_by_owner(owner) from dual) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID (SELECTF_GET_CNT_BY_OWNER(OWNER)FROMDUAL)
----------- -----------------------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows fetched.
Elapsed: 00:00:09.910把SELECT调用自定义函数改写为SELECT (SELECT 自定义函数 FROM DUAL) FROM写法，SQL耗时从15.2秒降低到9.9秒
笔者对9.9秒耗时有个疑问，耗时应该是2.7*2=5.4-6秒之间，怎么是9.9秒呢，希望崖山后续优化一下这个场景
通过上面测试，发现崖山SELECT调用自定义函数，如果自定义函数是纯SQL，能自动把自定义函数转为标量子查询
如果自定义函数不是纯SQL，无法转换，还是要把SELECT调用自定义函数改写为SELECT (SELECT 自定义函数 FROM DUAL) FROM
我们现在切回到Oracle11.2.0.4
1.创建函数，不加IF 1=1
sql
CREATE OR REPLACE FUNCTION f_get_cnt_by_owner(p_owner varchar) RETURN int AS
v_cnt int;
BEGIN
select count(*) into v_cnt from test01 where owner = p_owner;
RETURN v_cnt;
END;
/2.设置statistics_level=TYPICAL，之前设置的是ALL，还原回来
sql
alter session set statistics_level=TYPICAL; 扫描函数1次，耗时1.25秒
sql
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2);
OBJECT_ID F_GET_CNT_BY_OWNER(OWNER)
---------- -------------------------
2 19348992
Elapsed: 00:00:01.25扫描函数6次，耗时7.75秒
sql
SQL> select object_id,f_get_cnt_by_owner(owner) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID F_GET_CNT_BY_OWNER(OWNER)
---------- -------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows selected.
Elapsed: 00:00:07.75将函数调用改写到标量子查询中，耗时从7.75秒降低到2.54秒，也就是说函数只调用了2次
sql
SQL> select object_id,(select f_get_cnt_by_owner(owner) from dual) from test02 where object_id in(2,3,4,117,280,367);
OBJECT_ID (SELECTF_GET_CNT_BY_OWNER(OWNER)FROMDUAL)
---------- -----------------------------------------
2 19348992
3 19348992
4 19348992
117 17408512
280 17408512
367 17408512
6 rows selected.
Elapsed: 00:00:02.54结论：
1.崖山标量子查询有CACHE功能，值得表扬，很多国产数据库没有这个特性
2.崖山SELECT调用自定义函数，如果自定义函数是纯SQL会被优化器自动改写为标量子查询，如果自定义函数不是纯SQL，还是和Oracle一样要人工改写为标量子查询写法
3.希望崖山早日修复A-TIME，Loops瑕疵，最好把Loops改成Oracle Starts的效果
4.在Oracle中FILTER算法和标量子查询算法一模一样，本文不讨论FILTER CACHE，因为FILTER崖山优化器行为和ORACLE差异很大，需要单独写篇文章讨论
5.崖山SELECT (SELECT 非纯SQL自定义函数 FROM DUAL) FROM可能需要内核再优化优化