问题
I cannot figure out why sometimes, the total cost of a plan can be a very small number whereas looking inside the plan we can find huge costs. (indeed the query is very slow).
Can somebody explain me that?
Here is an example. Apparently the costful part comes from a field in the main select that does a listagg on a subview and the join condition with this subview contains a complex condition (we can join on one field or another).
| Id | Operation | Name | Rows | Bytes | Cost |
----------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 875 | 20 |
| 1 | SORT GROUP BY | | 1 | 544 | |
| 2 | VIEW | | 1 | 544 | 3 |
| 3 | SORT UNIQUE | | 1 | 481 | 3 |
| 4 | NESTED LOOPS | | | | |
| 5 | NESTED LOOPS | | 3 | 1443 | 2 |
| 6 | TABLE ACCESS BY INDEX ROWID | | 7 | 140 | 1 |
| 7 | INDEX RANGE SCAN | | 7 | | 1 |
| 8 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 9 | TABLE ACCESS BY INDEX ROWID | | 1 | 461 | 1 |
| 10 | SORT GROUP BY | | 1 | 182 | |
| 11 | NESTED LOOPS | | | | |
| 12 | NESTED LOOPS | | 8 | 1456 | 3 |
| 13 | NESTED LOOPS | | 8 | 304 | 2 |
| 14 | TABLE ACCESS BY INDEX ROWID | | 7 | 154 | 1 |
| 15 | INDEX RANGE SCAN | | 7 | | 1 |
| 16 | INDEX RANGE SCAN | | 1 | 16 | 1 |
| 17 | INDEX RANGE SCAN | | 1 | | 1 |
| 18 | TABLE ACCESS BY INDEX ROWID | | 1 | 144 | 1 |
| 19 | SORT GROUP BY | | 1 | 268 | |
| 20 | VIEW | | 1 | 268 | 9 |
| 21 | SORT UNIQUE | | 1 | 108 | 9 |
| 22 | CONCATENATION | | | | |
| 23 | NESTED LOOPS | | | | |
| 24 | NESTED LOOPS | | 1 | 108 | 4 |
| 25 | NESTED LOOPS | | 1 | 79 | 3 |
| 26 | NESTED LOOPS | | 1 | 59 | 2 |
| 27 | TABLE ACCESS BY INDEX ROWID | | 1 | 16 | 1 |
| 28 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 29 | TABLE ACCESS BY INDEX ROWID | | 1 | 43 | 1 |
| 30 | INDEX RANGE SCAN | | 1 | | 1 |
| 31 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 32 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 33 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 34 | TABLE ACCESS BY INDEX ROWID | | 1 | 29 | 1 |
| 35 | NESTED LOOPS | | | | |
| 36 | NESTED LOOPS | | 1 | 108 | 4 |
| 37 | NESTED LOOPS | | 1 | 79 | 3 |
| 38 | NESTED LOOPS | | 1 | 59 | 2 |
| 39 | TABLE ACCESS BY INDEX ROWID | | 4 | 64 | 1 |
| 40 | INDEX RANGE SCAN | | 2 | | 1 |
| 41 | TABLE ACCESS BY INDEX ROWID | | 1 | 43 | 1 |
| 42 | INDEX RANGE SCAN | | 1 | | 1 |
| 43 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 44 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 45 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 46 | TABLE ACCESS BY INDEX ROWID | | 1 | 29 | 1 |
| 47 | SORT GROUP BY | | 1 | 330 | |
| 48 | VIEW | | 1 | 330 | 26695 |
| 49 | SORT UNIQUE | | 1 | 130 | 26695 |
| 50 | CONCATENATION | | | | |
| 51 | HASH JOIN ANTI | | 1 | 130 | 13347 |
| 52 | NESTED LOOPS | | | | |
| 53 | NESTED LOOPS | | 1 | 110 | 4 |
| 54 | NESTED LOOPS | | 1 | 81 | 3 |
| 55 | NESTED LOOPS | | 1 | 61 | 2 |
| 56 | TABLE ACCESS BY INDEX ROWID | | 1 | 16 | 1 |
| 57 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 58 | TABLE ACCESS BY INDEX ROWID | | 1 | 45 | 1 |
| 59 | INDEX RANGE SCAN | | 1 | | 1 |
| 60 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 61 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 62 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 63 | TABLE ACCESS BY INDEX ROWID | | 1 | 29 | 1 |
| 64 | VIEW | | 164K| 3220K| 13341 |
| 65 | NESTED LOOPS | | | | |
| 66 | NESTED LOOPS | | 164K| 11M| 13341 |
| 67 | NESTED LOOPS | | 164K| 8535K| 10041 |
| 68 | TABLE ACCESS BY INDEX ROWID | | 164K| 6924K| 8391 |
| 69 | INDEX SKIP SCAN | | 2131K| | 163 |
| 70 | INDEX UNIQUE SCAN | | 1 | 10 | 1 |
| 71 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 72 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 73 | HASH JOIN ANTI | | 2 | 260 | 13347 |
| 74 | NESTED LOOPS | | | | |
| 75 | NESTED LOOPS | | 2 | 220 | 4 |
| 76 | NESTED LOOPS | | 2 | 162 | 3 |
| 77 | NESTED LOOPS | | 2 | 122 | 2 |
| 78 | TABLE ACCESS BY INDEX ROWID | | 4 | 64 | 1 |
| 79 | INDEX RANGE SCAN | | 2 | | 1 |
| 80 | TABLE ACCESS BY INDEX ROWID | | 1 | 45 | 1 |
| 81 | INDEX RANGE SCAN | | 1 | | 1 |
| 82 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 83 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 84 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 85 | TABLE ACCESS BY INDEX ROWID | | 1 | 29 | 1 |
| 86 | VIEW | | 164K| 3220K| 13341 |
| 87 | NESTED LOOPS | | | | |
| 88 | NESTED LOOPS | | 164K| 11M| 13341 |
| 89 | NESTED LOOPS | | 164K| 8535K| 10041 |
| 90 | TABLE ACCESS BY INDEX ROWID | | 164K| 6924K| 8391 |
| 91 | INDEX SKIP SCAN | | 2131K| | 163 |
| 92 | INDEX UNIQUE SCAN | | 1 | 10 | 1 |
| 93 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 94 | TABLE ACCESS BY INDEX ROWID | | 1 | 20 | 1 |
| 95 | NESTED LOOPS OUTER | | 1 | 875 | 20 |
| 96 | NESTED LOOPS OUTER | | 1 | 846 | 19 |
| 97 | NESTED LOOPS OUTER | | 1 | 800 | 18 |
| 98 | NESTED LOOPS OUTER | | 1 | 776 | 17 |
| 99 | NESTED LOOPS OUTER | | 1 | 752 | 16 |
| 100 | NESTED LOOPS OUTER | | 1 | 641 | 15 |
| 101 | NESTED LOOPS OUTER | | 1 | 576 | 14 |
| 102 | NESTED LOOPS OUTER | | 1 | 554 | 13 |
| 103 | NESTED LOOPS OUTER | | 1 | 487 | 12 |
| 104 | NESTED LOOPS OUTER | | 1 | 434 | 11 |
| 105 | NESTED LOOPS | | 1 | 368 | 10 |
| 106 | NESTED LOOPS | | 1 | 102 | 9 |
| 107 | NESTED LOOPS OUTER | | 1 | 85 | 8 |
| 108 | NESTED LOOPS | | 1 | 68 | 7 |
| 109 | NESTED LOOPS | | 50 | 2700 | 6 |
| 110 | HASH JOIN | | 53 | 1696 | 5 |
| 111 | INLIST ITERATOR | | | | |
| 112 | TABLE ACCESS BY INDEX ROWID| | 520 | 10400 | 3 |
| 113 | INDEX RANGE SCAN | | 520 | | 1 |
| 114 | INLIST ITERATOR | | | | |
| 115 | TABLE ACCESS BY INDEX ROWID| | 91457 | 1071K| 1 |
| 116 | INDEX UNIQUE SCAN | | 2 | | 1 |
| 117 | TABLE ACCESS BY INDEX ROWID | | 1 | 22 | 1 |
| 118 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 119 | TABLE ACCESS BY INDEX ROWID | | 1 | 14 | 1 |
| 120 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 121 | TABLE ACCESS BY INDEX ROWID | | 1 | 17 | 1 |
| 122 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 123 | TABLE ACCESS BY INDEX ROWID | | 1 | 17 | 1 |
| 124 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 125 | TABLE ACCESS BY INDEX ROWID | | 1 | 266 | 1 |
| 126 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 127 | TABLE ACCESS BY INDEX ROWID | | 1 | 66 | 1 |
| 128 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 129 | TABLE ACCESS BY INDEX ROWID | | 1 | 53 | 1 |
| 130 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 131 | TABLE ACCESS BY INDEX ROWID | | 1 | 67 | 1 |
| 132 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 133 | INDEX RANGE SCAN | | 1 | 22 | 1 |
| 134 | TABLE ACCESS BY INDEX ROWID | | 1 | 65 | 1 |
| 135 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 136 | TABLE ACCESS BY INDEX ROWID | | 1 | 111 | 1 |
| 137 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 138 | TABLE ACCESS BY INDEX ROWID | | 1 | 24 | 1 |
| 139 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 140 | TABLE ACCESS BY INDEX ROWID | | 1 | 24 | 1 |
| 141 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 142 | TABLE ACCESS BY INDEX ROWID | | 1 | 46 | 1 |
| 143 | INDEX UNIQUE SCAN | | 1 | | 1 |
| 144 | TABLE ACCESS BY INDEX ROWID | | 1 | 29 | 1 |
| 145 | INDEX UNIQUE SCAN | | 1 | | 1 |
----------------------------------------------------------------------------------------------------------
回答1:
The total cost of a statement is usually equal to or greater than the cost of any of its child operations. There are at least 4 exceptions to this rule.
Your plan looks like #3 but we can't be sure without looking at code.
1. FILTER
Execution plans may depend on conditions at run-time. These conditions cause FILTER operations that will dynamically decide which query block to execute. The example below uses a static condition but still demonstrates the concept. Part of the subquery is very expensive but the condition negates the whole thing.
explain plan for select * from dba_objects cross join dba_objects where 1 = 2;
select * from table(dbms_xplan.display(format => 'basic +cost'));
Plan hash value: 3258663795
--------------------------------------------------------------------
| Id | Operation | Name | Cost (%CPU)|
--------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 0 (0)|
| 1 | FILTER | | |
| 2 | MERGE JOIN CARTESIAN | | 11M (3)|
...
2. COUNT STOPKEY
Execution plans sum child operations up until the final cost. But child operations will not always finish. In the example below it may be correct to say that part of the plan costs 214. But because of the condition where rownum <= 1 only part of that child operation may run.
explain plan for
select /*+ no_query_transformation */ *
from (select * from dba_objects join dba_objects using (owner))
where rownum <= 1;
select * from table(dbms_xplan.display(format => 'basic +cost'));
Plan hash value: 2132093199
-------------------------------------------------------------------------------
| Id | Operation | Name | Cost (%CPU)|
-------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 4 (0)|
| 1 | COUNT STOPKEY | | |
| 2 | VIEW | | 4 (0)|
| 3 | VIEW | | 4 (0)|
| 4 | NESTED LOOPS | | 4 (0)|
| 5 | VIEW | DBA_OBJECTS | 2 (0)|
| 6 | UNION-ALL | | |
| 7 | HASH JOIN | | 3 (34)|
| 8 | INDEX FULL SCAN | I_USER2 | 1 (0)|
| 9 | VIEW | _CURRENT_EDITION_OBJ | 1 (0)|
| 10 | FILTER | | |
| 11 | HASH JOIN | | 214 (3)|
...
3. Subqueries in the SELECT column list
Cost aggregation does not include subqueries in the SELECT column list. A query like select ([expensive query]) from dual; will have a very small total cost. I don't understand the reason for this; Oracle estimates the subquery and he number of rows in the FROM, surely it could multiply them together for a total cost.
explain plan for
select dummy,(select count(*) from dba_objects cross join dba_objects) from dual;
select * from table(dbms_xplan.display(format => 'basic +cost'));
Plan hash value: 3705842531
---------------------------------------------------------------
| Id | Operation | Name | Cost (%CPU)|
---------------------------------------------------------------
| 0 | SELECT STATEMENT | | 2 (0)|
| 1 | SORT AGGREGATE | | |
| 2 | MERGE JOIN CARTESIAN | | 11M (3)|
...
4. Other - rounding? bugs?
About 0.01% of plans still have unexplainable cost issues. I can't find any pattern among them. Perhaps it's just a rounding issue or some rare optimizer bugs. There will always be some weird cases with a any model as complicated as the optimizer.
Check for more exceptions
This query can find other exceptions, it returns all plans where the first cost is less than the maximum cost.
select *
from
(
--First and Max cost per plan.
select
sql_id, plan_hash_value, id, cost
,max(cost) keep (dense_rank first order by id)
over (partition by sql_id, plan_hash_value) first_cost
,max(cost)
over (partition by sql_id, plan_hash_value) max_cost
,max(case when operation = 'COUNT' and options = 'STOPKEY' then 1 else 0 end)
over (partition by sql_id, plan_hash_value) has_count_stopkey
,max(case when operation = 'FILTER' and options is null then 1 else 0 end)
over (partition by sql_id, plan_hash_value) has_filter
,count(distinct(plan_hash_value))
over () total_plans
from v$sql_plan
--where sql_id = '61a161nm1ttjj'
order by 1,2,3
)
where first_cost < max_cost
--It's easy to exclude FILTER and COUNT STOPKEY.
and has_filter = 0
and has_count_stopkey = 0
order by 1,2,3;
来源：https://stackoverflow.com/questions/25179760/oracle-total-plan-cost-is-really-less-than-some-of-its-elements