I'm pretty confused by this behavior from Oracle optimizer. This has to do with union all operations from a CTE. If anyone has any ideas I'm all game.
--Relevant data structures:
--t_positionperf (index: POSITIONPERF_X1 (account_id, hist_date, security_id)
--t_positionhist (index: POSITIONHIST_X1 (position_id, hist_date, hist_type)
-- PK_T_POSITIONHIST (hist_date, position_id, hist_type)
--v_positiontype (very simple "case-when" translation of a tiny lookup table)
with q as (
select
pp.position_id, pp.hist_date, pp.account_id, pp.income, pp.expense,
ph.position_type_id, ph.price, ph.quantity, ph.factor, ph.daily_accrual,
n.daily_accrual as new_daily_accrual, nvl(n.is_loan, v.is_loan) as new_is_loan
from
t_positionperf pp
left outer join t_positionhist ph
on pp.position_id = ph.position_id
and pp.hist_date = ph.hist_date
and ph.hist_type = 'O' --the 'O' join set from t_positionhist
left outer join v_positiontype v
on ph.position_type_id = v.position_type_id
left outer join (
select
x.position_id, x.hist_date, x.daily_accrual, v2.is_loan
from t_positionhist x
join v_positiontype v2 on x.position_type_id = v2.position_type_id
where x.hist_type = 'N'
) n --the 'N' join set from t_positionhist
on ph.position_id = n.position_id
and ph.hist_date = n.hist_date
where pp.account_id in (5018,5312)
and pp.hist_date between to_date('01-jan-14') and to_date('31-jan-14')
)
select
q.account_id,q.hist_date,q.position_id,q.income,
q.expense,q.position_type_id,q.price,q.quantity,q.factor
from q
This uses a sensible access path.
Statistics
----------------------------------------------------------
29 recursive calls
0 db block gets
37640 consistent gets
5115 physical reads
0 redo size
227442 bytes sent via SQL*Net to client
4952 bytes received via SQL*Net from client
409 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
6116 rows processed
Execution Plan:
--------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU) | Time |
--------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 941 | 88454 | 5283 (1)| 00:01:04 |
|* 1 | FILTER | | | | | |
| 2 | NESTED LOOPS OUTER | | 941 | 88454 | 5283 (1)| 00:01:04 |
|* 3 | HASH JOIN RIGHT OUTER | | 941 | 71516 | 2456 (1)| 00:00:30 |
| 4 | TABLE ACCESS FULL | T_POSITIONTYPE | 64 | 1216 | 3 (0)| 00:00:01 |
| 5 | NESTED LOOPS OUTER | | 941 | 53637 | 2452 (1)| 00:00:30 |
| 6 | INLIST ITERATOR | | | | | |
| 7 | TABLE ACCESS BY INDEX ROWID| T_POSITIONPERF | 941 | 22584 | 567 (1)| 00:00:07 |
|* 8 | INDEX RANGE SCAN | POSITIONPERF_X1 | 941 | | 6 (0)| 00:00:01 |
| 9 | TABLE ACCESS BY INDEX ROWID | T_POSITIONHIST | 1 | 33 | 2 (0)| 00:00:01 |
|* 10 | INDEX UNIQUE SCAN | POSITIONHIST_X1 | 1 | | 1 (0)| 00:00:01 |
| 11 | VIEW PUSHED PREDICATE | | 1 | 18 | 3 (0)| 00:00:01 |
| 12 | NESTED LOOPS | | 1 | 27 | 4 (0)| 00:00:01 |
| 13 | TABLE ACCESS BY INDEX ROWID | T_POSITIONHIST | 1 | 20 | 3 (0)| 00:00:01 |
|* 14 | INDEX UNIQUE SCAN | POSITIONHIST_X1 | 1 | | 2 (0)| 00:00:01 |
| 15 | TABLE ACCESS BY INDEX ROWID | T_POSITIONTYPE | 1 | 7 | 1 (0)| 00:00:01 |
|* 16 | INDEX UNIQUE SCAN | PK_T_POSITIONTYPE | 1 | | 0 (0)| 00:00:01 |
--------------------------------------------------------------------------------------------------------
However, if you add a UNION ALL to this select from the CTE, so that that the query looks like:
with q as (...)
select q.account_id, q... etc.
UNION ALL
select q.account_id, max(q...) etc.
group by q.account_id
The Optimizer now wants to materialize the CTE (TEMP TABLE TRANSFORMATION). This part makes sense to me because of the UNION ALL from the same CTE. What I don't understand is why materializing the CTE causes it to choose a different (and much inferior) access path. It seems to lose the ability to push the predicate to the T_POSITIONHIST table when building the 'N' join set. You can see that it now does a Full Table Scan on T_POSITIONHIST, and as a result the disk I/O has gone through the roof:
Statistics
----------------------------------------------------------
362 recursive calls
53 db block gets
546116 consistent gets
521527 physical reads
688 redo size
250541 bytes sent via SQL*Net to client
4952 bytes received via SQL*Net from client
409 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
6117 rows processed
Execution Plan
------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1882 | 205K| 13 (54)| 00:00:01 |
| 1 | TEMP TABLE TRANSFORMATION | | | | | |
| 2 | LOAD AS SELECT | SYS_TEMP_0FD9D6CB3_54C7EBF8 | | | | |
|* 3 | FILTER | | | | | |
|* 4 | HASH JOIN OUTER | | 941 | 81867 | 147K (3)| 00:29:32 |
|* 5 | HASH JOIN RIGHT OUTER | | 941 | 63047 | 2456 (1)| 00:00:30 |
| 6 | VIEW | V_POSITIONTYPE | 64 | 448 | 3 (0)| 00:00:01 |
| 7 | TABLE ACCESS FULL | T_POSITIONTYPE | 64 | 448 | 3 (0)| 00:00:01 |
| 8 | NESTED LOOPS OUTER | | 941 | 56460 | 2452 (1)| 00:00:30 |
| 9 | INLIST ITERATOR | | | | | |
| 10 | TABLE ACCESS BY INDEX ROWID| T_POSITIONPERF | 941 | 22584 | 567 (1)| 00:00:07 |
|* 11 | INDEX RANGE SCAN | POSITIONPERF_X1 | 941 | | 6 (0)| 00:00:01 |
| 12 | TABLE ACCESS BY INDEX ROWID | T_POSITIONHIST | 1 | 36 | 2 (0)| 00:00:01 |
|* 13 | INDEX UNIQUE SCAN | POSITIONHIST_X1 | 1 | | 1 (0)| 00:00:01 |
| 14 | VIEW | | 2358K| 44M| 145K (3)| 00:29:02 |
| 15 | VIEW | | 2358K| 51M| 145K (3)| 00:29:02 |
|* 16 | HASH JOIN | | 2358K| 67M| 145K (3)| 00:29:02 |
| 17 | VIEW | V_POSITIONTYPE | 64 | 448 | 3 (0)| 00:00:01 |
| 18 | TABLE ACCESS FULL | T_POSITIONTYPE | 64 | 448 | 3 (0)| 00:00:01 |
|* 19 | TABLE ACCESS FULL | T_POSITIONHIST | 2358K| 51M| 145K (3)| 00:29:01 |
| 20 | UNION-ALL | | | | | |
| 21 | VIEW | | 941 | 106K| 6 (0)| 00:00:01 |
| 22 | TABLE ACCESS FULL | SYS_TEMP_0FD9D6CB3_54C7EBF8 | 941 | 56460 | 6 (0)| 00:00:01 |
| 23 | HASH GROUP BY | | 941 | 99K| 7 (15)| 00:00:01 |
| 24 | VIEW | | 941 | 99K| 6 (0)| 00:00:01 |
| 25 | TABLE ACCESS FULL | SYS_TEMP_0FD9D6CB3_54C7EBF8 | 941 | 56460 | 6 (0)| 00:00:01 |
------------------------------------------------------------------------------------------------------------------
The weird part is, you can partially get around this issue by duplicating the CTE, like so:
with q as (...), q2 as (.../*duplicate of q*/)
select q.account_id, q... etc.
UNION ALL
select q2account_id, max(q2...) etc.
group by q2.account_id
Because it no longer wants to materialize the result set, it's able to use the smarter access path, except it has to do it twice, so the cost is approximately doubled (but still far less than the really bad access path it uses above).
Statistics
57 recursive calls
0 db block gets
73855 consistent gets
0 physical reads
0 redo size
202937 bytes sent via SQL*Net to client
4953 bytes received via SQL*Net from client
409 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
6117 rows processed