Re: Execution path having full scan in a nested loop
- From: Jonathan Lewis <jlewisoracle@xxxxxxxxx>
- To: Oracle L <oracle-l@xxxxxxxxxxxxx>
- Date: Tue, 12 Oct 2021 10:28:19 +0100
So I want to understand if we can modify this query to make it go for one
time full table scan, maybe with a hash join kind of operation, so that
this can complete in a faster time?
I knew that if I answered the question you'd ask you'd only ask a load
more. ;)
I suggested the hints because that would demonstrate whether or not the
transformation was possible (it almost always is in 19c) and purely
cost-based (which is not always the case - sometimes the optimizer will
pick a higher cost path if a lower cost path is based in guesswork that
may introduce a lot of variability at run-time).
If the hinted code does what you expect and the cost of the query is HIGHER
as a result this almost always means that there's a statistics issue
involved. (If the cost is LOWER this tends to mean that Oracle has a rule
that has rejected the path.) This means you may be able to find a suitable
way to fix the statistics to get the plan you want without hinting.
I suggested the NO_MERGE hint because you may find that hinted with only
UNNEST a change in the data results in the optimizer deciding to do complex
view merging - which means it will join all the tables and aggregate late,
and that could be a disaster. (Here's a link to a very old blog note of
mine demonstrating the concept:
https://jonathanlewis.wordpress.com/2007/03/08/transformation-and-optimisation/
)
You started with a comment about an index on RID, but that's irrelevant, it
could help the join from MASTER to DETAIL, but that doesn't help with the
correlation between the subquery and PBRF. This becomes cleared when Oracle
rewrites the NOT IN to NOT EXISTS and you can see the necessary correlation
in the FILTER In the original plan:
3 - filter( NOT EXISTS
(SELECT 0 FROM "USER1"."DETAIL" "DTL","USER1"."MASTER" "REQ"
WHERE "REQ"."RID"="DTL"."RID" AND "REQ"."RTYP"=:B4
AND ("REQ"."RSTS"='YY' OR "REQ"."RSTS"='XX')
AND "DTL"."FL_TYP"='DP'
AND LNNVL("DTL"."FL_NM"<>:B1) AND
LNNVL(NVL(:B2,'-')<>NVL("DTL"."SID",'-'))))
and compare it with the original NOT IN predicate
AND (rf.fattr3, NVL (rf.fattr, '-')) NOT IN(
SELECT DISTINCT dtl.FL_NM, NVL (dtl.SID, '-')
The connection between RF and DTL is based on fattr3/fl_nm and fattr/sid -
and because the new plan report a "null aware" antijoin, you can see that
the optimizer has embedded the correlation predicates inside then lnnvl()
function. There might be a way to add some NOT NULL declarations and
function-based indexes to make this more efficient, but don't ask me to
guess about things that might or might not be valid for your data. It's
also possible that with the right indexing and stats the subquery could
drive a nested loop from MASTER to DETAIL, but that would at the least
require an index on (rtyp, rsts) and maybe some statistics that made it
look as if that matched only a few rows.
Another benefit of seeing the results of the unnested subquery is that we
can see why the optimizer didn't choose it.
The cost of the single SORT JOIN for the subquery is over 8,000; the cost
of the tablescan of DETAIL is 607, which means Oracle could do the
tablescan about 13 times before it should choose to unnest. However its
estimate of rows from the driving table is 2, (although the actual is
9.626). If you can get the optimizer to get a better estimate of
cardinality of the driving table it would unnest automatically. It's
possible that creating a column group on (F_GRP, rf.F_CATG) would be
sufficient if the number of combinations is significantly less than the
product of the individual num_distinct.
Regards
Jonathan Lewis
On Tue, 12 Oct 2021 at 07:05, Lok P <loknath.73@xxxxxxxxx> wrote:
Thank you Jonathan.
Your suggested hints(even with only UNNEST hints) to the subquery making
the query finish in ~2seconds as opposed to ~143 seconds before and so it
seems it's possible as per optimizer transformations is concerned. So I
wanted to understand why optimizers by default are not doing this
because of any restriction or we should tweak the query someway Or any
issue with stats itself?
Initially I was thinking if we are hitting the below restriction as it
is in the blog due to which the optimizer is not able to go for a hash
join, so tried putting that subquery in a case statement , but it
didn't work.
https://ctandrewsayer.wordpress.com/2018/06/06/conditional-outer-joins-forcing-nested-loops/
Global Information
------------------------------
Status : DONE (ALL ROWS)
Instance ID : 1
SQL ID : 0v7fsttsqubh5
SQL Execution ID : 16777216
Execution Started : 10/12/2021 01:30:02
First Refresh Time : 10/12/2021 01:30:02
Last Refresh Time : 10/12/2021 01:30:04
Duration : 2s
Module/Action : SQL*Plus/-
Program : sqlplus.exe
Fetch Calls : 2
Global Stats
======================================================================================================
| Elapsed | Cpu | IO | Application | Concurrency | Fetch |
Buffer | Read | Read | Cell |
| Time(s) | Time(s) | Waits(s) | Waits(s) | Waits(s) | Calls | Gets
| Reqs | Bytes | Offload |
======================================================================================================
| 1.02 | 0.98 | 0.04 | 0.00 | 0.00 | 2 |
31621 | 228 | 206MB | 81.41% |
======================================================================================================
SQL Plan Monitoring Details (Plan Hash Value=3386919479)
=============================================================================================================================================================================================
| Id | Operation | Name |
Rows | Cost | Time | Start | Execs | Rows | Read | Read |
Cell | Mem | Activity | Activity Detail |
| | | |
(Estim) | | Active(s) | Active | | (Actual) | Reqs | Bytes |
Offload | (Max) | (%) | (# samples) |
=============================================================================================================================================================================================
| 0 | SELECT STATEMENT | |
| | 1 | +2 | 1 | 421 | | |
| | | |
| 1 | WINDOW BUFFER | |
1 | 8150 | 1 | +2 | 1 | 421 | | |
| 32768 | | |
| 2 | HASH GROUP BY | |
1 | 8150 | 1 | +2 | 1 | 421 | | |
| 1M | | |
| 3 | FILTER | |
| | 1 | +2 | 1 | 6985 | | |
| | | |
| 4 | MERGE JOIN ANTI NA | |
1 | 8149 | 1 | +2 | 1 | 6985 | | |
| | | |
| 5 | SORT JOIN | |
2 | 120 | 1 | +2 | 1 | 9626 | | |
| 886K | | |
| 6 | TABLE ACCESS BY INDEX ROWID | PBRF |
2 | 119 | 1 | +2 | 1 | 9626 | | |
| | | |
| 7 | INDEX SKIP SCAN | PBRF_IX1 |
4 | 118 | 1 | +2 | 1 | 9626 | | |
| | | |
| 8 | SORT UNIQUE | |
868K | 8029 | 2 | +1 | 9627 | 2641 | | |
| | 100.00 | Cpu (1) |
| 9 | VIEW | VW_NSO_1 |
868K | 635 | 1 | +2 | 1 | 1M | | |
| | | |
| 10 | HASH JOIN | |
868K | 635 | 1 | +2 | 1 | 1M | | |
| 2M | | |
| 11 | JOIN FILTER CREATE | :BF0000 |
19381 | 18 | 1 | +2 | 1 | 18647 | | |
| | | |
| 12 | TABLE ACCESS STORAGE FULL | MASTER |
19381 | 18 | 1 | +2 | 1 | 18647 | | |
| | | |
| 13 | JOIN FILTER USE | :BF0000 |
1M | 607 | 1 | +2 | 1 | 1M | | |
| | | |
| 14 | TABLE ACCESS STORAGE FULL | DETAIL |
1M | 607 | 1 | +2 | 1 | 1M | 228 | 206MB | 81.41%
| 7M | | |
=============================================================================================================================================================================================
Predicate Information (identified by operation id):
---------------------------------------------------
3 - filter(TO_DATE(:B6,'MM/DD/YYYY')>=TO_DATE(:B5,'MM/DD/YYYY'))
7 - access("RF"."W_DATE">=TO_DATE(:B5,'MM/DD/YYYY') AND
"RF"."F_GRP"=:B1 AND "RF"."W_DATE"<=TO_DATE(:B6,'MM/DD/YYYY'))
filter("RF"."F_GRP"=:B1 AND ("RF"."F_CATG"=:B2 OR
"RF"."F_CATG"=:B3))
8 - access(INTERNAL_FUNCTION("RF"."fattr3")=INTERNAL_FUNCTION("FL_NM")
AND INTERNAL_FUNCTION("NVL(DTL.SID,'-')")=NVL("RF"."fattr2",'-'))
filter(INTERNAL_FUNCTION("NVL(DTL.SID,'-')")=NVL("RF"."fattr2",'-') AND
INTERNAL_FUNCTION("RF"."fattr3")=INTERNAL_FUNCTION("FL_NM"))
10 - access("REQ"."RID"="DTL"."RID")
12 - storage("REQ"."RTYP"=:B4 AND ("REQ"."RSTS"='FP' OR
"REQ"."RSTS"='IP'))
filter("REQ"."RTYP"=:B4 AND ("REQ"."RSTS"='FP' OR
"REQ"."RSTS"='IP'))
14 - storage("DTL"."FL_TYP"='DP' AND
SYS_OP_BLOOM_FILTER(:BF0000,"DTL"."RID"))
filter("DTL"."FL_TYP"='DP' AND
SYS_OP_BLOOM_FILTER(:BF0000,"DTL"."RID"))
On Tue, Oct 12, 2021 at 2:31 AM Jonathan Lewis <jlewisoracle@xxxxxxxxx>
wrote:
The shape of the plan you're describing is one where the subquery is
unnested to a non-mergeable aggregate view.
To see if this is legal as far as the optimizer is concerned you could
add the hints /*+ unnest no_merge */ to the subquery.
Regards
Jonathan Lewis
On Mon, 11 Oct 2021 at 20:50, Lok P <loknath.73@xxxxxxxxx> wrote:
Hello Listers, We have one database on version 11.2.0.4 of oracle. And
below query is spending quite a lot of time while scanning table 'DETAIL'
in nested loop path i.e. step-9 below. This table does have an index with
the leading column as RID(which is joined column) but still it's going for
a 'TABLE ACCESS STORAGE FULL FIRST ROWS' within a nested loop. Not sure if
it's just because we are reading a lot of rows from that table or if we are
hitting any optimizer restriction. So I want to understand if we can modify
this query to make it go for one time full table scan, maybe with a hash
join kind of operation, so that this can complete in a faster time?
Table DETAIL having ~1.7million rows in it. And column RID having 31K
distinct values in it. And table MASTER having ~34k in it.
SELECT TRIM (rf.fattr3) ,TRIM (rf.fattr) , rf.W_DATE ,rf.CODE ,NVL (SUM
(rf.txn_cnt), 0) AS cnt,NVL (SUM (DECODE (TTYP, 'S', 1, -1) * rf.amt),0)
amt,count(*) over () count1
FROM PBRF rf
WHERE rf.F_GRP = :b1
AND rf.F_CATG IN ( :b2, :b3)
AND (rf.fattr3, NVL (rf.fattr, '-')) NOT IN
(SELECT DISTINCT dtl.FL_NM, NVL (dtl.SID, '-')
FROM MASTER req, DETAIL dtl
WHERE req.RID = dtl.RID
AND req.RSTS IN ('XX', 'YY')
AND req.RTYP = :b4
AND dtl.FL_TYP = 'DP')
AND rf.W_DATE BETWEEN TO_DATE ( :b5,'MM/DD/YYYY') AND TO_DATE (
:b6,'MM/DD/YYYY')
GROUP BY rf.CODE,TRIM (rf.fattr),TRIM (rf.fattr3), rf.W_DATE;
Global Information
------------------------------
Status : DONE (ALL ROWS)
Instance ID : 1
SQL ID : 7yc5x3uzx7vzm
SQL Execution ID : 16777216
Execution Started : 10/11/2021 15:18:40
First Refresh Time : 10/11/2021 15:18:40
Last Refresh Time : 10/11/2021 15:21:03
Duration : 143s
Module/Action : SQL*Plus/-
Program : sqlplus.exe
Fetch Calls : 2
Global Stats
===================================================================================================
| Elapsed | Cpu | IO | Application | Other | Fetch | Buffer
| Read | Read | Cell |
| Time(s) | Time(s) | Waits(s) | Waits(s) | Waits(s) | Calls | Gets
| Reqs | Bytes | Offload |
===================================================================================================
| 144 | 131 | 11 | 1.13 | 0.18 | 2 | 38M
| 348K | 293GB | 80.58% |
===================================================================================================
SQL Plan Monitoring Details (Plan Hash Value=3512223473)
==========================================================================================================================================================================================================================
| Id | Operation | Name
| Rows | Cost | Time | Start | Execs | Rows | Read | Read
| Cell | Mem | Activity | Activity Detail |
| | |
| (Estim) | | Active(s) | Active | | (Actual) | Reqs | Bytes
| Offload | (Max) | (%) | (# samples) |
==========================================================================================================================================================================================================================
| 0 | SELECT STATEMENT |
| | | 1 | +143 | 1 | 421 | |
| | | | |
| 1 | WINDOW BUFFER |
| 1 | 125 | 1 | +143 | 1 | 421 | |
| | 32768 | | |
| 2 | HASH GROUP BY |
| 1 | 125 | 142 | +2 | 1 | 421 | |
| | 1M | | |
| 3 | FILTER |
| | | 142 | +2 | 1 | 6985 | |
| | | | |
| 4 | FILTER |
| | | 142 | +2 | 1 | 9626 | |
| | | | |
| 5 | TABLE ACCESS BY INDEX ROWID | PBRF
| 2 | 119 | 142 | +2 | 1 | 9626 | |
| | | | |
| 6 | INDEX SKIP SCAN | PBRF_IX1
| 4 | 118 | 142 | +2 | 1 | 9626 | 1 |
8192 | | | | |
| 7 | NESTED LOOPS |
| 3 | 5 | 142 | +2 | 1712 | 532 | |
| | | | |
| 8 | NESTED LOOPS |
| 3 | 5 | 142 | +2 | 1712 | 1417 | |
| | | | |
| 9 | TABLE ACCESS STORAGE FULL FIRST ROWS | DETAIL
| 6436 | 2 | 144 | +1 | 1712 | 1417 | 348K |
293GB | 80.58% | 17M | 100.00 | enq: KO - fast object checkpoint (2) |
| | |
| | | | | | | |
| | | | Cpu (124) |
| | |
| | | | | | | |
| | | | reliable message (4) |
| | |
| | | | | | | |
| | | | cell smart table scan (14) |
| 10 | INDEX UNIQUE SCAN | MASTER_PK
| 1 | | 142 | +2 | 1417 | 1417 | |
| | | | |
| 11 | TABLE ACCESS BY INDEX ROWID | MASTER
| 1 | 1 | 142 | +2 | 1417 | 532 | |
| | | | |
==========================================================================================================================================================================================================================
Predicate Information (identified by operation id):
---------------------------------------------------
3 - filter( NOT EXISTS (SELECT 0 FROM "USER1"."DETAIL"
"DTL","USER1"."MASTER" "REQ" WHERE "REQ"."RID"="DTL"."RID" AND
"REQ"."RTYP"=:B4 AND
("REQ"."RSTS"='YY' OR "REQ"."RSTS"='XX') AND
"DTL"."FL_TYP"='DP' AND LNNVL("DTL"."FL_NM"<>:B1) AND
LNNVL(NVL(:B2,'-')<>NVL("DTL"."SID",'-'))))
4 - filter(TO_DATE(:B6,'MM/DD/YYYY')>=TO_DATE(:B5,'MM/DD/YYYY'))
6 - access("RF"."W_DATE">=TO_DATE(:B5,'MM/DD/YYYY') AND
"RF"."F_GRP"=:B1 AND "RF"."W_DATE"<=TO_DATE(:B6,'MM/DD/YYYY'))
filter("RF"."F_GRP"=:B1 AND ("RF"."F_CATG"=:B2 OR
"RF"."F_CATG"=:B3))
9 - storage("DTL"."FL_TYP"='DP')
filter("DTL"."FL_TYP"='DP' AND LNNVL("DTL"."FL_NM"<>:B1) AND
LNNVL(NVL(:B2,'-')<>NVL("DTL"."SID",'-')))
10 - access("REQ"."RID"="DTL"."RID")
11 - filter("REQ"."RTYP"=:B4 AND ("REQ"."RSTS"='YY' OR
"REQ"."RSTS"='XX'))
Statistics
----------------------------------------------------------
1 recursive calls
0 db block gets
38469369 consistent gets
38459185 physical reads
52 redo size
14543 bytes sent via SQL*Net to client
472 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
421 rows processed
Other related posts:
