RE: FW: Error related to hash memory
Date: Sat, 4 Dec 2021 07:23:27 -0500
Message-ID: <65ac01d7e909$bdd63ab0$3982b010$_at_rsiz.com>
IF your memory still spills AND IF all your aggregations are SUM and COUNTs, you can do the individual partition summations into a union all with an outer select summing the returned columns with your group by, which will be one row per partition per final row. This should reduce the hash requirement to the whatever you would get for the partition returning the largest number of matched rows.
You aggregations other than count and sum are unlikely to give the correct answer. Any counts in the internal aggregation will need to be converted to SUM in the outer select of the group by sum of the partitionwise result.
That is a fragile hand hack with plenty of opportunity to err. Do all of the stuff JL recommended before you head down that route.
Good luck,
mwf
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Jonathan Lewis
Sent: Friday, December 03, 2021 9:54 AM
To: Oracle L
Subject: Re: FW: Error related to hash memory
Comments inline.
Regards
Jonathan Lewis
>>> As you said exactly there are many decodes/cases in the query and scalar subqueries just before 'group by' applies. There is a function(Pkg1.Fun1) used which is not parallel enabled. Also parallel hints used in the Select part of the query. And this query does insert+append into a global temporary table. Also in the FROM clause all are global temporary tables except TFA, TFMA and TSLV.
>>> In regards to having scalar subqueries operated after group by , would that not make the results different?
In some cases yes, in others no. The example I linked to doesn't highlight the threat you're thinking of since it doesn't AGGREGATE by department id before joining to get the department name; but if there had been an original "aggregate by department name" and I had aggregated by department ID and then joined to get the department name the result would have been different if two departments had had the same name. So there is a uniqueness requirement that you have to think about (8 times) in your case.You may find that none of your scalar subqueries can be moved in the same way. Another variation on the theme (which, again, may not be appropriate because it changes the result) is to see if you can aggregate any of the three tables on their join columns (in a no_merge view) before doing the join. (Oracle can do this automatically in later versions, but not - I think - in 11.g).
I've added a /*+ cardinality() */ hint to the start of your code that I think is in the correct place. It's easier that remembering the syntax for the opt_estimate()
This is simply to dump a trace file, not to work around the problem.
INSERT /*+ APPEND */
INTO GTT_TABLE (.....)
SELECT ......, SUM (crt), SUM (art), SUM (iamt), SUM (amt),...... ~46 different columns
FROM (SELECT /*+ cardinality(168000000) */ ....33 columns,
Regards
Jonathan Lewis
On Fri, 3 Dec 2021 at 13:53, Lok P <loknath.73_at_gmail.com> wrote:
Thank You So Much Jonathan and Mark.
As you said exactly there are many decodes/cases in the query and scalar subqueries just before 'group by' applies. There is a function(Pkg1.Fun1) used which is not parallel enabled. Also parallel hints used in the Select part of the query. And this query does insert+append into a global temporary table. Also in the FROM clause all are global temporary tables except TFA, TFMA and TSLV.
It's a big Insert query and I tried to not put all the individual columns but just key functions(like decode, case etc). It groups ~46 columns in that 'hash group by' but before that it fetches+transforms, ~55+ columns data. I will try to make that function parallel enabled + enable parallel DML for the query and then hopefully that "hash group by" can be operated in parallel.
Apology if these are silly questions,
In regards to having scalar subqueries operated after group by , would that not make the results different?
I am still trying to understand how to find the exact opt_estimate hints with a proper query block to make the optimizer estimate ~168million before "hash group by". And by setting the "event 32699" will that make the error disappear OR is it for debugging to find out the reason around the Ora- 32690?
INSERT /*+ APPEND */
INTO GTT_TABLE (.....)
SELECT ......, SUM (crt), SUM (art), SUM (iamt), SUM (amt),...... ~46 different columns
FROM (SELECT ....33 columns,
DECODE ( Amt1, 1, (SELECT Pkg1.Fun1 (...) FROM DUAL), NULL),
..13 columns..,
(SELECT DECODE (SUBSTR (Col_data,1,INSTR (Col_data, '*',1, 1)- 1),'XX', SUBSTR (Col_data,INSTR (Col_data,'*',1, 1)+ 1,( INSTR (Col_data,'*',1,2)- INSTR (Col_data,'*', 1,1))- 1))FROM DUAL)/ 100 Amt,
(SELECT DECODE (MPCD,'YY', SUBSTR (Col_data,INSTR (Col_data, '*',1, 2) + 1,( INSTR (Col_data,'*',1, 3)- INSTR (Col_data, '*',1,2))- 1)) FROM DUAL) aty,,
(SELECT DECODE (MPCD, 'YY', SUBSTR (Col_data, INSTR (Col_data, '*', 1,3) + 1, ( INSTR (Col_data,'*', 1, 4)- INSTR (Col_data, '*', 1, 3))- 1)) FROM DUAL) vnm,
(SELECT DECODE (SUBSTR (Col_data, 1, INSTR (Col_data, '*', 1, 1) - 1), 'M1', SUBSTR (Col_data, INSTR (Col_data, '*', 1, 4)+ 1,( INSTR (Col_data, '*',1,5)- INSTR (Col_data, '*', 1, 4)) - 1), NULL) FROM DUAL) INM,
(SELECT DECODE (MCID,9999, MDID, 9999, MDID, DECODE (SUBSTR (Col_data, 1, INSTR (Col_data,'*', 1,1) - 1),'M1', SUBSTR (Col_data,INSTR (Col_data, '*',1,2) + 1, ( INSTR (Col_data, '*',1,3) - INSTR (Col_data, '*', 1,2))-1))) FROM DUAL)
FROM (SELECT /*+ parallel(TTTD,4) leading(TTSFA TTTD)*/
DECODE (TTFE.C1, 'BBB', TTFE.CRT, NULL)
.......9 columns..
DECODE (TFMA.MACD, 'XX', TFMA.text1, 'XX', TFMA.text2, NULL) RDFI,
... 18 columns..,
DECODE (TFMA.MACD, 'XX', TFMA.text1, NULL) cbcd,
CASE
WHEN INSTR ( :B7, PTCD) = 0 THEN TTSFA.ILVL
ELSE
CASE
WHEN (SELECT PSS FROM tmfs FS WHERE FS.ptsid = TTFE.fsid) = 'N'
THEN CASE
WHEN (SELECT FEE.PSS FROM FEE FEE WHERE FEE.ptsid = TTFE.fsid AND FEE.PTMCD = TTTD.MPCD) = 'Y'
THEN TTSFA.ILVL
ELSE
CASE
WHEN INSTR ( :B6, PTCD) > 0
THEN DECODE (TTTD.MPCD,'XX', 'XX', 'XX', 'XXX','XX', 'XX','XX', 'XX', NULL)
WHEN INSTR ( :B5, PTCD) > 0
THEN DECODE (TTTD.MPCD, 'XX', 'XX', TTSFA.ILVL)
END
END
ELSE
CASE
WHEN (SELECT PSS FROM tmfs FS WHERE FS.ptsid = TTFE.fsid) = 'Y'
THEN CASE
WHEN (SELECT FEE.PSS FROM FEE FEE WHERE FEE.ptsid =TTFE.fsid AND FEE.PTMCD = TTTD.MPCD) ='N'
THEN CASE
WHEN INSTR (:B6,PTCD) > 0
THEN DECODE (TTTD.MPCD, 'XX', 'XX','VI', 'XX', 'XX', 'XX', 'XX', 'XX', NULL)
WHEN INSTR ( :B5,PTCD) >0
THEN DECODE (TTTD.MPCD, 'XX', 'XX', TTSFA.ILVL)
END
ELSE
TTSFA.ILVL
END
END
END
END
ILVL,
DECODE (TTFE.C1, 'XX', TTFE.CRT, 'XX', TTFE.CRT, 0) CRT,
DECODE (TTFE.C1, 'XX', TTFE.ARATE, 'XX', TTFE.amrt, 0) ARATE,
CASE
WHEN TTFE.C1 = 'XX' OR TTFE.C1 = 'XX'
THEN ROUND ( TTFE.CRT+ ( NVL (TTTD.samt, TTTD.TAMT) * TTFE.amrt), 2) * -1
ELSE 0
END
IN_AMT,
CASE
WHEN (SELECT COUNT (*) FROM TLSV WHERE LNM = :B9 AND DESCR = :B8 AND LV = TTFE.C1) >0 AND TTFE.TAMT <> 0
THEN TTFE.AMT * ( NVL (TTTD.samt, TTTD.TAMT)/ TTFE.TAMT)
ELSE 0
END
as_amt,
CASE
WHEN TLSV.LV IS NOT NULL
THEN (TTFE.AMT / TTFE.FCNT)
END
DAMT,
TFA.D_CHG / 100 D_CHG,
TFPA.PCD|| '*'| TRIM (TFPA.text1)|| '*'|| TFPA.TEXT3|| '*'|| TFPA.TEXT4|| '*'|| TFPA.TEXT5|| '*' Col_data,
DECODE (TTSFA.CT_AMT, 0, NULL, NULL, NULL, 1) CO_AMT,
DECODE (MACD, 'XX', TFMA.text2, NULL) S_CD,
DECODE (MACD, 'XX', TFMA.TEXT5, NULL) ,
DECODE (MACD, 'XX', TFMA.TEXT6, NULL) BC,
DECODE (MACD, 'XX', TFMA.TEXT8, NULL) MID,
DECODE (MACD, 'XX', TFMA.TEXT7, NULL) MTYP,
DECODE ( MACD, 'XX', TO_CHAR ( TO_DATE (TRIM (TFMA.TEXT9),'yyyymmdd'), 'mm/dd/yyyy'), NULL) S_DT,
DECODE (TFMA.text1, NULL, 'N', 'Y') ,
TTTD.DT_MTHD
FROM TTSFA , TTTD , TFA , TFMA, TTFE,
(SELECT .. ~7 columns.. FROM TFA WHERE PCD IN ('XX', 'XX')) TFPA,
(SELECT LV FROM TLSV WHERE LNM = 'XXXX' AND DESCR = 'XXXXXXXXX') TLSV
WHERE TTSFA.did = TTTD.did AND TFPA.SDT(+) = TTTD.SDT AND TFPA.TDID(+) = TTTD.TDID
AND TLSV.LV(+) = TTFE.C1 AND TTSFA.FDT = TTTD.SDT AND TTSFA.OSID = TTTD.SID AND TTTD.TDID = TFA.TDID(+)
AND TTTD.SDT = TFA.SDT(+) AND TFA.UBD(+) = :B4 AND TTTD.TDID = TFMA.TDID(+) AND TTTD.SDT = TFMA.SDT(+)
AND TFMA.MACD(+) = DECODE (TTTD.MPCD, 'XX', :B3, :B2) AND TTTD.DEP_IND = :B1
AND TTSFA.did = TTTD.did AND TTSFA.SMID = TTTD.SMID AND TTFE.EID = TTSFA.did AND TTFE.IDT = TTSFA.SDT
AND TTFE.SID = TTSFA.SID AND TTFE.OTYP = 'TTFE' AND TTFE.OID = TTTD.SMID AND TTFE.BST = 'XX'
AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX'))
AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX'))
AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX')) AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX'))
AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX'))
AND ( ( PTCD = 'XX' AND TTFE.fsid IS NOT NULL)
OR (PTCD <> 'XX'))))
GROUP BY .... ~46 columns
On Thu, Dec 2, 2021 at 11:17 PM Mark W. Farnham <mwf_at_rsiz.com> wrote:
all the stuff JL mentioned, plus a little bit that didn’t make it to oracle-l before, probably because my remember to snip circuit is failing:
Things I would try:
- force parallel local for the query
- if 1) alone fails, do 1) and on the instance of choice run a scan of a non-indexed column of key tables of the query SIMPLY on the instance of choice
- if 1-2) fails, try a prelude of enumerating the disjoint partition enumerators from “43 | PARTITION RANGE ITERATOR “ and generate the union all query of the separate partitions individually, thus making each hash memory set requirement smaller, doing these serially, rather than in parallel (which might not be possible in 11 anyway) because your problem is memory, not elapsed time. IF a single one of your partitions is nearly everything, it still will probably go splat, but if they are relatively flat each one being significantly smaller might avoid the error. IF figuring out the enumerated list of partition predicates is fast, this also has a *chance* to make everything faster, especially if the CBO finds a better plan for a single partition at a time but also because each hash area creation and probe might be a fraction of the current single one. IF you have time to experiment, this *might* be a consistent winner on both speed and footprint, but your mileage may vary.
- forcing a sort group by *might* also work and would be easier for most folks to code up.
From: Mark W. Farnham [mailto:mwf_at_rsiz.com]
Sent: Thursday, December 02, 2021 10:04 AM
To: 'loknath.73_at_gmail.com'; 'Pap'
Cc: 'Oracle L'
Subject: RE: Error related to hash memory
Things I would try:
<snip>
-- http://www.freelists.org/webpage/oracle-lReceived on Sat Dec 04 2021 - 13:23:27 CET