Обсуждение: Sequential Scans On Complex Query With UNION - see why this fails
PostgreSql 8.0.3
I have a complex query that I want to read the data in a hierarchical manner - ie from master table filtered rows to sub set1, from subset1 to subset2 etc. The query is in the attached file, as is the explain. What I do not understand is why I get sequential scans on certain tables - ie the purchaseOrdDet, product, supplierProduct when each of the joins linking the tables hits a valid 'record number' based index.
The query is trying to find all purchase orders within a period, then all the lines for those orders, products for those lines etc. The union selects from the supplierproduct/product tables on both sides - in one case it uses the index and on the other does a sequential scan.
I am obviously doing something wrong in the structure of the query - any ideas?
Also,...as a quick 'method' question..when writing embedded joins, which syntax is easier for the optimiser? Should you do:
FROM table tableA AS A
JOIN tableB AS B
JOIN tableC AS C
JOIN tableD AS D
ON C.joinD = D.join
JOIN tableE AS E
ON c.joinE = E.join
ON B.joinC = C.join
ON A.joinB = B.join
OR
FROM table tableA AS A
JOIN tableB AS B
ON A.joinB = B.join
JOIN tableC AS C
ON B.joinC = C.join
JOIN tableD AS D
ON C.joinD = D.join
JOIN tableE AS E
ON c.joinE = E.join
OR is it immaterial and just a matter of style?
| Regards, Steve Tucknott |
Вложения
Did anyone have any ideas on this? If not, I'll look at rewriting the query completely.
On Mon, 2006-01-02 at 09:10 +0000, Steve Tucknott wrote:
On Mon, 2006-01-02 at 09:10 +0000, Steve Tucknott wrote:
PostgreSql 8.0.3
I have a complex query that I want to read the data in a hierarchical manner - ie from master table filtered rows to sub set1, from subset1 to subset2 etc. The query is in the attached file, as is the explain. What I do not understand is why I get sequential scans on certain tables - ie the purchaseOrdDet, product, supplierProduct when each of the joins linking the tables hits a valid 'record number' based index.
The query is trying to find all purchase orders within a period, then all the lines for those orders, products for those lines etc. The union selects from the supplierproduct/product tables on both sides - in one case it uses the index and on the other does a sequential scan.
I am obviously doing something wrong in the structure of the query - any ideas?
Also,...as a quick 'method' question..when writing embedded joins, which syntax is easier for the optimiser? Should you do:
FROM table tableA AS A
JOIN tableB AS B
JOIN tableC AS C
JOIN tableD AS D
ON C.joinD = D.join
JOIN tableE AS E
ON c.joinE = E.join
ON B.joinC = C.join
ON A.joinB = B.join
OR
FROM table tableA AS A
JOIN tableB AS B
ON A.joinB = B.join
JOIN tableC AS C
ON B.joinC = C.join
JOIN tableD AS D
ON C.joinD = D.join
JOIN tableE AS E
ON c.joinE = E.join
OR is it immaterial and just a matter of style?
Regards,
Steve Tucknott
---------------------------(end of broadcast)--------------------------- TIP 9: In versions below 8.0, the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
| Regards, Steve Tucknott ReTSol Ltd DDI: 01903 828769 Mobile: 0773 671 5772 |
Steve,
I don't have any explanation as to why your query is using an index
in one case and a sequential scan in another. However, I do have a
few observations after looking at your code that you might find helpful.
On Jan 2, 2006, at 3:10 , Steve Tucknott wrote:
> I am obviously doing something wrong in the structure of the query
> - any ideas?
When I first saw your FROM clause, I didn't even know that moving the
ON clause away from the JOIN it was related to would even work. I was
surprised to find that it does! I find it quite counterintuitive to
separate a join and its condition. As I understand it,
select *
from a
join b
join c on c.b_name = b.b_name
on b.a_name = a.a_name;
is just another way of writing
select *
from a
join b on b.a_name = a.a_name
join c on c.b_name = b.b_name;
and doesn't make a difference in query planning or performance. It's
definitely not 'hierarchical' as I understand it. Comparing EXPLAIN
ANALYZE output for the two query forms will show you if this is the
case.
In the queries you attached, I saw that you mix restrictions (e.g.,
c.b_name = 'some value') and join conditions (e.g., c.b_name =
b.b_name) in both the FROM clause and the WHERE clause. I personally
find this quite hard to read, though I don't think it makes much
difference to the server. I believe both join conditions and
restrictions are rewritten as they'd appear in the WHERE clause, so
you *could* put them all in the WHERE clause.
However, I find it useful to separate them out, and build queries
using JOIN ... ON (and actually I prefer the JOIN ... USING syntax
because only one set of join columns appears in the result). This
helps me clearly see how the tables are related. I then place all
restrictions in the WHERE clause. This helps me know where to look in
the query to find what I'm looking for. The optimizer should be able
to perform restrictions as necessary—not necessarily in the order the
original query was written. After all, that's the reason for its
existence :).
I also noticed was that you've used column numbers rather than column
names in your ORDER BY clause. I recommend using column names, as it
makes it easier to really see what you're ordering by without having
to refer to the SELECT target list. (In relational theory attribute
order does not matter. The SQL standard *does* refer to column order,
but you can go a long way and not have to worry about it, other than
for UNION compatibility, and even then, you don't need to use column
numbers explicitly.)
I've provided a rewritten query at the bottom of this post. I'd be
interested to see if it performs any differently that your original
query (e.g., how their EXPLAIN ANALYZE output compares).
The output also includes quite a few columns with duplicate column
names (e.g., clnt.name, clntB.name; srcD.recNo, clnt.recNo,
clntB.recNo, srcU.recNo). If feasible, I'd use AS to provide unique
column names, as this can make it easier to reference the columns
either in your client application or if you use this query as a
subquery in another statement. (I haven't done this below. There
could be a number of additional duplicates lurking in sProd.*,
prod.*, VAT.*, and mrkUP.* .)
What does the 1 = 1 do in your WHERE clause?
Also, the mixed case column and table names you use get lowered in
PostgreSQL, but you've probably already noticed that. To maintain
case, you need to double-quote these identifiers.
> Also,...as a quick 'method' question..when writing embedded joins,
> which syntax is easier for the optimiser? Should you do:
EXPLAIN ANALYZE will help you determine which form is more efficient
in your situation. Though as I've indicated above, I'd rewrite the
queries to make it easier for the programmer. Start by writing your
query as you think it *should* be written to make it easiest for
*you* to use. Then see how it performs. If you find it doesn't
perform well, you'll have a baseline against which you can compare
any improvement, and you'll learn what does and does not work. The
optimizer is pretty good—don't second guess it. Code should be easy
for the programmer to read and use first.
Hope this helps!
Michael Glaesemann
grzm myrealbox com
--- original query
--- rewritten query
SELECT clnt.name
, clntB.name
, periodOf(pOHdr.orderDate)
, srcD.recNo
, pODet.purchaseOrdHdrRecNo
, clnt.recNo
, clntB.recNo
, srcU.recNo, srcU.replacementCost, srcU.requiredQty
, sProd.*
, prod.*
, VAT.*
, pODet.orderedQty
, mrkUp.*
FROM purchaseOrdHdr AS pOHdr
JOIN purchaseOrdDet AS pODet ON (pODet.purchaseOrdHdrRecNo =
pOHdr.recNo)
JOIN sourceDetUpgr AS srcU ON (srcU.recNo = pODet.sourceDetUpgrRecNo)
JOIN sourceDet AS srcD ON (srcD.recNo = srcU.sourceDetRecNo)
JOIN sourceDetExtref AS srcE ON (srcE.foreignRecNo = srcD.recNo)
JOIN clientBranch AS clntB ON (clntB.recNo = srcE.ownerForeignRecNo)
JOIN client AS clnt ON (clntB.clientRecNo = clnt.recNo)
JOIN markUp AS mrkUp ON (mrkUp.foreignRecNo = clnt.recNo)
JOIN supplierProduct AS sProd ON (sProd.recNo =
srcU.supplierProductRecNo)
JOIN product AS prod ON (prod.recNo = sProd.productRecNo)
JOIN VAT AS VAT ON (sProd.VATRecNo = VAT.recNo)
WHERE pOHdr.orderDate BETWEEN '01/12/2005' AND '21/12/2005'
AND 1=1 -- what does this do?
AND pODet.lineStatus != 'V'
AND srcD.actionStatus != 'V'
AND srcD.serviceCoBranchRecNo = 2
AND srcU.lineStatus != 'V'
AND srcE.tableName = 'sourcedet' -- moved down from join condition
AND srcE.ownerForeignTableName = 'clientbranch' -- moved down
from join condition
AND mrkUp.foreignTableName = 'client' -- moved down from join
condition
UNION
SELECT clnt.name
, clntB.name
, periodOf(srcD.enteredDate)
, srcD.recNo
, NULL
, clnt.recNo
, clntB.recNo
, srcU.recNo, srcU.replacementCost, srcU.requiredQty
, sProd.*
, prod.*
, VAT.*
, 0
, mrkUp.*
FROM sourceDet AS srcD
JOIN sourceDetUpgr AS srcU ON (srcU.sourceDetRecNo = srcD.recNo)
JOIN supplierProduct AS sProd ON (srcU.supplierProductRecNo =
sProd.recNo)
JOIN product AS prod ON (sProd.productRecNo = prod.recNo)
JOIN VAT AS VAT ON (sProd.VATRecNo = VAT.recNo)
JOIN sourceDetExtref AS srcE ON (srcE.foreignRecNo = srcD.recNo)
JOIN clientBranch AS clntB ON (clntB.recNo = srcE.ownerForeignRecNo)
JOIN client AS clnt ON (clntB.clientRecNo = clnt.recNo)
JOIN markUp AS mrkUp ON (mrkUp.foreignRecNo = clnt.recNo)
WHERE srcD.enteredDate BETWEEN '01/12/2005' AND '21/12/2005'
AND srcD.serviceCoBranchRecNo = 2
AND srcD.actionStatus != 'V'
AND srcU.lineStatus = 'S'
AND mrkUp.foreignTableName = 'client' -- from join condition
AND srcE.tableName = 'sourcedet' -- from join condition
AND srcE.ownerForeignTableName = 'clientbranch' -- from join
condition
AND (
SELECT COUNT(*)
FROM sourceDetUpgr AS srcU2
WHERE srcU2.sourceDetRecNo = srcD.recNo
AND srcU2.lineStatus IN ('S','T')
) = (
SELECT COUNT(*)
FROM sourceDetUpgr AS srcU2
WHERE srcU2.sourceDetRecNo = srcD.recNo
AND srcU2.lineStatus != 'V'
)
AND 1=1 -- what does this do?
ORDER BY --3,2,1,4,5
periodof
, clntB.name
, clnt.name
, srD.recNo
, pODet.purchaseOrdHdrRecNo
--- original query
SELECT clnt.name,clntB.name,periodOf
(pOHdr.orderDate),srcD.recNo,pODet.purchaseOrdHdrRecNo,clnt.recNo,clntB.
recNo,srcU.recNo,srcU.replacementCost,srcU.requiredQty,sProd.*,prod.*,VA
T.*,pODet.orderedQty,mrkUp.*
FROM purchaseOrdHdr AS pOHdr
JOIN purchaseOrdDet AS pODet
JOIN sourceDetUpgr AS srcU
JOIN sourceDet AS srcD
JOIN sourceDetExtref AS srcE
JOIN clientBranch AS clntB
JOIN client AS clnt
JOIN markUp AS mrkUp
ON mrkUp.foreignRecNo = clnt.recNo
AND mrkUp.foreignTableName = 'client'
ON clntB.clientRecNo = clnt.recNo
ON clntB.recNo =
srcE.ownerForeignRecNo
ON srcE.foreignRecNo = srcD.recNo
AND srcE.tableName = 'sourcedet'
AND srcE.ownerForeignTableName = 'clientbranch'
ON srcD.recNo = srcU.sourceDetRecNo
JOIN supplierProduct AS sProd
JOIN product AS prod
ON prod.recNo = sProd.productRecNo
JOIN VAT AS VAT
ON sProd.VATRecNo = VAT.recNo
ON sProd.recNo = srcU.supplierProductRecNo
ON srcU.recNo = pODet.sourceDetUpgrRecNo
ON pODet.purchaseOrdHdrRecNo = pOHdr.recNo
WHERE pOHdr.orderDate BETWEEN '01/12/2005' AND '21/12/2005'
AND 1=1
AND pODet.lineStatus != 'V'
AND srcD.actionStatus != 'V'
AND srcD.serviceCoBranchRecNo = 2
AND srcU.lineStatus != 'V'
UNION
SELECT clnt.name,clntB.name,periodOf
(srcD.enteredDate),srcD.recNo,NULL,clnt.recNo,clntB.recNo,srcU.recNo,
srcU.replacementCost,srcU.requiredQty,sProd.*,prod.*,VAT.*,
0,mrkUp.*
FROM sourceDet AS srcD
JOIN sourceDetUpgr AS srcU
JOIN supplierProduct AS sProd
JOIN product AS prod
ON sProd.productRecNo = prod.recNo
JOIN VAT AS VAT
ON sProd.VATRecNo = VAT.recNo
ON srcU.supplierProductRecNo = sProd.recNo
ON srcU.sourceDetRecNo = srcD.recNo
JOIN sourceDetExtref AS srcE
JOIN clientBranch AS clntB
JOIN client AS clnt
JOIN markUp AS mrkUp
ON mrkUp.foreignRecNo = clnt.recNo
AND mrkUp.foreignTableName = 'client'
ON clntB.clientRecNo = clnt.recNo
ON clntB.recNo = srcE.ownerForeignRecNo
ON srcE.foreignRecNo = srcD.recNo
AND srcE.tableName = 'sourcedet'
AND srcE.ownerForeignTableName = 'clientbranch'
WHERE srcD.enteredDate BETWEEN '01/12/2005' AND '21/12/2005'
AND srcD.serviceCoBranchRecNo = 2
AND srcD.actionStatus != 'V'
AND srcU.lineStatus = 'S'
AND (SELECT COUNT(*) FROM sourceDetUpgr AS srcU2
WHERE srcU2.sourceDetRecNo = srcD.recNo
AND srcU2.lineStatus IN ('S','T')) = (SELECT COUNT(*) FROM
sourceDetUpgr AS srcU2
WHERE
srcU2.sourceDetRecNo = srcD.recNo
AND
srcU2.lineStatus != 'V')
AND 1=1
ORDER BY 3,2,1,4,5
Michael Glaesemann <grzm@myrealbox.com> writes:
> When I first saw your FROM clause, I didn't even know that moving the
> ON clause away from the JOIN it was related to would even work. I was
> surprised to find that it does! I find it quite counterintuitive to
> separate a join and its condition. As I understand it,
I think what's happening is that the parser implicitly parenthesizes
like this:
from ((a join b on a.x=b.y) join c on b.y=c.z)
versus this:
from (a join (b join c on b.y=c.z) on a.x=b.y)
These are equivalent formulations for inner joins, but they are
decidedly not equivalent if one or both joins is OUTER. I'd tend
to agree with Michael's advice to keep the join condition close
to the pair of tables you think it's joining. Even better, use
parentheses to make sure the parser reads it the same as you do.
> I believe both join conditions and
> restrictions are rewritten as they'd appear in the WHERE clause, so
> you *could* put them all in the WHERE clause.
Again, this is true for inner joins but very much not the case if
any outer joins are involved.
regards, tom lane
On Jan 12, 2006, at 12:21 , Tom Lane wrote: > I think what's happening is that the parser implicitly parenthesizes > like this: > > from ((a join b on a.x=b.y) join c on b.y=c.z) Any idea off hand if the SQL spec has anything to say on the subject? I haven't dug into my local version of the final draft to see. >> I believe both join conditions and >> restrictions are rewritten as they'd appear in the WHERE clause, so >> you *could* put them all in the WHERE clause. > > Again, this is true for inner joins but very much not the case if > any outer joins are involved. Thanks for the clarification, Tom. While I was writing I thought that OUTER JOIN was an exception, but I neglected to look into it or include my thoughts. Michael Glaesemann grzm myrealbox com
Michael Glaesemann <grzm@myrealbox.com> writes:
> On Jan 12, 2006, at 12:21 , Tom Lane wrote:
>> I think what's happening is that the parser implicitly parenthesizes
>> like this:
>>
>> from ((a join b on a.x=b.y) join c on b.y=c.z)
> Any idea off hand if the SQL spec has anything to say on the subject?
SQL92 has this BNF:
<from clause> ::= FROM <table reference> [ { <comma> <table reference> }... ]
<table reference> ::=
<table name> [ [ AS ] <correlation name>
[ <left paren> <derived column list> <right paren> ] ]
| <derived table> [ AS ] <correlation name>
[ <left paren> <derived column list> <right paren> ]
| <joined table>
<derived table> ::= <table subquery>
<derived column list> ::= <column name list>
<joined table> ::=
<cross join>
| <qualified join>
| <left paren> <joined table> <right paren>
<cross join> ::=
<table reference> CROSS JOIN <table reference>
<qualified join> ::=
<table reference> [ NATURAL ] [ <join type> ] JOIN
<table reference> [ <join specification> ]
<join specification> ::=
<join condition>
| <named columns join>
<join condition> ::= ON <search condition>
<named columns join> ::=
USING <left paren> <join column list> <right paren>
<join type> ::=
INNER
| <outer join type> [ OUTER ]
| UNION
<outer join type> ::=
LEFT
| RIGHT
| FULL
What we're talking about is the <qualified join> production, whose
inputs are <table reference>s, and a <table reference> can be another
<qualified join> with or without surrounding parentheses. So AFAICS
SQL92 specifically allows both of these constructions.
regards, tom lane