Re: top-level DML under CTEs

The patch attached is based on the one rejected at the last CF for 9.0
last year.

http://archives.postgresql.org/message-id/16303.1266023203@sss.pgh.pa.us

This patch implements the feature that allows top-level DMLs under CTE
WITH clause. For example:

WITH t AS (SELECT * FROM x)
UPDATE y SET val = t.val FROM t
WHERE y.key = t.key;

This feature is part of writeable CTEs proposed by David Fetter originally.

There were two issues at the CF.

1. WITH clause atop INSERT
Although the previous discussion got the consensus that we forbid WITH
atop INSERT, it seems to me that it can be allowed. I managed to do it
by treating the top WITH clause (of INSERT) as if the one of SELECT
(or VALUES). It is possible to disallow the CTE over INSERT statement,
but the lack for INSERT, though there are for UPDATE and DELETE,
sounds inconsistent enough.

2. OLD/NEW in rules
Following the subsequent discussion after the post linked above, I add
code to throw an appropriate error when OLD/NEW is used in WITH
clauses. It is true that OLD/NEW references look sane to general
users, but actually (at least in our implementation) they are located
in the top-level query's Range Table List. Consequently, they are
invisible inside the WITH clause. To allow them, we should rewrite the
rule systems overall. Thus, we forbid them in WITH though we should
throw an error indicating appropriate message.

I'll add the entry to CF app later. Any feedback is welcome.

Regards,

--
Hitoshi Harada

On Mon, Sep 13, 2010 at 9:15 AM, Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> wrote:
> The patch attached is based on the one rejected at the last CF for 9.0
> last year.
>
> http://archives.postgresql.org/message-id/16303.1266023203@sss.pgh.pa.us
>
> This patch implements the feature that allows top-level DMLs under CTE
> WITH clause. For example:
>
> WITH t AS (SELECT * FROM x)
> UPDATE y SET val = t.val FROM t
> WHERE y.key = t.key;
>
> This feature is part of writeable CTEs proposed by David Fetter originally.

Thanks for pursuing this. I think this will be a useful feature if we
can get it committed, and plus David Fetter will be very, very happy.
:-)

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company

On Mon, Sep 13, 2010 at 9:20 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Mon, Sep 13, 2010 at 9:15 AM, Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> wrote:
>> The patch attached is based on the one rejected at the last CF for 9.0
>> last year.
>>
>> http://archives.postgresql.org/message-id/16303.1266023203@sss.pgh.pa.us
>>
>> This patch implements the feature that allows top-level DMLs under CTE
>> WITH clause. For example:
>>
>> WITH t AS (SELECT * FROM x)
>> UPDATE y SET val = t.val FROM t
>> WHERE y.key = t.key;
>>
>> This feature is part of writeable CTEs proposed by David Fetter originally.
>
> Thanks for pursuing this.  I think this will be a useful feature if we
> can get it committed, and plus David Fetter will be very, very happy.
> :-)

Just to be clear, the attached patch is missing the part of the wCTE
that allows queries of the form:
WITH foo AS (DELETE * FROM bar RETURNING *) <any query using foo>

IOW, your CTE query has to be a select. This is still highly useful
however. The patch itself looks very clean after a quick glance
(which is all I can offer ATM unfortunately).

merlin

On Mon, Sep 13, 2010 at 2:43 PM, Merlin Moncure <mmoncure(at)gmail(dot)com> wrote:
> Just to be clear, the attached patch is missing the part of the wCTE
> that allows queries of the form:
> WITH foo AS (DELETE * FROM bar RETURNING *) <any query using foo>

Understood.

> IOW, your CTE query has to be a select.  This is still highly useful
> however.

Agreed.

> The patch itself looks very clean after a quick glance
> (which is all I can offer ATM unfortunately).

Yeah, I haven't had a chance to look at it either, just wanted to send props.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company

2010/9/14 Merlin Moncure <mmoncure(at)gmail(dot)com>:
> On Mon, Sep 13, 2010 at 9:20 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> On Mon, Sep 13, 2010 at 9:15 AM, Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> wrote:
>>> The patch attached is based on the one rejected at the last CF for 9.0
>>> last year.
>>>
>>> http://archives.postgresql.org/message-id/16303.1266023203@sss.pgh.pa.us
>>>
>>> This patch implements the feature that allows top-level DMLs under CTE
>>> WITH clause. For example:
>>>
>>> WITH t AS (SELECT * FROM x)
>>> UPDATE y SET val = t.val FROM t
>>> WHERE y.key = t.key;
>>>
>>> This feature is part of writeable CTEs proposed by David Fetter originally.
>>
>> Thanks for pursuing this.  I think this will be a useful feature if we
>> can get it committed, and plus David Fetter will be very, very happy.
>> :-)
>
> Just to be clear, the attached patch is missing the part of the wCTE

Yes, the main part of wCTE that allows DMLs in WITH is still under
Marko Tikkaja. To work parallel, we split the tasks into pieces.

Regards,

--
Hitoshi Harada

On 2010-09-13 4:15 PM +0300, Hitoshi Harada wrote:
> 1. WITH clause atop INSERT
> Although the previous discussion got the consensus that we forbid WITH
> atop INSERT, it seems to me that it can be allowed. I managed to do it
> by treating the top WITH clause (of INSERT) as if the one of SELECT
> (or VALUES).

In the email you referred to, Tom was concerned about the case where
these WITH lists have different RECURSIVE declarations. This patch
makes both RECURSIVE if either of them is. I can think of cases where
that might lead to surprising behaviour, but the chances of any of those
happening in real life seem pretty slim.

> It is possible to disallow the CTE over INSERT statement,
> but the lack for INSERT, though there are for UPDATE and DELETE,
> sounds inconsistent enough.

Also because wCTEs are not allowed below the top level, not being able
to use INSERT as the top level statement would force people to wrap that
INSERT in another CTE.

Regards,
Marko Tiikkaja

2010/9/15 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
> On 2010-09-13 4:15 PM +0300, Hitoshi Harada wrote:
>>
>> 1. WITH clause atop INSERT
>> Although the previous discussion got the consensus that we forbid WITH
>> atop INSERT, it seems to me that it can be allowed. I managed to do it
>> by treating the top WITH clause (of INSERT) as if the one of SELECT
>> (or VALUES).
>
> In the email you referred to, Tom was concerned about the case where these
> WITH lists have different RECURSIVE declarations.  This patch makes both
> RECURSIVE if either of them is.  I can think of cases where that might lead
> to surprising behaviour, but the chances of any of those happening in real
> life seem pretty slim.

I might not understand the RECURSIVE issue correctly. I put my effort
to make such query

WITH RECURSIVE r AS (SELECT 1 i UNION ALL SELECT i + 1 FROM r WHERE i
< 10) INSERT INTO WITH t AS (SELECT 0) VALUES((SELECT * FROM r LIMIT
1)),((SELECT * FROM t));

look like

INSERT INTO WITH RECURSIVE r AS (SELECT 1 i UNION ALL SELECT i + 1
FROM r WHERE i < 10), t AS (SELECT 0) VALUES((SELECT * FROM r LIMIT
1)),((SELECT * FROM t));

Does that cause surprising behavior?

> but the chances of any of those happening in real
> life seem pretty slim.

The OLD/NEW issue is also near impossible to be problem in the real
life, except for the misleading error message. But once users see the
non-understandable behavior, they make lines to claim as it's a "bug".
So we need to put effort to avoid it as possible, I believe.

Regards,

--
Hitoshi Harada

Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> writes:
> 2010/9/15 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
>> In the email you referred to, Tom was concerned about the case where these
>> WITH lists have different RECURSIVE declarations. This patch makes both
>> RECURSIVE if either of them is. I can think of cases where that might lead
>> to surprising behaviour, but the chances of any of those happening in real
>> life seem pretty slim.

> Does that cause surprising behavior?

My recollection is that whether a CTE is marked RECURSIVE or not affects
its scope of visibility, so that confusing the two cases can result in
flat-out incorrect parser behavior.

It would probably be all right to combine the cases internally, at the
rewriter or planner stage. It's not okay to do it in the parser, not
even after doing parse analysis of the individual CTEs, because then it
would be impossible for ruleutils.c to reverse-list the query correctly.

regards, tom lane

On 2010-09-14 10:51 PM, Tom Lane wrote:
> Hitoshi Harada<umi(dot)tanuki(at)gmail(dot)com> writes:
>> 2010/9/15 Marko Tiikkaja<marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
>>> In the email you referred to, Tom was concerned about the case where these
>>> WITH lists have different RECURSIVE declarations. This patch makes both
>>> RECURSIVE if either of them is. I can think of cases where that might lead
>>> to surprising behaviour, but the chances of any of those happening in real
>>> life seem pretty slim.
>
>> Does that cause surprising behavior?
>
> My recollection is that whether a CTE is marked RECURSIVE or not affects
> its scope of visibility, so that confusing the two cases can result in
> flat-out incorrect parser behavior.

The worst I can think of is:

CREATE TABLE foo(a int);

WITH t AS (SELECT * FROM foo)
INSERT INTO bar
WITH RECURSIVE foo (SELECT 1 AS a)
SELECT * FROM t;

t will actually be populated with the results of the CTE, not the table foo.

I don't think this is a huge problem in real life, but if someone thinks
otherwise, I think we could just error out if the lists have a different
RECURSIVE definition.

Anyone have a worse example? Thoughts?

Regards,
Marko Tiikkaja

On Tue, Sep 14, 2010 at 4:28 PM, Marko Tiikkaja
<marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi> wrote:
> On 2010-09-14 10:51 PM, Tom Lane wrote:
>>
>> Hitoshi Harada<umi(dot)tanuki(at)gmail(dot)com>  writes:
>>>
>>> 2010/9/15 Marko Tiikkaja<marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
>>>>
>>>> In the email you referred to, Tom was concerned about the case where
>>>> these
>>>> WITH lists have different RECURSIVE declarations.  This patch makes both
>>>> RECURSIVE if either of them is.  I can think of cases where that might
>>>> lead
>>>> to surprising behaviour, but the chances of any of those happening in
>>>> real
>>>> life seem pretty slim.
>>
>>> Does that cause surprising behavior?
>>
>> My recollection is that whether a CTE is marked RECURSIVE or not affects
>> its scope of visibility, so that confusing the two cases can result in
>> flat-out incorrect parser behavior.
>
> The worst I can think of is:
>
> CREATE TABLE foo(a int);
>
> WITH t AS (SELECT * FROM foo)
> INSERT INTO bar
> WITH RECURSIVE foo (SELECT 1 AS a)
> SELECT * FROM t;
>
> t will actually be populated with the results of the CTE, not the table foo.

Unless I'm confused, that seems pretty clearly wrong.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company

Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi> writes:
> On 2010-09-14 10:51 PM, Tom Lane wrote:
>> My recollection is that whether a CTE is marked RECURSIVE or not affects
>> its scope of visibility, so that confusing the two cases can result in
>> flat-out incorrect parser behavior.

> The worst I can think of is:

> CREATE TABLE foo(a int);

> WITH t AS (SELECT * FROM foo)
> INSERT INTO bar
> WITH RECURSIVE foo (SELECT 1 AS a)
> SELECT * FROM t;

> t will actually be populated with the results of the CTE, not the table foo.

> I don't think this is a huge problem in real life, but if someone thinks
> otherwise, I think we could just error out if the lists have a different
> RECURSIVE definition.

Wrong is wrong. Doesn't matter whether it's "a huge problem in real life".

Why is it so difficult to do this correctly?

regards, tom lane

2010/9/15 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
> On 2010-09-14 10:51 PM, Tom Lane wrote:
>> My recollection is that whether a CTE is marked RECURSIVE or not affects
>> its scope of visibility, so that confusing the two cases can result in
>> flat-out incorrect parser behavior.
>
> The worst I can think of is:
>
> CREATE TABLE foo(a int);
>
> WITH t AS (SELECT * FROM foo)
> INSERT INTO bar
> WITH RECURSIVE foo (SELECT 1 AS a)
> SELECT * FROM t;
>
> t will actually be populated with the results of the CTE, not the table foo.

Hmmm, that's true. But it seems unrelated to RECURSIVE option, right?

Regards,

--
Hitoshi Harada

2010/9/15 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi> writes:
>> On 2010-09-14 10:51 PM, Tom Lane wrote:
>>> My recollection is that whether a CTE is marked RECURSIVE or not affects
>>> its scope of visibility, so that confusing the two cases can result in
>>> flat-out incorrect parser behavior.
>
>> The worst I can think of is:
>
>> CREATE TABLE foo(a int);
>
>> WITH t AS (SELECT * FROM foo)
>> INSERT INTO bar
>> WITH RECURSIVE foo (SELECT 1 AS a)
>> SELECT * FROM t;
>
>> t will actually be populated with the results of the CTE, not the table foo.
>
>> I don't think this is a huge problem in real life, but if someone thinks
>> otherwise, I think we could just error out if the lists have a different
>> RECURSIVE definition.
>
> Wrong is wrong.  Doesn't matter whether it's "a huge problem in real life".
>
> Why is it so difficult to do this correctly?

Because INSERT INTO ... (SELECT|VALUES) is already a collection of
kludge (as comments say). It was possible to parse the two WITHs
separately, but it results in ambiguous naming issue;
parseWithClause() asserts there's only one WITH clause in the Stmt and
detects duplicated CTE name in it. It seems possible to call
parseWithClause() twice by cheating ParseState and to try to find name
duplication outside it, though it is another kludge :-(

Now that we find the worst situation, I start to think I have to take
the kludy way anyway.

Regards,

--
Hitoshi Harada

Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> writes:
> 2010/9/15 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
>> Why is it so difficult to do this correctly?

> Because INSERT INTO ... (SELECT|VALUES) is already a collection of
> kludge (as comments say). It was possible to parse the two WITHs
> separately, but it results in ambiguous naming issue;
> parseWithClause() asserts there's only one WITH clause in the Stmt and
> detects duplicated CTE name in it.

Well, I would think that the no-duplication rule applies to each WITH
list separately, not both together. If you do something like

with t1 as (select * from foo)
select * from
(with t2 as (select * from foo)
select * from t1, t2) ss;

there's no expectation that the WITH clauses can't both define the same
name.

regards, tom lane

2010/9/15 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com> writes:
>> 2010/9/15 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
>>> Why is it so difficult to do this correctly?
>
>> Because INSERT INTO ... (SELECT|VALUES) is already a collection of
>> kludge (as comments say). It was possible to parse the two WITHs
>> separately, but it results in ambiguous naming issue;
>> parseWithClause() asserts there's only one WITH clause in the Stmt and
>> detects duplicated CTE name in it.
>
> Well, I would think that the no-duplication rule applies to each WITH
> list separately, not both together.  If you do something like
>
> with t1 as (select * from foo)
>  select * from
>    (with t2 as (select * from foo)
>       select * from t1, t2) ss;
>

Well, I didn't know it is allowed. That would look like the way to go.

Regards,

--
Hitoshi Harada

2010/9/15 Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com>:
> 2010/9/15 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
>>
>> Well, I would think that the no-duplication rule applies to each WITH
>> list separately, not both together.  If you do something like
>>
>> with t1 as (select * from foo)
>>  select * from
>>    (with t2 as (select * from foo)
>>       select * from t1, t2) ss;
>>
>
> Well, I didn't know it is allowed. That would look like the way to go.

I made changes to the previous version, so that it avoids to resolve
CTE name duplication.

regression=# with t as (select 1 as i) insert into z with t as(select
2 as i )values ((select * from t));
INSERT 0 1
Time: 1.656 ms
regression=# table z;
f3
----
2
(1 row)

Also, the sample Marko gave is OK.

> CREATE TABLE foo(a int);
>
> WITH t AS (SELECT * FROM foo)
> INSERT INTO bar
> WITH RECURSIVE foo (SELECT 1 AS a)
> SELECT * FROM t;
>

Hope this covers all the cases.

Regards,

--
Hitoshi Harada

On 2010-09-17 4:48 AM, Hitoshi Harada wrote:
> 2010/9/15 Hitoshi Harada<umi(dot)tanuki(at)gmail(dot)com>:
>> Well, I didn't know it is allowed. That would look like the way to go.
>
> I made changes to the previous version, so that it avoids to resolve
> CTE name duplication.

This patch still doesn't address the issue Tom raised here:
http://archives.postgresql.org/pgsql-hackers/2010-09/msg00753.php

For WITH .. INSERT .. WITH .. SELECT ..; this patch works OK, but not so
much for VALUES:

=# CREATE RULE barrule AS ON UPDATE TO bar DO INSTEAD
-# WITH RECURSIVE t AS (SELECT -1)
-# INSERT INTO bar
-# WITH t AS (SELECT 1)
-# VALUES((SELECT * FROM t));
CREATE RULE

=# \d bar
Table "public.bar"
Column | Type | Modifiers
--------+---------+-----------
a | integer |
Rules:
barrule AS
ON UPDATE TO bar DO INSTEAD WITH RECURSIVE t AS (
SELECT 1
), t AS (
SELECT (-1)
)
INSERT INTO bar (a) WITH RECURSIVE t AS (
SELECT 1
), t AS (
SELECT (-1)
)

VALUES (( SELECT t."?column?"
FROM t))

Regards,
Marko Tiikkaja

2010/9/23 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
> On 2010-09-17 4:48 AM, Hitoshi Harada wrote:
>>
>> 2010/9/15 Hitoshi Harada<umi(dot)tanuki(at)gmail(dot)com>:
>>>
>>> Well, I didn't know it is allowed. That would look like the way to go.
>>
>> I made changes to the previous version, so that it avoids to resolve
>> CTE name duplication.
>
> This patch still doesn't address the issue Tom raised here:
> http://archives.postgresql.org/pgsql-hackers/2010-09/msg00753.php
>
> For WITH .. INSERT .. WITH .. SELECT ..; this patch works OK, but not so
> much for VALUES:
>
> =# CREATE RULE barrule AS ON UPDATE TO bar DO INSTEAD
> -# WITH RECURSIVE t AS (SELECT -1)
> -# INSERT INTO bar
> -# WITH t AS (SELECT 1)
> -# VALUES((SELECT * FROM t));
> CREATE RULE
>
> =# \d bar
>      Table "public.bar"
>  Column |  Type   | Modifiers
> --------+---------+-----------
>  a      | integer |
> Rules:
>    barrule AS
>    ON UPDATE TO bar DO INSTEAD  WITH RECURSIVE t AS (
>         SELECT 1
>        ), t AS (
>         SELECT (-1)
>        )
>  INSERT INTO bar (a)  WITH RECURSIVE t AS (
>                 SELECT 1
>                ), t AS (
>                 SELECT (-1)
>                )
>
>  VALUES (( SELECT t."?column?"
>           FROM t))

I ran the sql and recognized what is wrong. In VALUES case, the WITH
table "t" is copied in one list and shown up in the both of
INSERT-level WITH and SELECT-level WITH. Since the transformation of
WITH clause to cheat postgres is in the parser stage currently, I
wonder if this should be done in the rewriter or the planner stage.

Thanks for the report. Next time, please point the clear problem in
English aside the sample.

Regards,

--
Hitoshi Harada

On 2010-09-23 9:12 AM +0300, Hitoshi Harada wrote:
> Thanks for the report. Next time, please point the clear problem in
> English aside the sample.

I apologize. The problem was exactly the one pointed out in the email I
referred to, so I assumed that further explanation was not necessary.

I will try to be more clear in the future.

Regards,
Marko Tiikkaja

On 2010-09-23 9:12 AM +0300, Hitoshi Harada wrote:
> Since the transformation of
> WITH clause to cheat postgres is in the parser stage currently, I
> wonder if this should be done in the rewriter or the planner stage.

It's been about a week now. Should we expect a new patch soon?

Regards,
Marko Tiikkaja

2010/9/30 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
> On 2010-09-23 9:12 AM +0300, Hitoshi Harada wrote:
>>
>> Since the transformation of
>> WITH clause to cheat postgres is in the parser stage currently, I
>> wonder if this should be done in the rewriter or the planner stage.
>
> It's been about a week now.  Should we expect a new patch soon?
>
>

Yep, I'm working it now. You'll see the conclusion in a day or so.

Regards,

--
Hitoshi Harada

2010/10/1 Hitoshi Harada <umi(dot)tanuki(at)gmail(dot)com>:
> 2010/9/30 Marko Tiikkaja <marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi>:
>> On 2010-09-23 9:12 AM +0300, Hitoshi Harada wrote:
>>>
>>> Since the transformation of
>>> WITH clause to cheat postgres is in the parser stage currently, I
>>> wonder if this should be done in the rewriter or the planner stage.
>>
>> It's been about a week now.  Should we expect a new patch soon?
>>
>>
>
> Yep, I'm working it now. You'll see the conclusion in a day or so.

...and attached is the latest patch. It contains LIMIT etc. bug of
INSERT fixes and I confirmed the barrule case correctly in this
version.

> =# CREATE RULE barrule AS ON UPDATE TO bar DO INSTEAD
> -# WITH RECURSIVE t AS (SELECT -1)
> -# INSERT INTO bar
> -# WITH t AS (SELECT 1)
> -# VALUES((SELECT * FROM t));

Regards,

--
Hitoshi Harada

On Sun, October 3, 2010 15:47, Hitoshi Harada wrote:

[...]

> ...and attached is the latest patch. It contains LIMIT etc. bug of
> INSERT fixes and I confirmed the barrule case correctly in this
> version.
>

(HEAD from git://git.postgresql.org/git/postgresql.git)

The patch applies only with error.
If that error is ignored, the regression 'with' test failes.
If that is also ignored, it runs.

I thought I'd give you the errors anyway:

patch --strip=1 < toplevel-dml-cte.20101003.diff
patching file doc/src/sgml/ref/delete.sgml
patching file doc/src/sgml/ref/insert.sgml
patching file doc/src/sgml/ref/update.sgml
patching file src/backend/nodes/copyfuncs.c
patching file src/backend/nodes/equalfuncs.c
patching file src/backend/parser/analyze.c
Hunk #2 FAILED at 350.
Hunk #3 succeeded at 397 (offset 9 lines).
Hunk #5 succeeded at 630 (offset 9 lines).
Hunk #7 succeeded at 1800 (offset 9 lines).
1 out of 7 hunks FAILED -- saving rejects to file src/backend/parser/analyze.c.rej
patching file src/backend/parser/gram.y
patching file src/backend/parser/parse_utilcmd.c
patching file src/backend/rewrite/rewriteManip.c
patching file src/backend/utils/adt/ruleutils.c
patching file src/include/nodes/parsenodes.h
patching file src/include/rewrite/rewriteManip.h
patching file src/test/regress/expected/with.out
patching file src/test/regress/sql/with.sql
-------------------8<-------------------
***************
*** 343,354 ****
qry->commandType = CMD_INSERT;
pstate->p_is_insert = true;

/*
* We have three cases to deal with: DEFAULT VALUES (selectStmt == NULL),
- * VALUES list, or general SELECT input. We special-case VALUES, both for
- * efficiency and so we can handle DEFAULT specifications.
*/
- isGeneralSelect = (selectStmt && selectStmt->valuesLists == NIL);

/*
* If a non-nil rangetable/namespace was passed in, and we are doing
--- 350,375 ----
qry->commandType = CMD_INSERT;
pstate->p_is_insert = true;

+ /* process the WITH clause independently of all else */
+ if (stmt->withClause)
+ {
+ qry->hasRecursive = stmt->withClause->recursive;
+ qry->cteList = transformWithClause(pstate, stmt->withClause);
+ }
+
/*
* We have three cases to deal with: DEFAULT VALUES (selectStmt == NULL),
+ * simple VALUES list, or general SELECT input including complex VALUES.
+ * We special-case VALUES, both for efficiency and so we can handle
+ * DEFAULT specifications. In a complex VALUES case, which means the list
+ * has any of ORDER BY, OFFSET, LIMIT or WITH, we don't accept DEFAULT
+ * in it; The spec may require it but for now we reject it from point of
+ * code base and expected use cases.
*/
+ isGeneralSelect = (selectStmt &&
+ (selectStmt->valuesLists == NIL ||
+ selectStmt->sortClause || selectStmt->limitOffset ||
+ selectStmt->limitCount || selectStmt->withClause));

/*
* If a non-nil rangetable/namespace was passed in, and we are doing
-------------------8<-------------------

Continuing after that error:
make OK;
make check:

[...]
largeobject ... ok
with ... FAILED
xml ... ok
[...]

regression.diffs:

*** /var/data1/pg_stuff/pg_sandbox/pgsql.dml_cte/src/test/regress/expected/with.out 2010-10-04
13:25:26.000000000 +0200
--- /var/data1/pg_stuff/pg_sandbox/pgsql.dml_cte/src/test/regress/results/with.out 2010-10-04
13:28:20.000000000 +0200
***************
*** 747,783 ****
)
INSERT INTO y
SELECT a+20 FROM t RETURNING *;
! a
! ----
! 21
! 22
! 23
! 24
! 25
! 26
! 27
! 28
! 29
! 30
! (10 rows)
!
WITH t AS (
SELECT a FROM y
)
UPDATE y SET a = y.a-10 FROM t WHERE y.a > 20 AND t.a = y.a RETURNING y.a;
! a
! ----
! 11
! 12
! 13
! 14
! 15
! 16
! 17
! 18
! 19
! 20
! (10 rows)

WITH RECURSIVE t(a) AS (
SELECT 11
--- 747,762 ----
)
INSERT INTO y
SELECT a+20 FROM t RETURNING *;
! ERROR: relation "t" does not exist
! LINE 5: SELECT a+20 FROM t RETURNING *;
! ^
WITH t AS (
SELECT a FROM y
)
UPDATE y SET a = y.a-10 FROM t WHERE y.a > 20 AND t.a = y.a RETURNING y.a;
! a
! ---
! (0 rows)

WITH RECURSIVE t(a) AS (
SELECT 11
***************
*** 785,803 ****
SELECT a+1 FROM t WHERE a < 50
)
DELETE FROM y USING t WHERE t.a = y.a RETURNING y.a;
! a
! ----
! 11
! 12
! 13
! 14
! 15
! 16
! 17
! 18
! 19
! 20
! (10 rows)

SELECT * FROM y;
a
--- 764,772 ----
SELECT a+1 FROM t WHERE a < 50
)
DELETE FROM y USING t WHERE t.a = y.a RETURNING y.a;
! a
! ---
! (0 rows)

SELECT * FROM y;
a

======================================================================

hth,

Erik Rijkers

On 2010-10-04 2:46 PM +0300, Erik Rijkers wrote:
> (HEAD from git://git.postgresql.org/git/postgresql.git)
>
> The patch applies only with error.
> If that error is ignored, the regression 'with' test failes.
> If that is also ignored, it runs.

This patch conflicted with Tom's WITH .. INSERT change. I tweaked that
patch just a bit and it now passes all regression tests so I can review
it. New version attached for documentation purposes.

Regards,
Marko Tiikkaja

On Mon, Oct 4, 2010 at 5:59 PM, Marko Tiikkaja
<marko(dot)tiikkaja(at)cs(dot)helsinki(dot)fi> wrote:
> On 2010-10-04 2:46 PM +0300, Erik Rijkers wrote:
>>
>> (HEAD from git://git.postgresql.org/git/postgresql.git)
>>
>> The patch applies only with error.
>> If that error is ignored, the regression 'with' test failes.
>> If that is also ignored, it runs.
>
> This patch conflicted with Tom's WITH .. INSERT change.  I tweaked that
> patch just a bit and it now passes all regression tests so I can review it.
>  New version attached for documentation purposes.

Guys -

PLEASE redirect this conversation to -hackers!

Thanks,

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company