On 2010-02-08 18:42 +0200, Robert Haas wrote:
> On Thu, Feb 4, 2010 at 11:57 AM, Marko Tiikkaja
> <marko.tiikkaja@cs.helsinki.fi> wrote:
>> Here's an updated patch. Only changes from the previous patch are
>> fixing the above issue and a regression test for it.
>
> - I'm not sure that canSetTag is the right name for the additional
> argument to ExecInsert/ExecUpdate/ExecDelete. OTOH, I'm not sure it's
> the wrong name either. But should we use something like
> isTopLevelQuery?
I'm going to have to take back my previous statement; this doesn't make
a lot of sense in the case of DO ALSO rules (or multiple statements in a
DO INSTEAD RULE). Those will have canSetTag=false, but they will be at
the top level.
> - It appears that we pull out all of the DML statements first and run
> them in order, but I'm not sure that's the right thing to do.
> Consider:
>
> WITH x AS (INSERT ...), y AS (SELECT ...), z AS (INSERT ...) SELECT ...
>
> I would assume we would do x, CCI, do y, do z, CCI, do main query, but
> I don't think that's what this implements. The user might be
> surprised to find out that y sees the effects of z.
I went ahead and implemented this, but there seems to be one small
problem: RECURSIVE. If there is a recursive query between those, it
might loop forever even if the top-level SELECT only wanted to see a few
rows from it. The docs already discourage writing recursive ctes like
that, but still this is a small caveat.
> - It seems like the gram.y changes for common_table_expr might benefit
> from some factoring; that is, create a production (or find a suitable
> existing one) for "statements of the sort that can appear within
> CTEs", and then use that in common_table_expr. Or maybe this doesn't
> work; I haven't tried it.
This seems to work. I used PreparableStmt, but I'm not sure how good
idea that really is. Maybe I should create a new one?
Regards,
Marko Tiikkaja
