Skip site navigation (1) Skip section navigation (2)

Re: Scanning for insert

From: James William Pye <pgsql(at)jwp(dot)name>
To: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Cc: Hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: Scanning for insert
Date: 2006-02-28 01:48:21
Message-ID: (view raw, whole thread or download thread mbox)
Lists: pgsql-hackers
On Mon, Feb 27, 2006 at 05:44:20PM -0500, Tom Lane wrote:
> Is this really a good idea?  The fundamental problem I see with it is
> that it takes away the index AM's ability to make any guarantees about
> its locking behavior, ie, how long locks will be held or what other
> operations might intervene while holding them.

Yeah, I saw that as well. :(

My only thoughts on that issue so far have been that the user must tread
carefully while holding these scans; it becomes the user's responsibility.
Weak at best, perhaps, so I can understand if that does not move you or
anyone else into thinking positively of these proposed interfaces. ;)

> It'll also require the
> AM to save potentially large amounts of state between the two calls
> (eg, an entire search path might be needed for GiST).  Currently any
> such state can live on the stack in local variables, but that won't
> work if it has to be remembered until a later AM call.

Hrm, certainly, implementing these methods for AMs that use such state
keeping techniques may be extra difficult. Massive changes may be necessary.
However, I don't imagine that making such a change would be impossible, nor
would it necessarily be required at all. Similar to uniqueness, the index need
not be forced to implement these new interfaces--surely not upon this patch's
introduction into the source(if/when). If a feature were to demand the use of scan
insertions on a specific index that does not provide the interfaces, it could
easily raise an informative exception about the shortcoming.

Hehe, perhaps a Truly Foolish Notion, but could we use siglongjmp for such
cases(state living on the stack)? Seems questionable and evil, tho; I know. ;)

> Lastly, what
> happens if the caller loses control (due to elog) and never makes the
> followup AM call?

I imagine/hope it would get cleaned up similar to how nbtree gets cleaned up at
the end of the transaction that was rolled back due to a unique constraint
violation. [I seem to recall having to do some special _bt_wrtbuf to get my
experimental insert unless patch working, so I guess that at eox some cleanup is
done w.r.t. to those locks and whatnot. (yeah, technical term, "whatnot" ;)]

Additionally, if the caller is very concerned with potential exceptions,
perhaps a PG_TRY() block should be exercised in those areas of worry.

I guess in most cases it simply comes back to becoming the scan's user's
responsibility to be sure to keep things kosher. :\

> If that's what you want, maybe a better answer is to simply allow
> aminsert to return a uniqueness-conflict indication, rather than
> raising the ereport for itself.

Maybe so. However, I guess I was thinking along lines that many times multiple
insert scans may need to be done before the final decision to actually do the
insertion is made(yeah, uh, the insert unless thing is what I've had in mind ;).

Also, I think part of this point is to be able to avoid the actual insertion
into heap, so as to avoid a superfluous heap_insert & heap_delete, and other
"unwinding code" if a uniqueness-conflict indication were made to a user that
needs to respond to such signals for the feature being implemented.
(The more unique constraint violations that occur, the more I/O that gets saved
with insert scans. I imagine this could be a very good thing for certain
Regards, James William Pye

In response to

pgsql-hackers by date

Next:From: Bruce MomjianDate: 2006-02-28 02:09:56
Subject: Re: pg_config, pg_service.conf, postgresql.conf ....
Previous:From: Tatsuo IshiiDate: 2006-02-28 01:14:33
Subject: Re: possible design bug with PQescapeString()

Privacy Policy | About PostgreSQL
Copyright © 1996-2017 The PostgreSQL Global Development Group