Re: GUC parameter cursors_tuple_fraction - Mailing list pgsql-patches

On Thu, 2008-04-03 at 16:45 +0200, Hell, Robert wrote:
> This patch adds a GUC parameter for tuple_fraction of cursors (discussed
> earlier here:
> http://archives.postgresql.org/pgsql-performance/2008-04/msg00018.php).
> By setting this parameter the planner's favor to use fast-start plans
> for cursors can be affected.

I think this patch looks OK coding-wise, but not tested, yet.

If we did apply this patch it would need significantly more
documentation, probably examples and the like. But I think writing that
docs would open up a can of worms, hence copying -hackers.

Other RDBMS allow users to specify whether they want fast-start or
all-data plans. We should discuss whether we want to set the fraction
directly or whether we should have 2 (or more) specific settings such as
"fast" and "all".

Also, if we did have this parameter then I don't think it should be
included in postgresql.conf. I don't see any need to change the default
setting for *all* cursors, but I can see the need to change the cursor
fraction for *one* specific query. Which raises wider issues.

* We could add to DECLARE syntax that says something like OPTIMIZE FOR
FIRST ROWS or OPTIMIZE FOR ALL ROWS. But our policy AIUI is that we do
not want to further decorate SQL Standard commands.

* We've said here http://www.postgresql.org/docs/faqs.TODO.html that we
"Don't want hints". If that's what we really think, then this patch must
surely be rejected because its a hint... That isn't my view. I *now*
think we do need hints of various kinds.

Decorating queries with *all* necessary information is not always good,
but there are some kinds of information that *do* belong on specific
queries. The cursor fraction is a great example of information that
really does live on a specific query.

But in a wider sense, I think support of hints is actually the only way
long term of making large applications work within a reasonable
timeframe and cost. If we change information at the database object
level in order to correct one issue, we are likely to find that more
problems are raised elsewhere. Same thing is true of altering optimiser
cost models. Few users can wait 2 years while we solve the problem and
fix it permanently, or even a few days while they resolve the inner
workings of the planner and work out how to re-write it.

I had spoken strongly against hints for general use in Postgres
previously. Many attendees on recent PostgreSQL performance courses have
successfully argued in favour of hints and as a result my viewpoint is
now changed. Though we need a central no-hints-allowed GUC for those
cases where application programmers need restraining...

I think we need Hints. (And not this patch, sorry about that, Robert).

(This could well lead to me losing work doing performance tuning, though
I believe its the wish of the majority that we should support hints).

--
  Simon Riggs
  2ndQuadrant  http://www.2ndQuadrant.com