Peter Geoghegan <pg@heroku.com> writes:
> On Tue, Dec 10, 2013 at 3:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I'm wondering whether this doesn't indicate that we need to rethink where
>> the fingerprinter has been plugged in. I'm not sure that somewhere in
>> the planner, post-constant-folding, would be a better place; but it's
>> worth thinking about.
> ... if you're not talking about blaming plans and not queries, you
> must be talking about making the planner do the constant folding in a
> way that facilitates tools like pg_stat_statements in getting the
> "essential nature" of the query (*not* the plan) post constant
> folding.
Yeah; if we were going to take this seriously, we'd need to do some
refactoring to separate normalization-type activities from other
planning activities. I'm not sure it's worth the trouble. Right
now, for instance, eval_const_expressions() also handles inlining
of SQL functions, which is a behavior we'd almost certainly *not*
want in front of query fingerprinting. But it's hard to see how
we disentangle that without either lost optimization capacity
or duplicative processing. A significant fraction of the point of
const-folding is to exploit opportunities revealed by inlining.
Anyway, I'm not volunteering to do this ;-) ... just idly speculating
about whether it'd be worth doing.
regards, tom lane
