Re: Avoiding bad prepared-statement plans.

Tom Lane wrote:

> Well, no, consider the situation where planning takes 50 ms, the generic
> plan costs 100ms to execute, but a parameter-specific plan would take 1ms
> to execute. Planning is very expensive compared to execution but it's
> still a win to do it.

I think that's a fun and worthwhile problem. But my limited personal
aim right now is a stopgap for pathological cases. I'd like to pick the
low-hanging fruit; actually squeezing the fat out of prepared statements
is a job I wouldn't get around to completing. Sorry for mixing metaphors.

Here's what I like about the really slow plans. (Now why does that sound
so strange?) We don't know if re-planning will help, but we do know
that (1) it won't hurt much relative to execution time, so we really
don't _care_; and (2) there is lots of potential for improvement, so
catching just one execution that can be much faster might pay for all
the extra time spent re-planning.

Where do we draw the line between costly and pathological? I still like
Bart's idea of a fixed ratio to planning time, because it reflects a
self-tuning sense of proportion. Sure, planning time can vary a lot but
we're talking about an order-of-magnitude difference, not an exact 19:21
optimum. We can be sloppy and still expect to win.

AFAIC a statement could go to "re-planning mode" if the shortest
execution time for the generic plan takes at least 10x longer than the
longest planning time. That gives us a decent shot at finding
statements where re-planning is a safe bet. A parameter that we or the
user would have to tweak would just be a fragile approximation of that.

> A possible scheme is to try it and keep track of whether we ever
> actually do get a better plan. If, after N attempts, none of the custom
> plans were ever more than X% cheaper than the generic one, then give up
> and stop attempting to produce custom plans. Tuning the variables might
> be challenging though.

A simple stopgap implementation may also be a useful experimentation
platform for refinements. It shouldn't be too complex to rip out when
something better comes along.

Jeroen