On Wed, Jan 13, 2016 at 10:47:18AM -0500, Tom Lane wrote:
> Vladimir Sitnikov <sitnikov.vladimir@gmail.com> writes:
> > Note: I state that mixing "kinds" of bind values is a bad application
> > design anyway. In other words, application developer should understand
> > if a query is DWH-like (requires replans) or OLTP-like (does not
> > require replans). Agreed?
>
> No, not agreed. As was already pointed out upthread, such information
> is not available in many use-cases for the plancache.
>
> The real problem here IMO is inaccurate plan cost estimates, and that's
> not something that there is any easy fix for.
>
> However ... one specific aspect of that is that to some extent, the cost
> estimate made for the generic plan is incommensurate with the estimates
> for the custom plans because the latter are made with more information.
> I don't remember the details of your specific case anymore, but we've
> seen cases where the generic plan is falsely estimated to be cheaper
> than custom plans because of this.
I never understood why we don't just keep the selectivity estimates of
previous plans and just reuse the plan if the selectivity estimates are
similar. Isn't parameter selectivity the only thing that distinguishes
on plan with parameter from another?
Checking selectivity estimates must be cheaper than replanning. This
could be done at the second use of the prepared plan, and maybe for all
plan reuses, rather than waiting for five and then perhaps getting this
bad behavior.
-- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB
http://enterprisedb.com
+ As you are, so once was I. As I am, so you will be. +
+ Roman grave inscription +
