Re: More stable query plans via more predictable column statistics

On Mon, Mar 7, 2016 at 3:17 AM, Shulgin, Oleksandr
<oleksandr.shulgin@zalando.de> wrote:
> On Fri, Mar 4, 2016 at 7:27 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>>
>> On Thu, Mar 3, 2016 at 2:48 AM, Shulgin, Oleksandr
>> <oleksandr.shulgin@zalando.de> wrote:
>> > On Wed, Mar 2, 2016 at 7:33 PM, Alvaro Herrera
>> > <alvherre@2ndquadrant.com>
>> > wrote:
>> >> Shulgin, Oleksandr wrote:
>> >>
>> >> > Alright.  I'm attaching the latest version of this patch split in two
>> >> > parts: the first one is NULLs-related bugfix and the second is the
>> >> > "improvement" part, which applies on top of the first one.
>> >>
>> >> So is this null-related bugfix supposed to be backpatched?  (I assume
>> >> it's not because it's very likely to change existing plans).
>> >
>> > For the good, because cardinality estimations will be more accurate in
>> > these
>> > cases, so yes I would expect it to be back-patchable.
>>
>> -1.  I think the cost of changing existing query plans in back
>> branches is too high.  The people who get a better plan never thank
>> us, but the people who (by bad luck) get a worse plan always complain.
>
>
> They might get that different plan when they upgrade to the latest major
> version anyway.  Is it set somewhere that minor version upgrades should
> never affect the planner?  I doubt so.

People with meticulous standards are expected to re-validate their
application, including plans and performance, before doing major
version updates into production. They can continue to use a *fully
patched* server from a previous major release while they do that.

This is not the case for minor version updates.  We do not want to put
people in the position where getting a security or corruption-risk
update forces them to also accept changes which may destroy the
performance of their system.

I don't know if it is set out somewhere else, but there are many
examples in this list of us declining to back-patch performance bug
fixes which might negatively impact some users.  The only times we
have done it that I can think of are when there is almost no
conceivable way it could have a meaningful negative effect, or if the
bug was tied in with security or stability bugs that needed to be
fixed anyway and couldn't be separated.

Cheers,

Jeff