On 06/25/2015 11:04 AM, Tom Lane wrote:
> Josh Berkus <josh@agliodbs.com> writes:
>> On 06/25/2015 08:12 AM, Tom Lane wrote:
>>> I can't see doing a release just for this. If we were due for releases
>>> anyway, sure, but we've considerably overstressed our poor packagers of
>>> late. Previous discussion was to the effect that we'd anticipate another
>>> set of releases in a month or so, after some more multixact fixes have
>>> landed.
>
>> FWIW, I know users who will not update because of this regression.
>> Consider applications which regularly need to spin up 200 new
>> connections in 90 seconds due to usage peaks.
>
> TBH, if your app is critically dependent on backend startup time,
> You're Doing It Wrong, because that means you're pissing away significant
> performance by not using a connection pooler. So I don't have a huge
> amount of sympathy.
Most of these apps are using a connection pooler. But usage is very
spiky, and one doesn't want to keep around 300 idle connections all the
time for the 5 minutes when you need them. Indeed, pgbouncer is
configured to spin down idle connections after a configured amount of
time idle, as are most connection poolers.
Regardless, I don't think it's a good idea to release again sooner than
planned. It is a good reason to look at making connection creation time
a regular test, though; you can be sure I'll be testing 9.4.5 and 9.5a!
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com