Re: [psycopg] Caching connection pool?

On Thu, Jun 8, 2017 at 7:28 PM, Israel Brewster <israel@ravnalaska.net> wrote:

To resolve this, I went ahead and wrote up a AbstractConnectionPool subclass
that implements the caching logic I laid out above
(https://github.com/psycopg/psycopg2/pull/565).

Thank you for your work, Israel

For the list: I've not been a heavy pool user for a long time, and the
previous logic always felt a bit weak (destroy immediately a
connection if > minconn?) So I welcome improvement on this object. I'd
be happy to know from other people what they expect as a reasonable
default behaviour for the connection pool.

At the moment I would
consider it to be a fairly basic implementation - pruning of expired
connections occurs only on a call to putconn (as opposed to running
periodically), and connections time out at a (user-specified) period of time
from when they were created, rather than last used, for example.

For me it is desirable that the pool is "passive": a periodic run
would require a separate thread, which then someone would like to be
an eventlet green thread and some other a gevent green thread... In
any case it wouldn't behave in a deterministic way.

Probably using the date of last usage would be better than the date of
creation, otherwise you would just delete all the connections every 10
minutes, or something like that.

It's
arguable that both those elements (and perhaps others) could be improved on,
but in my initial testing it appears to work. What do you guys think? Is
this worth my time, or am I re-creating the wheel or engaging in premature
optimization or something? Is the logic as I laid it out above good, or
would there be some improvements that could be suggested for it? One thing I
was thinking that could be helpful would be that if it saw the pool of
available connections getting low, it could open one or more additional
connections "in the background" so they are ready when needed. But maybe
that's going a bit overboard?

I think the pool provided by psycopg should still be a "mechanism"
rather than a "policy". It would be great to improve on what we
currently have, but not the pool too "intelligent": a super-smart pool
would be probably tied and tweaked to the application using it, which
may decide to use e.g. background threads to maintain the pool at a
certain level, may periodically check the state of the connections,
etc. But the pool is a weird object, it's a bit on the fence between
being mechanism and being policy, so I welcome a discussion about what
should be reasonable to expect from this object out of the psycopg
box, what may instead belong to an extension module from PyPI, what is
too closely tied to an appication/environment to be generically
useful.

Cheers,

-- Daniele