Psycopg2 pool clarification

[Israel Brewster]

I've been using the psycopg2 pool class for a while now, using code similar to the following:

>>> pool=ThreadedConnectionPool(0,5,<connection_args>)
>>> conn1=pool.getconn()
>>> <do whatever with conn1>
>>> pool.putconn(conn1)
.... repeat later, or perhaps "simultaneously" in a different thread.

and my understanding was that the pool logic was something like the following:

- create a "pool" of connections, with an initial number of connections equal to the "minconn" argument
- When getconn is called, see if there is an available connection. If so, return it. If not, open a new connection and return that (up to "maxconn" total connections)
- When putconn is called, return the connection to the pool for re-use, but do *not* close it (unless the close argument is specified as True, documentation says default is False)
- On the next request to getconn, this connection is now available and so no new connection will be made
- perhaps (or perhaps not), after some time, unused connections would be closed and purged from the pool to prevent large numbers of only used once connections from laying around.

However, in some testing I just did, this doesn't appear to be the case, at least based on the postgresql logs. Running the following code:

>>> pool=ThreadedConnectionPool(0,5,<connection_args>)
>>> conn1=pool.getconn()
>>> conn2=pool.getconn()
>>> pool.putconn(conn1)
>>> pool.putconn(conn2)
>>> conn3=pool.getconn()
>>> pool.putconn(conn3)

produced the following output in the postgresql log:

2017-06-02 14:30:26 AKDT LOG: connection received: host=::1 port=64786
2017-06-02 14:30:26 AKDT LOG: connection authorized: user=logger database=flightlogs
2017-06-02 14:30:35 AKDT LOG: connection received: host=::1 port=64788
2017-06-02 14:30:35 AKDT LOG: connection authorized: user=logger database=flightlogs
2017-06-02 14:30:46 AKDT LOG: disconnection: session time: 0:00:19.293 user=logger database=flightlogs host=::1 port=64786
2017-06-02 14:30:53 AKDT LOG: disconnection: session time: 0:00:17.822 user=logger database=flightlogs host=::1 port=64788
2017-06-02 14:31:15 AKDT LOG: connection received: host=::1 port=64790
2017-06-02 14:31:15 AKDT LOG: connection authorized: user=logger database=flightlogs
2017-06-02 14:31:20 AKDT LOG: disconnection: session time: 0:00:05.078 user=logger database=flightlogs host=::1 port=64790

Since I set the maxconn parameter to 5, and only used 3 connections, I wasn't expecting to see any disconnects - and yet as soon as I do putconn, I *do* see a disconnection. Additionally, I would have thought that when I pulled connection 3, there would have been two connections available, and so it wouldn't have needed to connect again, yet it did. Even if I explicitly say close=False in the putconn call, it still closes the connection and has to open

What am I missing? From this testing, it looks like I get no benefit at all from having the connection pool, unless you consider an upper limit to the number of simultaneous connections a benefit? :-) Maybe a little code savings from not having to manually call connect and close after each connection, but that's easily gained by simply writing a context manager. I could get *some* limited benefit by raising the minconn value, but then I risk having connections that are *never* used, yet still taking resources on the DB server.

Ideally, it would open as many connections as are needed, and then leave them open for future requests, perhaps with an "idle" timeout. Is there any way to achieve this behavior?

-----------------------------------------------
Israel Brewster
Systems Analyst II
Ravn Alaska
5245 Airport Industrial Rd
Fairbanks, AK 99709
(907) 450-7293
-----------------------------------------------

[israel]

Since I've gotten no replies to this, I was wondering if someone could
at least confirm which behavior (my expected or my observed) is
*supposed* to be the correct? Should a psycopg2 pool keep connections
open when returned to the pool (if closed is False), or should it close
them as long as there is more than minconn open? i.e is my observed
behavior a bug or a feature?

[dieter]

israel <isr...@ravnalaska.net> writes:
> Since I've gotten no replies to this, I was wondering if someone could
> at least confirm which behavior (my expected or my observed) is
> *supposed* to be the correct? Should a psycopg2 pool keep connections
> open when returned to the pool (if closed is False), or should it
> close them as long as there is more than minconn open? i.e is my
> observed behavior a bug or a feature?

You should ask the author[s] of "psycopg2" about the supposed behavior.


>From my point of view, everything depends on the meaning of the "min"
and "max" parameters for the pool.

You seem to interprete "max" as "keep as many connections as this open".
But it can also be a hard limit in the form "never open more than this
number of connections". In the latter case, "min" may mean "keep this
many connections open at all time".

[israel]

You are right about my interpretation of "max", and also about the
actual meaning. Thus the reason I was asking :-). I did post on the bug
report forum, and was informed that the observed behavior was the
correct behavior. As such, using psycopg2's pool is essentially
worthless for me (plenty of use for it, i'm sure, just not for me/my use
case).

So let me ask a different, but related, question: Is there a Python
library available that gives me the behavior I described in my first
post, where connections are "cached" for future use for a time? Or
should I just write my own? I didn't find anything with some quick
googling, other than middleware servers like pgpool which, while they
have the behavior I want (at least from my reading), will still require
the overhead of making a connection (perhaps less than direct to
postgres? Any performance comparisons out there?), not to mention
keeping yet another service configured/running. I would prefer to keep
the pool internal to my application, if possible, and simply reuse
existing connections rather than making new ones. Thanks!

[dieter]

israel <isr...@ravnalaska.net> writes:
> On 2017-06-06 22:53, dieter wrote:

> ...

> As such, using psycopg2's pool is essentially
> worthless for me (plenty of use for it, i'm sure, just not for me/my
> use case).

Could you not simply adjust the value for the "min" parameter?
If you want at least "n" open connections, then set "min" to "n".

[Israel Brewster]

-----------------------------------------------
Israel Brewster
Systems Analyst II
Ravn Alaska
5245 Airport Industrial Rd
Fairbanks, AK 99709
(907) 450-7293
-----------------------------------------------

Well, sure, if I didn't care about wasting resources (which, I guess many people don't). I could set "n" to some magic number that would always give "enough" connections, such that my application never has to open additional connections, then adjust that number every few months as usage changes. In fact, now that I know how the logic of the pool works, that's exactly what I'm doing until I am confident that my caching replacement is solid.

Of course, in order to avoid having to open/close a bunch of connections during the times when it is most critical - that is, when the server is under heavy load - I have to set that number arbitrarily high. Furthermore, that means that much of the time many, if not most, of those connections would be idle. Each connection uses a certain amount of RAM on the server, not to mention using up limited connection slots, so now I've got to think about if my server is sized properly to be able to handle that load not just occasionally, but constantly - when reducing server load by reducing the frequency of connections being opened/closed was the goal in the first place. So all I've done is trade dynamic load for static load - increasing performance at the cost of resources, rather than more intelligently using the available resources. All-in-all, not the best solution, though it does work. Maybe if load was fairly constant it would make more sense though. So like I said *my* use case, which is a number of web apps with varying loads, loads that also vary from day-to-day and hour-to-hour.

On the other hand, a pool that caches connections using the logic I laid out in my original post would avoid the issue. Under heavy load, it could open additional connections as needed - a performance penalty for the first few users over the min threshold, but only the first few, rather than all the users over a certain threshold ("n"). Those connections would then remain available for the duration of the load, so it doesn't need to open/close numerous connections. Then, during periods of lighter load, the unused connections can drop off, freeing up server resources for other uses. A well-written pool could even do something like see that the available connection pool is running low, and open a few more connections in the background, thus completely avoiding the connection overhead on requests while never having more than a few "extra" connections at any given time. Even if you left of the expiration logic, it would still be an improvement, because while unused connections wouldn't drop, the "n" open connections could scale up dynamically until you have "enough" connections, without having to figure out and hard-code that "magic number" of open connections.

Why wouldn't I want something like that? It's not like its hard to code - took me about an hour and a half to get to a working prototype yesterday. Still need to write tests and add some polish, but it works. Perhaps, though, the common thought is just "throw more hardware at it and keep a lot of connections open at all time?" Maybe I was raised to conservatively, or the company I work for is too poor.... :-D

>
> --
> https://mail.python.org/mailman/listinfo/python-list

[Ian Kelly]

Psycopg is first and foremost a database adapter. To quote from the
psycopg2.pool module documentation, "This module offers a few pure
Python classes implementing *simple* connection pooling directly in
the client application" (emphasis added). The advertised list of
features at http://initd.org/psycopg/features/ doesn't even mention
connection pooling. In short, you're getting what you paid for.

It sounds like your needs are beyond what the psycopg2.pool module
provides. I suggest looking into a dedicated connection pooler like
PgBouncer. You'll find that it's much more feature-rich and
configurable than psycopg2.pool. It's production-ready, unlike your
prototype. And since it's a proxy, it can take connections from
multiple client apps and tune the pool to your overall load rather
than on an app-by-app basis (and thus risk overloading the backend if
multiple apps unexpectedly peak together).

As for why psycopg2.pool is the way it is, maybe most users don't have
your situation of serving multiple apps with loads varying on
different cycles. Most are probably only serving a single app, or if
serving multiple apps then they likely have common user bases with
similar peak times. You can't dynamically adjust the amount of RAM in
your server, so saving resources like RAM at below-peak times only
matters if you're going to do something else with it. In the scenarios
I described there isn't much else to do with it, so I can understand
if saving RAM isn't a priority.

[israel]

Quite possible. Thus the reason I was looking for clarification on how
the module was intended to work - if it doesn't work in the way that I
want it to, I need to look elsewhere for a solution. My main reason for
posting this thread was that I was expecting it to work one way, but
testing showed it working another way, so I was trying to find out if
that was intentional or user error. Apparently it's intentional, so
there we go - in it's current form at least, my needs are beyond what
the psycopg2 pool provides. Fair enough.

> I suggest looking into a dedicated connection pooler like
> PgBouncer. You'll find that it's much more feature-rich and
> configurable than psycopg2.pool. It's production-ready, unlike your
> prototype. And since it's a proxy, it can take connections from
> multiple client apps and tune the pool to your overall load rather
> than on an app-by-app basis (and thus risk overloading the backend if
> multiple apps unexpectedly peak together).

Very true, and I've looked into that (as well as the related, but more
basic, PgPool product), but it seems to me that any external proxy
product like these would defeat *my* purpose for using a pool in the
first place: avoiding the overhead of making/breaking many connections
quickly. That is, all you have really done is gone from connecting to
Postgres to connecting to PgBouncer. You are still making and breaking
just as many connections. Unless connecting to PgBouncer is
significantly cheaper than connecting to Postgres? This may well be the
case, but I haven't yet seen anything to support that. Haven't seen
anything to refute that either, however :)

Of course, there may be many other features provided by such tools that
would make them worthwhile, even for my use case. However, my primary
goal in using a pool was avoiding the connection overhead with each
request, so if a tool doesn't do that, then it isn't the right tool for
me :)

>
> As for why psycopg2.pool is the way it is, maybe most users don't have
> your situation of serving multiple apps with loads varying on
> different cycles. Most are probably only serving a single app, or if
> serving multiple apps then they likely have common user bases with
> similar peak times. You can't dynamically adjust the amount of RAM in
> your server, so saving resources like RAM at below-peak times only
> matters if you're going to do something else with it. In the scenarios
> I described there isn't much else to do with it, so I can understand
> if saving RAM isn't a priority.

True, but you would still have to deal with the minconn "magic number",
unless you just adjusted it so high from the start that even if your
load/use grows over time you never have to mess with it. Even in the
single app use case, where you don't care about RAM usage (since there
is nothing else trying to use the RAM) in order to get maximum benefit
from a pool you'd have to keep an eye on your usage and make sure it
never (or rarely) exceeds whatever arbitrary value you have set for
minconn. Not a big deal, especially if you tune it high to begin with,
but it is one more thing.

Honestly, some of that is just personal issues. I have problems with
code that is inefficient by design (even if there is nothing to be
gained from efficiency), or that makes assumptions about things when it
could be dynamic. I have often coded in such a way that a given value
can be adjusted dynamically, even when the people giving me the specs
say it will never change. More than once that has enabled me to respond
to a "feature request" or change order by saying "it already does that".
On the other hand, I am probably a poster child for "premature
optimization", and often have to stop myself from optimizing code just
because I can, when in reality it is not worth my time. By the same
token, the idea of wasting resources - even when, as you state, there is
nothing else to do with them - just rubs me the wrong way. As such, I
readily acknowledge that some of my requests/statements stem from my own
personal desires, and not from any actual lack/need in the product.