Thread: Connection pooling - Number of connections

Hi, Brett Wooldridge here, one of the principals of HikariCP.  I thought I'd wade into the conversation pool a little myself if you guys don't mind. 

Speaking to David's point... 
>> Reaching the maxPoolSize from the minPoolSize means creating the 
>> connections at the crucial moment where the client application is in the 
>> desperate need of completing an important query/transaction which the 
>> primary responsibility since it cannot hold the data collected. 

This was one of the reasons I was proposing the fixed pool design.  In my experience, even in pools that maintain a minimum number of idle connections, responding to spike demands is problematic.  If you have a pool with say 30 max. connections, and a 10 minimum idle connection goal, a sudden spike demand for 20 connections means the pool can satisfy 10 instantly but then is left to [try to] establish 10 connections before the application's connectionTimeout (read acquisition timeout from the pool) is reached.  This in turn generates a spike demand on the database slowing down not only the connection establishments themselves but also slowing down the completion of transactions that might actually return connections to the pool. 

As I think Tom noted is a slidestack I read somewhere, there is a "knee" in the performance curve beyond which additional connections cause a drop in TPS.  While users often think it is a good idea to have maxPoolSize of 100, the reality is they can retire/reuse connections faster with a much smaller pool.  I didn't see a pool of a 2 or 3 dozen connections actually impacting performance much when half of them are idle and half are executing transactions (ie. the idle ones don't impact the overall performance much). 

Finally, one of my contentions was, either your database server has resources or it doesn't.  Either it has enough memory and processing power for N connections or it doesn't.  If the pool is set below, near, or at that capacity what is the purpose of releasing connections in that case?  Yes, it frees up memory, but that memory is not really available for other use given that at any instant the maximum capacity of the pool may be demanded. 
Instead releasing resources only to try to reallocate them during a demand peak seems counter-productive. 

I'd appreciate any shared thoughts on my presuppositions.

-Brett 

On Tue, Mar 25, 2014 at 5:24 AM, Gavin Flower <GavinFlower@archidevsys.co.nz> wrote:

Surely no code changes are required, as one can simply set the min and max pool sizes to be the same?

Cheers,
Gavin

To be sure it can be implemented that way, but its a question of design targets.  For example, if a pool is allowed to grow and shrink, the design might encompass a pool of threads that try to maintain the configured-minimum idle connections to respond to spike demands.  And there is the additional setting in the pool for users to [typically] misconfigure.

However, if the pool is fixed size, and attrition from the pool is only by idle timeout (typically measured in minutes), the design does not need to account for spike demand.  Likely connections that have dropped out can either be restored on-demand rather than something running constantly in the background trying to maintain and idle level.

One of the attributes of HikariCP is a minimalistic set of configuration options with sane defaults, and minimalistic code.  There are many competitor pools, offering dozens of settings ranging form banal to unsafe.  HikariCP initially even offered some of these options, but one-ny-one they're getting the ax.  More and more we're trying to look at what is the true core functionality that user's need -- eliminate what is unnecessary and easily misconfigured.

Thus a debate started over in our group, with some similar views as expressed here (on both sides).  Guillaume Smet was concerned about the impact of idle connections on active ones in PostgreSQL (in a fixed pool scenario) and wanted to ask some of the experts over here.

On 03/24/2014 06:27 AM, Brett Wooldridge wrote:
> This was one of the reasons I was proposing the fixed pool design.  In my
> experience, even in pools that maintain a minimum number of idle
> connections, responding to spike demands is problematic.  If you have a
> pool with say 30 max. connections, and a 10 minimum idle connection goal, a
> sudden spike demand for 20 connections means the pool can satisfy 10
> instantly but then is left to [try to] establish 10 connections before the
> application's connectionTimeout (read acquisition timeout from the pool) is
> reached.  This in turn generates a spike demand on the database slowing
> down not only the connection establishments themselves but also slowing
> down the completion of transactions that might actually return connections
> to the pool.

So what makes sense really depends on what your actual connection
pattern is.  Idle connections aren't free; aside from PostgreSQL
lock-checking overhead, they hold on to any virtual memory allocated to
them when they were working.  In the aggregate, this can add up to quite
a bit of memory, which can then cause the OS to decide not to cache some
data you could really use.

Now, if your peak is 100 connections and your median is 50, this doesn't
signify.  But I know more than a few workloads where the peak is 1000
and the median is 25, and in that case you want to drop the idle
connections gradually.  The key is to keep enough of a buffer of ready
connections to deal with the next peak when it comes in.

That also means that even if the pool is a fixed size, you want to
rotate in and out the actual sessions, so that they don't hang onto
maximum virtual memory indefinitely.

--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com

Sent from my iPhone
> On Mar 27, 2014, at 9:35 AM, Josh Berkus <josh@agliodbs.com> wrote:
>
>> On 03/24/2014 06:27 AM, Brett Wooldridge wrote:
>> This was one of the reasons I was proposing the fixed pool design.  In my
>> experience, even in pools that maintain a minimum number of idle
>> connections, responding to spike demands is problematic.  If you have a
>> pool with say 30 max. connections, and a 10 minimum idle connection goal, a
>> sudden spike demand for 20 connections means the pool can satisfy 10
>> instantly but then is left to [try to] establish 10 connections before the
>> application's connectionTimeout (read acquisition timeout from the pool) is
>> reached.  This in turn generates a spike demand on the database slowing
>> down not only the connection establishments themselves but also slowing
>> down the completion of transactions that might actually return connections
>> to the pool.
>
> Now, if your peak is 100 connections and your median is 50, this doesn't
> signify.  But I know more than a few workloads where the peak is 1000
> and the median is 25, and in that case you want to drop the idle
> connections gradually.

In the end we've gone with a maxPoolSize + minIdle model where the default is that they are equal (fixed pool).

Though I won't dispute that such workloads (1000 active connections) exist, in that someone created it, I would love to
heartheir justification. Unless they have >128 CPU cores and solid state storage they are basically spinning their
wheels.

> That also means that even if the pool is a fixed size, you want to
> rotate in and out the actual sessions, so that they don't hang onto
> maximum virtual memory indefinitely.

We do this, there is a maxLifeTime setting to rotate out connections.

-Brett