Re: Built-in connection pooler

> I attached new version of the patch with fixed indentation problems and
> Win32 specific fixes.

Great, this latest patch applies cleanly to master. installcheck world
still passes.

> "connections_proxies" is used mostly to toggle connection pooling.
> Using more than 1 proxy is be needed only for huge workloads (hundreds
> connections).

My testing showed using only one proxy performing very poorly vs not using
the pooler, even at 300 connections, with -3% TPS. At lower numbers of
connections it was much worse than other configurations I tried. I just
shared my full pgbench results [1], the "No Pool" and "# Proxies 2" data is
what I used to generate the charts I previously shared. The 1 proxy and 10
proxy data I had referred to but hadn't shared the results, sorry about
that.

> And "session_pool_size" is core parameter which determine efficiency of
> pooling.
> The main trouble with it now, is that it is per database/user
> combination. Each such combination will have its own connection pool.
> Choosing optimal value of pooler backends is non-trivial task. It
> certainly depends on number of available CPU cores.
> But if backends and mostly disk-bounded, then optimal number of pooler
> worker can be large than number of cores.

I will do more testing around this variable next. It seems that increasing
session_pool_size for connection_proxies = 1 might help and leaving it at
its default was my problem.

> PgPRO EE version of connection pooler has "idle_pool_worker_timeout"
> parameter which allows to terminate idle workers.

+1

> It is possible to implement it also for vanilla version of pooler. But
> primary intention of this patch was to minimize changes in Postgres core

Understood.

I attached a patch to apply after your latest patch [2] with my suggested
changes to the docs. I tried to make things read smoother without altering
your meaning. I don't think the connection pooler chapter fits in The SQL
Language section, it seems more like Server Admin functionality so I moved
it to follow the chapter on HA, load balancing and replication. That made
more sense to me looking at the overall ToC of the docs.

Thanks,

[1]
https://docs.google.com/spreadsheets/d/11XFoR26eiPQETUIlLGY5idG3fzJKEhuAjuKp6RVECOU
[2]
https://www.postgresql.org/message-id/attachment/102848/builtin_connection_proxy-11.patch

*Ryan*

>