We use Postgres for a lot of stuff here. Since it's a multi-process database server (unlike MySQL which is multi-threaded), creating a direct connection to Postgres is relatively slow.

This can be worked around by using a connection pool server in front of the DB. Up until recently we used pgpool for this. But it has a few disadvantages:

* It's multi-process itself (although it preforks), so it eats quite a lot of RAM.
* Once you reach the maximum number of processes, it "queues" incoming connections in the OS accept queue, which is ugly.
* It has poor (but improving) instrumentation.

Recently we switched to using PgBouncer from our friends at Skype. It fixes all these problems quite nicely:

* It uses single-process event-based socket handling, using the excellent libevent like memcached. This means it uses several orders of magnitude less RAM than pgpool.
* It has configurable internal queuing, so your clients' connections don't time out if they're in the queue.
* It has good instrumentation in the form of an internal pseudo-database you can query.

Now here are some pretty graphs. Spot where we changed: