Using Connection Pooling (pgBouncer)#
This document captures one set of experiments with pgBouncer 1.14 and RelStorage 3.1. pgBouncer is not a regularly tested configuration.
It’s written in the context of a gunicorn web application serving a high-level of traffic using RelStorage and PostgreSQL.
In the default ‘session’ pooling mode, RelStorage works with pgBouncer as-is. In session pooling mode, there’s a one-to-one relationship between each RelStorage connection and a PostgreSQL connection. Idle ZODB and RelStorage connections still occupy a PG connection. On the surface that doesn’t seem to buy very much, but it might be helpful at the extreme end of the scale by allowing us to keep a somewhat larger gunicorn accept value: If we reduce the ZODB connection pools beneath what PG allows, then when new temporary connections are needed, they could come out of the pgBouncer pool, do their thing, then go back to the pool. That would speed up (a very little bit) using temporary connections, and it would prevent us from ever seeing the ‘no more connections’ error, because we would configure pgBouncer to queue incoming connections until a PG backend is available.
I’m not sure it’s worth it though, especially because there is also a small speed hit to using pgBouncer, as expected. Some of the RelStorage tests are timing tests with strict tolerances. For example, we set a lock timeout of 0.1s, fail to get a conflicted lock, and assert that the time taken to fail was no longer than 0.1s. With pgBouncer in the way, that test (sometimes) fails by taking anywhere from 0.11 to 0.14s.
The pooling mode that multiplexes connections from RelStorage to PG (and hence could allow for fewer overall PG connections), ‘transaction,’ DOES NOT WORK.
RelStorage pretty deeply embeds the idea that it owns the connection state and sets up prepared statements, temporary tables and triggers and assumes they persist for the duration of the connection. I was able to hack around some of the resulting issues, but eventually I hit a wall where some tests failed with transaction state errors whose cause was not immediately apparent. At least once it also crashed the Python process.
Future Implementation Ideas#
I could imagine making RS play nicely with pgBouncer, or even pool and share connections internally (which would benefit all databases). But either way would take some work.
I could also imagine adding a setting to RS to make it close its connections at the end of every transaction. That would practically require pgBouncer to help reduce the overhead of opening new connections for every transaction. But it should let pgBouncer work in either ‘session’ or ‘transaction’ mode, and either mode would result in effective multiplexing (because there would be no open idle connections). That would also take some work, but probably somewhat less than sharing sessions would (because of everything already in place to gracefully deal with closed connections).