Armin from Sentry here. We talked a bit about where we could scale with postgres and part of our problem is that we're quite write and also delete heavy which means our vacuum takes very long. A potential option would be to partition some of our time based tables and make a rolling window view over it. Then we could throw entire tables away after the hold period and our vacuums should be faster at least. We could also move some tables to another database which might help, there however we then need to deal with higher connection counts unfortunately.
Downside for us is that we also want to keep the setup straightforward for our on-premise customers so we need to be careful with not introducing unnecessary complexity.
Is this still a problem after you tuned vacuum to be actually aggressive? It sure isn't perfect but you can go a long way with with just adjusting a couple settings.
I need to check how long vacuum takes at the moment, but I think it's in the high hours.
There is not a database in existence that allows you to be oblivious to the underlying organization while still giving good write/delete performance. PostgreSQL is no different in that regard.
This is quite the statement and inconsistent with your later point. Postgres never updates data in-place, and certain workloads can never be as fast on Postgres as using a different storage engine such as MySQL/TokuDB.
MyISAM didn't have such problem, but it also couldn't handle high update concurrency.
We're running a setup that employs postgres for a write-heavy (but far from write-only) application, and are experiencing similarly troubling vacuum jobs. We're considering moving back to good old MySQL because of it. Anyone who's been here before and has some insight?