There's pros/cons to in-app pooling vs. middleware pooling, but calling in-app pooling "worse" than no pooling at all is kind of misleading.
If I were using PGBouncer, I'd probably leave SQLA's pooling on anyway, so at least you have no TCP latency at all on connect. unless you're dealing with apps that go unevenly idle/active.
You don't need to get all worked up. Nowhere I affirm it's worse than no polling.
I'm talking about not reinventing the wheel, specially since pooling inside the worker is inferior than just setting up pgpool.
As well as, in-app pooling is the easiest way to limit the max simultaneous connection checkouts per application. If PGBouncer were to handle this, I'd assume you'd need to either send distinct usernames or some other instance-specific token over so that it could identify sources of connections.
> You'll have a different pool per worker process, and they won't persist when you restart your workers - right when you need low latency the most