The simple, maybe unhelpful answer is that fail2ban needs to have two things at once: the logs, and a way to block the network traffic.
Where exactly you want those things to coincide is really up to you, there might only be one point that simultaneously has access to both those things, or there might be multiple points depending on how your systems and services and network is configured, or if you’re in a bad situation you might find you don’t really have any single point where both those things are simultaneously possible, in which case you’ll need to reconfigure something until you do have at least one point where both those things are again coincident.
As far as best practices, I can’t really say for sure, but I know that one of the more convenient ways to run it is usually on the same system, I usually run it outside of docker, on the host, which can pretty easily get access to the container’s logs if necessary, and let fail2ban block traffic on the whole system. For me, any system running any publicly accessible network services that allow password login gets a fail2ban instance.
A whole-network approach where you block the traffic on the firewall is fine too, if that’s what you prefer and what you want to work towards, but it’s probably going to be significantly more complex to set up because now you need to either figure out how to get fail2ban to be able to access your firewall or a way for your firewall to get the logs it needs.