JPA implementations have "managed entities", sometimes called session or 1st level cache which is making sure that every entity is loaded at max. one time within a transaction. Like e.g. checking user/user permissions is something which typically has to be done in several places in course of a single request - you don't want to keep loading them for every check, you don't want to keep passing them across 20 layers, so some form of caching is needed. JPA implementations do it for you automatically (assuming you're fine with transaction-scoped cache) since this is such a core concept to how JPA works (the fact it's also a cache is kind of secondary consequence). JPA implementations typically provide more advanced caching capabilities, caching query results, distributed cache (with proper invalidation) etc.
Why is caching not a feature in DB connection pools? I mean, most databases have it on their side, why not have it as an option for the same query sets prior to hitting the db, with configurable invalidations? Or is it, and I've just never thought to look for it.
Integrating cache into connection pools brings little added value since connection pools don't have enough context/information to manage the cache intelligently. You'd have to do all the hard work (like invalidation) yourself anyway.
Example: if you execute "UPDATE orders SET x = 5 WHERE id = 10", the connection pool has no idea what entries to invalidate. ORM knows that since it tracks all managed entities, understands their structure, identity.
I guess I was thinking more of frequently run queries against infrequently modified data or where stale data doesn't matter so much. The sort of things that are ideal cache targets. You'd think you could tag queries like that. Sort of the things that a CDN caches but more granular. Sure if it's stuff that's frequently changed, an ORM could reason about it just like, well, the database does, but then you're back into all the bad things about running your shadow database with a badly fitting model, and you'd be better off just ensuring all or part of the database ran closer to your app with replication, say, in memory on same server.
That would be a result set layer, not a connection pool. Could make sense if you worked with rows, but if you use ORM, why mapping cached row again and again? ORMs cache hydrated objects, which seems to be more efficient.
Forgive my ignorance, but how do ORMs help with adding caching? Or are you implying they obviate or reduce the need for caching?