I don’t think there’s been a solid discussion of what the requirements are of predicate pushdown support, and I’d be interested in seeing the requirements and desires of various potential layers’ uses of a predicate pushdown API before considering how to implement one.
Specifically, if you and anyone else interested in the feature could please detail:
- What data model you have, and what predicate/computation you want to be able to push down to FDB?
- Why does this predicate/computation need to be pushed down to FDB, and can’t as effectively or easily be a part of a layer?
I’m concerned that there’s probably a number of issues that complicate creating a layer that’s very efficient and tightly integrated to FDB, and I’d like to make sure we understand specifically what slice of that would be best solved by predicate pushdown (and thus define our requirements), and what problems could be better improved by some other piece of work.