Query sharding is at the data source layer, which Haxl doesn't delve into. It's up to each data source integration with Haxl to do the appropriate routing/etc.

Hope you find it useful!

That said, the way it currently works is more like the first. You can think of the entire haxl run (program) as an AST that is given to the execution. It expands as much of the AST as possible (anything that's not IO), and anywhere it needs IO it enqueues those requests to be scheduled. Once it's explored as much as possible, it aggressively schedules the IO (deduping, batching, and overlapping the calls). Once it all comes back, it unblocks the AST where it can, and repeats the process.

This isn't necessarily the optimal scheduling (as you point out, unblocking each part of the tree as each result comes in might be better). It was specifically designed to make it easy to play with this kind of stuff later. Since the concurrency is entirely implicit the implementation is entirely abstracted away.

Interpreted code was no longer cutting it for perf reasons, and any time you create your own language you end up reinventing the entire tool chain (debuggers, profilers, etc.). Haskell provides so much functionality in the language itself and has mature solutions to the other issues plaguing us in FXL, so it was a natural choice.