Do we think the adhoc_tool improvements could change the equation on whether Pants could become purely Rust because fewer people need to write plugins? I know this is a wildcard idea

(my bad I thought this was private chat. Long day, lol) I think more could get ported for sure, but: • I don't think the adhoc improvements model the world in the same, precise way a plugin would. Consider writing a Fortran backend solely in adhoc • We'd be abandoning one real factor to adoption, which is Python plugins are easier to write and iterate on ◦ And this doesn't account for existing plugins in people's repos

An idea I had a year ago is to allow people to write Rust plugins. (Very very hand-wavey) Then we port most of Pants to Rust. But at the end of the day, we'd still need a Python interpreter so I think that has fundamental limits. Compared to pure Rust

People can Write Rust plugins today... in a way. It's just very unfortunate that the boundary between plugins is a Python object, so you'd have Rust <-> Python <-> Rust. But we already do that today with our own code. I think

nogil

is our best bet at making that handshaking not "halting the world" so to speak

LATE: Relying entirely on

adhoc_tool

+ macros would be a worse experience than Pants 1 😉

I could imagine a world where: • adhoc tool or similar BUILD/macros could be used for (basic) dependency inference too • macros can be distributed in some convenient ways eg as pypi packages or Git repos • iterating on adhoc tool as the stable api boundary for “plugins” is less effort/gives a better result than making pants internals stable • (More powerful adhoc tools + a rust backend gives rust “plugins”: have a tool that runs an in-repo rust binary.) I’m not sure how closely my imaginings match reality. (I think adhoc tool is really impactful: right now, it’ll cut our codegen times by like 10x, and simplify the worst parts of our BUILD files, all without us having to keep up with plug-in changes.)

in my experience, the actual language implementations (“plugins”) rarely show up in profiles: at least 80% of runtime is in core @rule code (

graph.py

, etc) to that end, i’m going to spend a few days trying to port core data structures like `Address`/`Field`/`Target` to Rust, and then see about porting some of

graph.py

to Rust as intrisics. made 600 lines of (…untested =P ) progress today.

I should insert (untested) in more of my messaging.

Yeah it (untested) works!

i think (hope?) i chose the hardest one first… so much string/path manipulation!: https://github.com/pantsbuild/pants/pull/18989