dev call 20220908

Dev Call: 8 September 2022

Simon, Ronan, Matti, Antonio, Du Toit, Tim, Stepan, Mohaned, Florian

In order to increase HPy's visibility, Tim and Carl Friedrich had the idea to do lightning talks about the topics/decisions/etc worked on during sprint at some dedicated time slot (like Thursday afternoon). Make it public.

Matti was working on HPy 0.0.4 branch in PyPy.
- Everything works except the debug mode tests about returning/accessing invalid handles. Accessing a closed handle is meant to be a fatal error. There is no way to recover. It works on CPython but it's not a strict requirement for other runtimes.
- Three options: (1) mark tests as CPython-only, (2) return some special marker object, or (3) don't allow a custom hook function.
- Stepan suggests that we could set a long-jump target when calling into native which allows to recover from that situation..
- There is a consens on making the custom hook function optional. Then runtimes could just disallow custom on_invalid_handle.
Stepan thinks about submitting an HPy talk to a CPython meetup (organized by Petr Viktorin) in Brno CZ.

Du Toit did some measurements with CPython and GraalPy.
- Compared original (not migrated) version to the migrated version.
- In all cases, the modules were built with the runtime they were in the end executed. This means we run with LLVM.

Not much due to vacations.
Florian concentrates on performance work on GraalPy side. GraalPy should get as close as possible to CPython performance.
Mohaned is working on making NumPy test and benchmarks run by removing cut-offs.
- Tests are already working.
- There is a problem because we switched to heap types. Slot inheritance is working a bit different in that case because tp_as_number is always not null in the case of heap types. This is something we need to figure out.

We are all a bit surprised about this project.
It seems that he learned a lot from HPy (but isn't mentioning us at all).
Mark has a completely different point of view. He doesn't care about migration, only about the CPython changes.
It's important for us to stay in touch with Mark. We expect that the project will struggle with similar problems as HPy concerning migration of packages. We already have a lot more experience in migration problems.
We also need to better communication what we are doing and why our solution is viable. It seems that our blog posts and conference talks were not enough so far.
Tim started to write a PEP but is now concerned that publishing this PEP could start a competition with Mark.
There is still no strict separation (in Mark's project) between internal and public API.
Stepan is not really sure if Mark wants a better internal or public API.
It seems that Mark (and his colleagues from Microsoft) are under high pressure and have tight deadlines. It seems they want to fix performance problems now which means that they also need to change the public API.
Antonio will invite Mark since he already had some contact in the last time.
Petr Viktorin (steering council member) pushed a lot back on Mark's efforts.