Start adding a performance section to the guide.

[adamreichold]

We discussed this quite often and I though that even if it is only a single thing, just making a start would improve our chances of getting it finished eventually.

[davidhewitt]

Nice, thanks for kicking this off!

One day it would be nice to remove this performance pitfall. Something to investigate after removing the pool 😄

[adamreichold]

@davidhewitt I added another subsection on implicit access to GIL token. Please have another look before this is merged.

Also do you remember any other topics when wanted to discuss here? I do think I remember that there about four different things we wanted to have here, but I am unable find a list anywhere in the older issues, pull requests or discussions.

[adamreichold]

Also do you remember any other topics when wanted to discuss here?

Is dictionary dispatch, i.e. look-up tables based on type object identity, something we want to discuss here or is this too specialized?

[davidhewitt]

Personally I think that's quite application-specific, but I'm not against having that section if you want to write it.

Some other topics I recalled over the day:

• Overhead of conversions (i.e. PyList vs Vec etc.)
• Threading / allow_threads
• String intern!
• Vec<u8> becoming list[int] (and thus being very slow)
• #[pyo3(get)] deep-cloning non-Py data

We can potentially move these into an issue rather than write them all in this PR?

[adamreichold]

Threading / allow_threads

Shouldn't this be covered by the "parallelism" section?

We can potentially move these into an issue rather than write them all in this PR?

Yes, I'd like to limit this PR to the two small sections already written.

Personally I think that's quite application-specific, but I'm not against having that section if you want to write it.

Ok, let's include a note in the issue. Certainly not in this PR.

[davidhewitt]

Sorry for the slow reply - I got a bit sidetracked by the various PRs I have pushed in the last 24h...

Shouldn't this be covered by the "parallelism" section?

Yes, though I think it can't hurt to call it out i.e. "how do I multithread? See parallelism section". More visibility :)