Hey@serhiy-storchaka, I believe you're pretty familiar with the pathlib internals. Penny for your initial impressions of this patch? It does cut out a key part of the initial pathlib design: the idea of skipping re-parsing and re-normalizing paths in case where we can be sure it's not necessary (e.g. inpath.parents,path.iterdir()). I'd love to be able to convince you this is a good idea!

This is now ready for review! My previous comment no longer applies, as we fixed the pathlib construction weirdness in#102789. This change should be pretty performance-neutral now.

My biggest concern is that we're adding a breaking change to_from_parsed_parts which is not a big problem on its own because the interface is private, but I would still take a look around for some relatively popular libraries that might actually be dependent on that interface.

Well, it's undocumented and underscore-prefixed. We've been breaking other things like that over the last few months, e.g. removingPath._accessor. It's hard to make progress without doing this.

I'm also slightly concerned that this design might be a little fragile. The following situation seems fairly plausible:

1. Barney takes a break from open source for a few months2. During Barney's break, a contributor files a PR to add `awesome_new_method` to pathlib. The method is so awesome, it's immediately clear to the core dev team that we desperately need the method.3. Alex merges the PR, but because Barney's on an open-source break, there's nobody to remind us that the new method should really be calling `self.makepath(*args)` instead of `self.__class__(*args)`. Support for sharing state between pathlib subclasses is now broken :(

Is there a way we can guard against this?

Hm. I expect a user will spot that it doesn't work and log a bug. It doesn't have any wider consequences, I don't think.

Okay, I chatted with Brandt at the sprint, and he persuaded me that I was worrying too much about this hypothetical scenario :)

FWIW, I am open to alternatives to this new method. There are things we can do like pass an opaque object around, or derive a new type to 'bind' the context. I went back and forth with these, and in the end went with this approach as it appears to be the simplest. But happy to look again if you have suggestions :)

Also as a bit of a testimonial I really like the design and have the changes for something I locally implemented for this PR and it makes my code significantly nicer

Thanks so much for the thorough review! I'll open a thread about the method name shortly.

Yeah, I was missing the wood for the trees there. Massive improvement.

I've created a branch/PR that uses a__newpath__() method, for comparison's sake:

• GH-100479: Addpathlib.PurePath.with_segments() #103975

Closing in favour of#103975.