bpo-44490: Add __parameters__ and __getitem__ to types.Union

[uriyyo]

https://bugs.python.org/issue44490

[uriyyo]

@Fidget-Spinner Could you please review this PR?

[Fidget-Spinner]

@uriyyo thank you for sending this PR so quickly! Excellent work. I only have a few minor readability nits below.

[uriyyo]

@Fidget-Spinner Thanks for review, I have added tests that you suggested.

[Fidget-Spinner]

@Fidget-Spinner Thanks for review, I have added tests that you suggested.

Thanks for applying the suggestions. This LGTM. I've requested a review from Guido and I'm waiting for him to take a look when he has the time. I can't merge PRs, sorry.

[Fidget-Spinner]

@uriyyo can you please rebase this against main branch? I fixed a GC bug in Union in 1097384, and I think you may have to add Py_VISIT(alias->parameters) in union_traverse Thanks.

[uriyyo]

@Fidget-Spinner Thanks, fixed

[gvanrossum]

LG, I would just update two comments slightly.

[Fidget-Spinner]

Eventually if this PR is merged, we may have to consider backporting this to 3.10. Since it's a somewhat big change, I'm pinging @pablogsal as the 3.10 RM for his awareness.

TLDR: Some uncommon but valid use cases not covered by PEP 604 were discovered in bpo. This PR adds support for them. Can it land in 3.10 beta 4 too?

[pablogsal]

Eventually if this PR is merged, we may have to consider backporting this to 3.10. Since it's a somewhat big change, I'm pinging @pablogsal as the 3.10 RM for his awareness.

TLDR: Some uncommon but valid use cases not covered by PEP 604 were discovered in bpo. This PR adds support for them. Can it land in 3.10 beta 4 too?

We are too far in the development cycle to add new features, specially stuff that is touching c code. I do understand that this is something new in 3.10 so we do not have possibility of regressions and I see that this is an important point. I think we are fine to get this in with the following requirements:

• Is merged before the last beta (beta4)
• At least 2 core Devs review and approve the PR

[gvanrossum]

Thanks!

[gvanrossum]

@Fidget-Spinner Can you think of another core dev to review this (not Pablo)?

[pablogsal]

@Fidget-Spinner Can you think of another core dev to review this (not Pablo)?

I can do a review if you don't find anyone else, but currently, I am flooded with PEP 657 stuff so it will probably be in a couple of days

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @pablogsal for commit 7799a15 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[Fidget-Spinner]

@Fidget-Spinner Can you think of another core dev to review this (not Pablo)?

@serhiy-storchaka , are you alright with reviewing this?

In the meantime, the Windows 10 buildbot failures seem potentially related so @uriyyo please look into them. I'll take a look too in a day or two.

[pablogsal]

In the meantime, the Windows 10 buildbot failures seem potentially related so @uriyyo please look into them. I'll take a look too in a day or two.

I am restarting them as I am quite sure this is due to corrupted pyc files from a previous run. New builds here:

https://buildbot.python.org/all/#/builders/577/builds/86
https://buildbot.python.org/all/#/builders/593/builds/89

[Fidget-Spinner]

In the meantime, the Windows 10 buildbot failures seem potentially related so @uriyyo please look into them. I'll take a look too in a day or two.

I am restarting them as I am quite sure this is due to corrupted pyc files from a previous run. New builds here:

https://buildbot.python.org/all/#/builders/577/builds/86
https://buildbot.python.org/all/#/builders/593/builds/89

Oh you're right. Sorry @uriyyo if I led you on a wild goose chase. I suspect one of two things

• maybe someone forgot to bump the magic number for a Windows-related change.
• the worker maybe didn't clean up old files properly -- I suspect it's this one because I cannot reproduce this locally on Windows machine.

[Fidget-Spinner]

After spending a few hours investigating this. I think our best bet is to merge latest main into this PR and try running the buildbots again. I have no clue why those 2 buildbots fail and any help would be much appreciated. Some findings:

• They are 2 different workers. I find it unlikely that 2 machines could both have corrupted/out of date pyc files coincidentally.
• I've tried building this locally with same build flags as the buildbots, but I can't reproduce it.
• I noticed the buildbot/clean.bat script fails on both. But they've been like this for at least a month and I suspect it's because there are no leftover pyc files to clean. Maybe the buildbots already clean it up themselves?

[gvanrossum]

Yeah, merge latest main is definitely a good first step.

[uriyyo]

@gvanrossum@Fidget-Spinner I have merged latest main into PR branch.

And I have a question should we fix more cases in the scope of this PR?

1. types.Union is not serializable and can not be used with pickle.
2. types.Union can not be used with typing.Annotated.
3. typing._collect_type_vars currently do not collect params from the types.Union.

>>>typing.List[list[T] |float].__parameters__ ()

UPD. Actually, there is a lot of case at typing module where types.Union is not supported but I believe it should be.

[gvanrossum]

Tests pass now, I propose to merge now and do another PR for the other issues.

We still need another core dev approval, but I'm merging now. ("Merged before b4" was one of the requirements, "two core devs" is the other.)

[uriyyo]

Great, I will open another PR to cover cases that I mentioned.

[serhiy-storchaka]

There are several bugs (bpo-44633, bpo-44653), but I am going to fix them soon.

I am +1 for backporting this feature to 3.10.

[uriyyo]

@serhiy-storchaka Do you need help with it? There is also issue with types.Union serialisation that I am going to fix.

[pablogsal]

There are several bugs (bpo-44633, bpo-44653), but I am going to fix them soon.

I am +1 for backporting this feature to 3.10.

Please, see my other comment: #27207 (comment)