Fix fixed offset timezone conversion bug.

[grantslatton]

Closes #3267

[adamreichold]

Since we are going through a license change, please also have a loot at #3108

[adamreichold]

Not sure what the Windows build failures are about. Is the Python installation missing the time zone database there?

[grantslatton]

Not sure what the Windows build failures are about. Is the Python installation missing the time zone database there?

It's because Windows only supports times in 1970-2038 I believe, but the proptest is using a much larger range, I am fixing this now

[grantslatton]

Hunting down some more bugs the more exhaustive proptest found. Found one in chrono: chronotope/chrono#1154

There is another one that results in the python vs rust datetimes being 1 second off when the time falls exactly on the daylight savings switchover. Unsure if pyo3 conversion bug (with the fold parameter) or just a discrepancy between python and chrono. Investigating further...

[pitdicker]

@grantslatton I am interested in the issues you may run into.

[grantslatton]

@grantslatton I am interested in the issues you may run into.

@pitdicker Does chrono have facilities for supporting ambiguous instants created at the end of daylight savings? i.e. https://peps.python.org/pep-0495/

[pitdicker]

Yes, you will want to look into methods that return a LocaleResult such as TimeZone::from_local_datetime.

[grantslatton]

I am wondering if we should also fix #3266 in the same change, because this change will result in fold information being lost on FromPyObject, because PyO3's doesn't have a conversion out to DateTime<Tz> -- only DateTime<Utc> and DateTime<FixedOffset.

But doing this would require changing the IntoPy for DateTime<Tz: TimeZone> path to different impls: one for DateTime<Utc>, DateTime<FixedOffset>, DateTime<Local>, and DateTime<Tz>. This would be a breaking change that would break any downstream code that has defined their own custom TimeZone impl. There is probably no such code, as everyone is probably using timezones defined by chrono and chrono_tz.

I think such a change increases the correctness of this crate, because right now the conversion is lossy and loses daylight savings information present in chrono_tz::Tz locale timezones.

[adamreichold]

This would be a breaking change that would break any downstream code that has defined their own custom TimeZone impl. There is probably no such code, as everyone is probably using timezones defined by chrono and chrono_tz.

This would mean the fix would also be limited to 0.20, the next release where breakage is possible. So I think the question is whether the side effects of losing the fold information is serious enough that one would say the backwards-compatible version of the fix cannot be used anyway?

Personally, I would be fine with try to fix this properly for 0.20 only because the chrono integration is still a relatively new feature of PyO3 and some churn seems expected.

[adamreichold]

P.S.: You seem to have included two unrelated commits during your rebase.

[Psykopear]

Ok, sorry for the delay. I've reviewed the issue and the PR, the bug is there, and this fixes it, nice catch.

I am wondering if we should also fix #3266 in the same change, because this change will result in fold information being lost on FromPyObject, because PyO3's doesn't have a conversion out to DateTime<Tz> -- only DateTime<Utc> and DateTime<FixedOffset>.

The problem with the chrono_tz integration IIRC, is that python<3.9 doesn't have the zoneinfo package in the standard library, and I'm not sure it's reasonable to implement the integration without that. So, unless there is an alternative solution I missed (that doesn't require adding a python dependency like backports.zoneinfo and/or tzdata), we'd probably need to put the chrono_tz integration under a separate feature (dependent on python>=3.9), and keep only the chrono one for python<3.9. The two features would also have to be mutually exclusive I think, because we can't implement the python traits for both chrono::Tz (the trait) and chrono_tz::Tz (the enum).

All that to say that it's probably ok to lose the fold info with just chrono (since we only handle fixedoffsets), and fix this for python>=3.9 in the chrono_tz integration.
Gating the new integration under a separate feature will also avoid breaking code that uses the current one, because you'd need to enable the chrono_tz feature.

[adamreichold]

@grantslatton @Psykopear Ok, you to summarize my understanding: We take as it is and it fixes #3267 but #3266 will need a more involved approach which might even use an independent code path. Do you both agree with this?

[grantslatton]

@grantslatton @Psykopear Ok, you to summarize my understanding: We take as it is and it fixes #3267 but #3266 will need a more involved approach which might even use an independent code path. Do you both agree with this?

Yes this is right