REGR: Index.union should pick correct dtype for combinations of tstamps with different tzones

[zogzog]

Here's a pdb session showing the issue:

ipdb> basei
DatetimeIndex(['2017-10-28 23:00:00+00:00', '2017-10-29 00:00:00+00:00',
               '2017-10-29 01:00:00+00:00', '2017-10-29 02:00:00+00:00'],
              dtype='datetime64[ns, UTC]', freq=None)
ipdb> diffi
DatetimeIndex(['2017-10-29 02:00:00+01:00', '2017-10-29 03:00:00+01:00',
               '2017-10-29 04:00:00+01:00', '2017-10-29 05:00:00+01:00'],
              dtype='datetime64[ns, Europe/Paris]', freq=None)
ipdb> basei.union(diffi)
Index([2017-10-28 23:00:00+00:00, 2017-10-29 00:00:00+00:00,
       2017-10-29 01:00:00+00:00, 2017-10-29 02:00:00+00:00,
       2017-10-29 04:00:00+01:00, 2017-10-29 05:00:00+01:00],
      dtype='object')

With pandas 0.24 I got back an index with a dtype='datetime64[ns, UTC]'

This looks a bit like a cousin of #26778

[TomAugspurger]

I think that was a deliberate change: https://pandas.pydata.org/pandas-docs/stable/whatsnew/v0.25.0.html#incompatible-index-type-unions.

cc @ArtinSarraf if you recall whether mix of datetimetz was discussed.

[zogzog]

I've seen the "incompatible types" discussion but I feel like this is going too far. Admittedly this is an edge case (like in the int vs float discussion).

[TomAugspurger]

The ints and float case is a reasonable analogy. But with ints & floats it's not clear (to me) that the loss of precision for integers is the best outcome. You could make a similar argument for datetimetz and daylight savings issues, when combining multiple timezones.

…

On Tue, Aug 20, 2019 at 11:16 AM Aurélien Campéas ***@***.***> wrote: I've seen the "incompatible types" discussion but I feel like this is going too far. Admittedly this is an edge case. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#28034?email_source=notifications&email_token=AAKAOIVRMBSA32VFWVXVQW3QFQKFLA5CNFSM4INWXMPKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4W23KQ#issuecomment-523087274>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAKAOIW3KR4SSYMPPVWTZYTQFQKFLANCNFSM4INWXMPA> .

[mroeschke]

I agree that this is a deliberate change in the right direction as to prevent the loss of original timezone information akin to this change back in 0.24.0: https://pandas.pydata.org/pandas-docs/stable/whatsnew/v0.24.0.html#parsing-datetime-strings-with-timezone-offsets

You can always to_datetime(..., utc=True) to coerce your result to UTC.

[zogzog]

Timezone information, as far as I am concerned, is a presentation information, a bit like the 'freq' attribute. Two timestamps in differing timezones are fundamentally compatible, and losing timezones is about dropping display information. Are there situations where this would be wrong ?

Another point: it already cost me an hour to write code to paper over the issue. I have a non-indecent test base, but real world tests reveal that I have non-handled edge cases. So it's not just a matter of .tz_converting the offending index and the cost of reimplementing myself what .union did previously is going to mount. In other words, this is a breaking change.

[jreback]

Timezone information, as far as I am concerned, is a presentation information, a bit like the 'freq' attribute. Two timestamps in differing timezones are fundamentally compatible, and losing timezones is about dropping display information. Are there situations where this would be wrong ?

this is not the case at all in pandas; time zones are first class; turning them to UTC is losing information; we have deliberately tried not to do this generally

mixing times zones while possible is often a bug in user code; it’s possible but very likely not something you want library code to do, rather it becomes a burden of the user code

Another point: it already cost me an hour to write code to paper over the issue. I have a non-indecent test base, but real world tests reveal that I have non-handled edge cases. So it's not just a matter of .tz_converting the offending index and the cost of reimplementing myself what .union did previously is going to mount. In other words, this is a breaking change.

not really sure what is your point here. pandas has lots of users; api changes are a fact of life

[ms7463]

@TomAugspurger - yes, this exact case was discussed explicitly.

#23538 (comment)

Also FWIW, I think that the fact alone, that converting from an dst aware tz for the Nov dst transition hour can be ambiguous, is enough to disqualify automatically converting to UTC, for mismatched tz merges.

[zogzog]

Thanks for the reference. Maybe then a small deprecation period would have been gentler ?

@ArtinSarraf can you expand on your dst transition ambiguity example ? For my personal education I'd like to understand the issue (if only because I currently believe tz-aware stamps don't have ambiguity issues).

[ms7463]

Ignore my comment, what I was thinking of wasn't an issue with already tz-aware date ranges, but rather the ambiguity than can occur from localizing a non tz-aware date range.

[zogzog]

The very fact tz-aware tstamps are non-abiguous is the reason I think this is a regression. The actual timezones are display information (just like freq). There is no fundamental information loss when the time zone is aligned to UTC.

[jreback]

@zogzog pls read my comments again. there is loss of information as pandas keeeps the tz meta data as part of the dtype; so by definition you are losing information.

[zogzog]

I've read your comment. "By definition" doesn't cut it. Did you read mine and try to understand the argument and address it ?

[mroeschke]

@zogzog this is what @jreback is referring to:

In [1]: utc = pd.Series(pd.date_range('2019', periods=3, freq='D', tz='UTC'))

In [2]: pacific = utc.dt.tz_convert('US/Pacific')

In [3]: utc
Out[3]:
0   2019-01-01 00:00:00+00:00
1   2019-01-02 00:00:00+00:00
2   2019-01-03 00:00:00+00:00
dtype: datetime64[ns, UTC]

In [4]: pacific
Out[4]:
0   2018-12-31 16:00:00-08:00
1   2019-01-01 16:00:00-08:00
2   2019-01-02 16:00:00-08:00
dtype: datetime64[ns, US/Pacific]

# What you are probably referring to
In [5]: utc == pacific
Out[5]:
0    True
1    True
2    True
dtype: bool

In [6]: utc.dtype
Out[6]: datetime64[ns, UTC]

In [7]: pacific.dtype
Out[7]: datetime64[ns, US/Pacific]

# The loss of information that would occur if Index.union coerced to UTC
In [8]: utc.dtype == pacific.dtype
Out[8]: False