Relax nanosecond datetime restriction in CF time coding

[kmuehlbauer]

• Closes Interoperability with Pandas 2.0 non-nanosecond datetime #7493
• Closes DataArray constructor still coerces to np.datetime64[ns], not cftime in 0.11.0 #2587
• Tests added/changed
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

This is another attempt to resolve #7493. This goes a step further than #9580.

The idea of this PR is to automatically infer the needed resolutions for decoding/encoding and only keep the constraints pandas imposes ("s" - lowest resolution, "ns" - highest resolution). There is still the idea of a default resolution, but this should only take precedence if it doesn't clash with the automatic inference. This can be discussed, though. Update: I've implemented time-unit-kwarg a first try to have default resolution on decode, which will override the current inferred resolution only to higher resolution (eg. 's' -> 'ns').

For sanity checking, and also for my own good, I've created a documentation page on time-coding in the internal dev section. Any suggestions (especially grammar) or ideas for enhancements are much appreciated.

There still might be room for consolidation of functions/methods (mostly in coding/times.py), but I have to leave it alone for some days. I went down that rabbit hole and need to relax, too 😬.

Looking forward to get your insights here, @spencerkclark, @ChrisBarker-NOAA, @pydata/xarray.

Todo:

• floating point handling
• Handling in Variable constructor
• update decoding tests to iterate over time_units (where appropriate)
• ...

[kmuehlbauer]

Nice, mypy 1.12 is out and breaks our typing, 😭.

[TomNicholas]

Nice, mypy 1.12 is out and breaks our typing, 😭

Can we pin it in the CI temporarily?

[kmuehlbauer]

Can we pin it in the CI temporarily?

Yes, 1.11.2 was the last version.

[kmuehlbauer]

This is now ready for a first round of review. I think this is already in a quite usable state.

But no rush, this should be thoroughly tested.

[spencerkclark]

Sounds good @kmuehlbauer! I’ll try and take an initial look this weekend.

[kmuehlbauer]

Thanks Stephan for the review. Looking into that next week.

[ChrisBarker-NOAA]

Not to throw too much of a wrench in the works here -- so feel free to disregard, but there's an issue I've faced with (single precision) float time encoding:

Folks (carelessly :-( ) sometimes encode times as "days since ..." using a single precision float. The problem here is not unnecessary precision, as you get with double, but too little -- if you go more than a few years out , you lose seconds precision. (the key problem is that float time -- its precision is a function of the magnitude -- not good for this use case)

The end result is that I get things like model timesteps that are supposed to be hourly, reporting as, e.g. 12:00:18, rather than 12:00:00

One way I've dealt with this is rounding to the minute, or even to hours (if I know the output is hourly), or perhaps to 10 minutes.

Could / should xarray provide a facility for doing this? maybe?

I guess what I'm proposing is that there be some way to tell xarray to store / save a time variable with e.g. second precision, but to round it to something more coarse when decoding.

maybe this could even be automatic / inferred:

if a time is in float days since -- it almost certainly is NOT millisecond precision, or even second precision -- and you could even look at the values (the first one?) and see what the minimum precision is for that timespan.

If I've done my math right, a float can only store second precision for a little over three years. So if the values are greater than three years, you don't have second precision.

Anyway, maybe way too much magic, but it would be nice for my use cases :-)

Example:

# 15 min timestep
In [57]: dates
Out[57]: 
[datetime.datetime(2024, 1, 1, 0, 0),
 datetime.datetime(2024, 1, 1, 0, 15),
 datetime.datetime(2024, 1, 1, 0, 30),
 datetime.datetime(2024, 1, 1, 0, 45)]

# common choice of units (though a bad one :-( )
In [58]: units
Out[58]: 'days since 1970-01-01T00:00:00'

# convert to numbers, used float64 by default
In [59]: nums_double = nc4.date2num(dates, units)

# truncate to float32
In [60]: nums_float = nums_double.astype(np.float32)

# convert back to datetimes:
In [61]: dates_float = nc4.num2date(nums_float, units)

In [62]: dates_float
Out[62]: 
array([cftime.DatetimeGregorian(2024, 1, 1, 0, 0, 0, 0, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 14, 3, 750000, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 30, 56, 250000, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 45, 0, 0, has_year_zero=False)],
      dtype=object)
In [67]: [str(dt) for dt in dates_float]
Out[67]: 
['2024-01-01 00:00:00',
 '2024-01-01 00:14:03.750000',
 '2024-01-01 00:30:56.250000',
 '2024-01-01 00:45:00']

Ouch! so what were 15 minute timesteps is now off by about one minute -- and what's too bad is that rounding to the minute wouldn't be right either -- you'd need to round to maybe 5 minutes?

Anyway, maybe this simply isn't xarray's problem to solve -- data providers shouldn't make such mistakes :-(

[spencerkclark]

@ChrisBarker-NOAA yeah, I agree this kind of situation is annoying, but my feeling is that trying to fix this automatically would be too much magic. Xarray has convenient functionality for rounding times, which can be used to correct this explicitly—that would be my preference. E.g. for your example it would look like:

>>> decoded
<xarray.DataArray 'time' (time: 4)> Size: 32B
array([cftime.DatetimeGregorian(2024, 1, 1, 0, 0, 0, 0, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 14, 3, 750000, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 30, 56, 250000, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 45, 0, 0, has_year_zero=False)],
      dtype=object)
Dimensions without coordinates: time
>>> decoded.dt.round("5min")
<xarray.DataArray 'round' (time: 4)> Size: 32B
array([cftime.DatetimeGregorian(2024, 1, 1, 0, 0, 0, 0, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 15, 0, 0, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 30, 0, 0, has_year_zero=False),
       cftime.DatetimeGregorian(2024, 1, 1, 0, 45, 0, 0, has_year_zero=False)],
      dtype=object)
Dimensions without coordinates: time

[ChrisBarker-NOAA]

Xarray has convenient functionality for rounding times

Oh, nice: I had missed that! -- you're probably right, too much magic to do for people.

[kmuehlbauer]

@spencerkclark @ChrisBarker-NOAA I've implemented automated decoding of floating point data to the needed resolution, even when the wanted resolution does not apply.

Unfortunately the above outlined behaviour is too much involved to be put into the decoder. Nevertheless maybe we can distill some best practices from your vast experience with data @ChrisBarker-NOAA and create a nice example how to handle these difficulties?

[ChrisBarker-NOAA]

create a nice example how to handle these difficulties?

Sure -- where would be a good home for that?

[kmuehlbauer]

Not sure, but https://docs.xarray.dev/en/stable/user-guide/time-series.html could have a dedicated floating point date section.