issue_comments

This logic has been around in xarray for a long time (I think it dates back to https://github.com/pydata/xarray/pull/12!), so it predates me. If I had to guess though, it would have to do with the fact that back then, a form of cftime.date2num was used to encode all times, even those that started as np.datetime64 values. I think that's significant for two reasons: 1. In the old days, date2num would only return floating point values, even if the times could in principle be encoded with integers. For that reason, for accuracy reasons, it was best to keep the encoded values as small as possible to avoid roundoff error. 2. Even if (1) was not the case back then, date2num did not -- and still does not -- support nanosecond units, because it relies on microsecond-precision datetimes.

This of course is not true anymore. We no longer use date2num to encode np.datetime64 values, and we no longer encode dates with floating point values by default (#4045); we use integers for optimal round-tripping accuracy, and are capable of encoding dates with nanosecond units.

To be honest, currently it seems the only remaining advantage to choosing a larger time encoding unit and proximate reference date is that it makes the raw encoded values a little more human-readable. However, encoding dates with units of "nanoseconds since 1970-01-01" is objectively optimal for np.datetime64[ns] values, as it guarantees the maximum range of possible encoded times, and maximum round-trip accuracy, so it could be worth revisiting our approach in light of the fact that it makes appending somewhat dangerous.

Please may I ask: Why not default to xarray encoding time as 'units': 'nanoseconds since 1970-01-01' to be consistent with np.datetime64[ns]?

It's choosing the highest resolution that matches the data, which has the benefit of allowing the maximum possible time range given the data's frequency: https://github.com/pydata/xarray/blob/5871637873cd83c3a656ee6f4df86ea6628cf68a/xarray/coding/times.py#L317-L319

I'm not sure if this is why it was originally chosen; but that is one advantage. Perhaps @spencerkclark has some insight here.

Cool, I agree that an error and a documentation change is likely to be sufficient :slightly_smiling_face: (and I'd be keen to write a PR to help out!)

But, before we commit to that path: Please may I ask: Why not default to xarray encoding time as 'units': 'nanoseconds since 1970-01-01' to be consistent with np.datetime64[ns]? Sorry if I've missed something obvious!

let's definitely add a note to the documentation to say that it might be a good idea for users to manually specify the encoding for datetimes if they wish to append to Zarrs.

:+1:

However I think xarray should raise an error when trying to append times that cannot be represented by the on-disk encoding.

Adding this error message would make it obvious that this is happening. PRs are very welcome!

I think I've bumped into a symptom of this issue (my issue is described in #5969). And I think #3379 may be another symptom of this issue.

Perhaps I'm biased (because I work with timeseries which only span a few years) but I wonder if xarray should default to encoding time as 'units': 'nanoseconds since 1970-01-01' (to be consistent with np.datetime64[ns]) unless the timeseries includes dates before the year 1677, or after the year 2262 :slightly_smiling_face:? Would that work?

If that's no good, then let's definitely add a note to the documentation to say that it might be a good idea for users to manually specify the encoding for datetimes if they wish to append to Zarrs.

Yep I managed to overcome this by manually setting encoding parameters, just wondering if there would be any downside in preferring float64 over int64 when automatically defining these? This seems to fix that issue. I guess it could result in some other precision losses due to float-point errors but these should be small..

I agree with Deepak. Xarray intelligently chooses its encoding when it write the initial dataset to make sure it has enough precision to resolve all times. It cannot magically know that, in the future, you plan to append data which requires greater precision. Your options are: - If you know from the outset that you will require greater precision in time encoding, you can manually specify your encoding before you write (http://xarray.pydata.org/en/stable/io.html#scaling-and-type-conversions) - If you don't know from the outset, you will have to overwrite the full time variable with new encoding

I also agree that we should definitely be raising a warning (or even an error) in your situation.

I have run in to this problem before.

The initial choice to use int64 and days since ... is perfectly valid. However I think xarray should raise an error when trying to append times that cannot be represented by the on-disk encoding.

Note that you can always specify an encoding to make sure that you can append properly.