• [x] Tests added
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst

Small PR to add a new "method" to convert to and from 360_day calendars. The current two methods (chosen with the align_on keyword) will always remove or add the same day-of-year for all years of the same length.

This new option will randomly chose the days, one for each fifth of the year (72-days period). It emulates the method of the LOCA datasets (see web page and article ). February 29th is always removed/added when the source/target is a leap year.

I copied the implementation from xclim (which I wrote), see code here .

• [ ] Closes #xxxx
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Simply add a encodings argument to save_mfdataset. As for the other args, it expects a list of dictionaries, with encoding information to be passed to to_netcdf for each dataset. Added a minimal test, simply to see if the argument was taken into account.

• [ ] Closes #xxxx
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

When playing with cftime objects a problem I encountered many times is that I can sub two arrays and them add it back to another. Subtracting to cftime datetime arrays result in an array of np.timedelta64. And when trying to add it back to another cftime array, we get a UFuncTypeError because the two arrays have incompatible dtypes : '<m8[ns]' and 'O'.

Example: ```python import xarray as xr da = xr.DataArray(xr.cftime_range('1900-01-01', freq='D', periods=10), dims=('time',))

An array of timedelta64[ns]

dt = da - da[0]

da[-1] + dt # Fails ```

However, if the two arrays were of 'O' dtype, then the subtraction would be made by cftime which supports datetime.timedelta objects.

This solution here adds a to_pytimedelta to the TimedeltaAccessor, mirroring the name of the similar function on pd.Series.dt. It uses a monkeypatching workaround to prevent xarray to case the array back into numpy objects.

The user still has to check if the data is in cftime or numpy to adapt the operation (calling dt.to_pytimedelta or not), but custom workaround were always overly complicated for such a simple problem, this helps.

Also, this doesn't work with dask arrays because loading a dask array triggers the variable constructor and thus recasts the array of datetime.timedelta to numpy.timedelta[64].

I realize I maybe should have opened an issue before, but I had this idea and it all rushed along.

• [x] Closes #5155
• [x] Tests added
• [x] Passes pre-commit run --all-files
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [x] New functions/methods are listed in api.rst

So:

• Added coding.cftime_offsets.date_range and coding.cftime_offsets.date_range_like The first simply swtiches between pd.date_range and xarray.cftime_range according to the arguments. The second infers start, end and freq from an existing datetime array and returns a similar range in another calendar.

• Added coding/calendar_ops.py with convert_calendar and interp_calendar Didn't know where to put them, so there they are.

• Added DataArray.dt.calendar. When the datetime objects are backed by numpy, it always return "proleptic_gregorian".

I'm not sure where to expose the function. Should the range-generators be accessible directly like xr.date_range?

The convert_calendar and interp_calendar could be implemented as methods of DataArray and Dataset, should I do that?

• [x] Closes #4190
• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] User visible changes (including notable bug fixes) are documented in whats-new.rst
• [ ] New functions/methods are listed in api.rst

Fixes #4190. In cases where the input matrix had a deficient rank (matrix rank != order) because of the number of NaN values, polyfit would fail, simply because numpy's lstsq returned an empty array for the residuals (instead of a size 1 array). This fixes the problem by catching the case and returning np.nan instead.

The other point in the issue was that RankWarning is also not raised in that case. That was due to the fact that da.polyfit was computing the rank from the coordinate (Vandermonde) matrix, instead of the masked data. Thus, is a given line has too many NaN values, its deficient rank was not detected. I added a test and warning at all places where a rank is computed (5 different lines). Also, to match np.polyfit behaviour of no warning when full=True, I changed the warning filters using a context manager, ignoring the RankWarning in that case. Overall, it feels a bi ugly because of the duplicated code and it will print the warning for every line of an array that has a deficient rank, which can be a lot...

• [x] Closes #4123
• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

Based on comments on dask/dask#3096, I fixed the dask indexing error that occurred when idxmax/idxmin were called on ND data (where N > 2). Added tests are very simplistic, I believe the 1D and 2D tests already cover most cases, I just wanted to test that is was indeed working on ND data, assuming that non-dask data was already treated properly.

I believe this doesn't conflict with #3936.

• [ ] Closes #xxxx
• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

Pull #3733 added da.polyfit and xr.polyval and is using xr.core.missing.get_clean_interp_index in order to get the fitting coordinate. However, this method is stricter than what polyfit needs: as in numpy.polyfit, non-unique and non-monotonic indexes are acceptable. This PR adds a strict keyword argument to get_clean_interp_index so we can skip the uniqueness and monotony tests.

ds.polyfit and xr.polyval were modified to use that keyword. I only added tests for get_clean_interp_index, could add more for polyfit if requested.

• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

This PR adds a xr.infer_freq method to copy pandas infer_freq but on CFTimeIndex objects. I tried to subclass pandas _FrequencyInferer and to only override as little as possible.

Two things are problematic right now and I would like to get feedback on how to implement them if this PR gets the dev's approval.

1) pd.DatetimeIndex.asi8 returns integers representing nanoseconds since 1970-1-1, while xr.CFTimeIndex.asi8 returns microseconds. In order not to break the API, I patched the _CFTimeFrequencyInferer to store 1000x the values. Not sure if this is the best, but it works.

2) As of now, xr.infer_freq will fail on weekly indexes. This is because pandas is using datetime.weekday() at some point but cftime objects do not implement that (they use dayofwk instead). I'm not sure what to do? Cftime could implement it to completly mirror python's datetime or pandas could use dayofwk since it's available on the TimeStamp objects.

Another option, cleaner but longer, would be to reimplement _FrequencyInferer from scratch. I may have time for this, cause I really think a xr.infer_freq method would be useful.

• [x] Closes #3349
• [x] Tests added
• [x] Passes isort -rc . && black . && mypy . && flake8
• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

Following discussions in #3349, I suggest here an implementation of polyfit and polyval for xarray. However, this is still work in progress, a lot of testing is missing, all docstrings are missing. But, mainly, I have questions on how to properly conduct this.

My implementation mostly duplicates the code of np.polyfit, but making use of dask.array.linalg.lstsq and dask.array.apply_along_axis for dask arrays. The same method as in xscale.signal.fitting.polyfit, but I add NaN-awareness in a 1-D manner. The version with numpy is also slightly different of np.polyfit because of the NaN skipping, but I wanted the function to replicate its behaviour. It returns a variable number of DataArrays, depending on the keyword arguments (coefficients, [ residuals, matrix rank, singular values ] / [covariance matrix]). Thus giving a medium-length function that has a lot of duplicated code from numpy.polyfit. I thought of simply using a xr.apply_ufunc, but that makes chunking along the fitted dimension forbidden and difficult to return the ancillary results (residuals, rank, covariance matrix...).

Questions: 1 ) Are the functions where they should go? 2 ) Should xarray's implementation really replicate the behaviour of numpy's? A lot of extra code could be removed if we'd say we only want to compute and return the residuals and the coefficients. All the other variables are a few lines of code away for the user that really wants them, and they don't need the power of xarray and dask anyway.