I think that masking the construct-ed array would make the full copy of this, while originally construct returns a view of the original array. Another option would be to support an additional method (get_mask?), but as this can be realized quite easily (rolling.count() > rolling.min_periods) I am not sure how worth it is.

Would it make sense to support min_periods in construct by masking the result? That would arguably be consistent with how we do iteration over rolling objects, though to be honest I'm not sure how useful it really is.

I think we need to document more clearly that min_periods option is just neglected if using construct method.

I also think we could label construct as advanced / experimental. It's extremely cool, but for traditional cases, it's probably more cool than required, and less well tested than the standard methods.

At the moment we say "You can use this for more advanced rolling operations, such as strided rolling, windowed rolling, convolution, short-time FFT, etc."; I'm imagining adding something like "For standard operations above, there is no need to use construct". Or adding an Experimental marker

Ok, so ds.rolling.mean allows for min_periods while ds.rolling.construct.mean allows for fill_value?

Yes. Originally, construct method is designed for more advanced use cases, such as strided moving average, weighted mean, short time FFT, etc.

On the other hand, using skipna=False resores the default behaviour or ds.rolling

It is not strictly true. min_periods also considers nan values that already exist in the original array. This behavior is the same with pandas rolling.

Ok, so ds.rolling.mean allows for min_periods while ds.rolling.construct.mean allows for fill_value?

On the other hand, using skipna=False resores the default behaviour or ds.rolling ``` ds.rolling(dim=3, center=True).construct('window').mean('window', skipna=False)

RESULT: array([nan, 1., 2., nan])

```

Actually I'm sure that min_periods and fill_value are not the only differences between the two - but that's how far my understanding got at the moment.

I think we need to document more clearly that min_periods option is just neglected if using construct method.

Hi, @mathause. Thanks for reporting.

I agree that this behavior is a little surprising, but this is something expected. The difference is not skipna=True (both use this), but actually in min_periods. If you specify min_periods=1, then the results become identical. python In [7]: ds.rolling(dim=3, center=True, min_periods=1).mean() Out[7]: <xarray.DataArray (dim: 4)> array([0.5, 1. , 2. , 2.5]) Dimensions without coordinates: dim

I am not sure how we can make both the behaviors the same with keeping the generality of construct method.