TimeSeries doesn't know which dimension is time

[ctw]

@mivade & @leondavis1 : I've realized that our plan to check that "time" is in the dims does not work, because the coords themselves are TimeSeries objects, as well. For example ts['channels'] is a TimeSeries instance without a "time" dimension (and would thus throw an error if such a check were in place). I think the solution is to no longer subclass xarray, but obviously that would be a fairly major refactor. Thoughts?

[ariel215]

If we wanted to leave TimeSeries as a DataArray subclass, we could override DataArray._replace to be something like this:

def _replace(self,variable=None,coords = None, name=__default):
        if variable is None:
            variable = self.variable
        if coords is None:
            coords = self._coords
        if name is self.__default:
            name = self.name
        if variable is self.variable:
            return type(self)(variable,coords,name=name,fastpath=True)
        else:
            return DataArray(variable,coords,name=name,fastpath=True)

[mivade]

I don't like the idea of overriding private methods. Instead, is there some trick we can do with properties to get a simple array back when we want to know the time coordinates?

EDIT: see this for example.

[ariel215]

A DataArray has two ways to access coordinates: either through the coords property, or directly using __getitem__. Both of these delegate to _getitem_coord, which (in the case of a coordinate name) delegate to _replace. If you don't want to override private methods and you also want the constructor to enforce having a coordinate named "time" , you probably to override both coords and __getitem__.

I don't think there's a good way to make sure every TimeSeries object has a dimension representing time, other than to enforce this in the constructor. The other options are:

• Have a property indicating which axis should be regarded as the 'time' axis
• Assume that every TimeSeries has a 'time' axis, with the caveat that some functions will crash if this is not the case. This is more or less what we're doing now.

[ariel215]

My suggestion doesn't work as written, because it has the side effect of allowing operations on TimeSeries objects with differing samplerates to succeed silently, which is a much more relevant issue than not having a 'time' axis.