Add an Cumulative aggregation, similar to Rolling

[max-sixty]

Is your feature request related to a problem? Please describe.

Pandas has a .expanding aggregation, which is basically rolling with a full lookback. I often end up supplying rolling with the length of the dimension, and this is some nice sugar for that.

Describe the solution you'd like
Basically the same as pandas — a .expanding method that returns an Expanding class, which implements the same methods as a Rolling class.

Describe alternatives you've considered
Some options:
– This
– Don't add anything, the sugar isn't worth the additional API.
– Go full out and write specialized expanding algos — which will be faster since they don't have to keep track of the window. But not that much faster, likely not worth the effort.

[mathause]

I don't entirely get what the function is supposed to do (even after looking at the pandas docstring) - can you give a an example or two?

[dcherian]

IIUC da.expanding(dim=dim).sum() is da.cumsum(dim) with support for min_periods and center like rolling.

I guess all expanding reductions are basically numpy.ufunc.accumulate(...). The dask versions will be "interesting" to write.
https://github.com/dask/dask/blob/f1f37cae96d5e98f8043ea430539a4fffbe62661/dask/array/reductions.py#L1389-L1413

Like @mathause I find "expanding" confusing. .accumulate().sum() or .cumulative().sum() sounds much better to me.

[max-sixty]

.cumulative is great! Much better.

The benefit is that the API surface is reduced — e.g. we can have a .cumulative().integrate(), rather than a separate .cumulative_integrate (and so on, for each aggregation), from #5153.

The implementation could be as simple as da.rolling(dim=da.sizes[dim]). How compatible would dask be with that? How does it compare to the numpy.ufunc.accumulate(...) suggestion?

[max-sixty]

I'd be up for doing .cumulative if I continue my recent contribution burst...

[mathause]

Go for it - we can hardly keep up with reviewing ;-)

[shoyer]

I am pretty stoked about the idea of supporting cumulative integrals with the same syntax as other cumulative operations :)