I had some debate over whether to include this in the DataFrame or Array documentation. I chose the Array documentation because the chunks are an Array object.

Thanks for adding this @stsievert! It might be worth including a note that lengths=True will trigger an immediate computation.

Thanks @jrbourbeau! I've added the note. I mention that this enables downstream computations, but don't point to any examples in case they're fixed (e.g., slicing an array raises a NotImplementedError at slicing.py#L933).

I can see another use case with arrays:

>>> x = np.random.choice([-1, 0, 1], size=100)
>>> y = da.from_array(x, chunks=50)
>>> y[y != -1]
# dask.array<getitem, shape=(nan,), dtype=int64, chunksize=(nan,)>

I think computing the chunk size could be useful (e.g., with the slicing example above). Looks like #3293 (comment) is the relevant work.

After looking at this again, converting a Dask DataFrame to a Dask array (and the issue of .values giving unknown chunk sizes) is brought up in the "From Dask DataFrame" section of the Dask array creation documentation (dask/docs/source/array-creation.rst). Do you think this additional to_dask_array documentation is a better fit in that section?

Thanks for that @jrbourbeau! I think that's a better place, and I still link to it from chunks page.

I also improved that page a bit – that page didn't have a mention of to_dask_array, so I highlighted it there.

Thanks @stsievert!