issue_comments

On a somewhat related note, I am now proposing extending xarray's "lazy array" functionality to include limited support for arithmetic, without necessarily using dask: https://github.com/pydata/xarray/issues/2298

@rabernat actually in #1532 we switched to not displaying a preview of any lazily loaded data on disk -- even if it isn't loaded with Dask. (I was not sure about this change, but I was alone in my reservations.)

I do agree that our lazy arrays serve a useful purpose currently. I would only consider removing them if we can improve Dask so it works just as well for this use case.

Yeah, we could solve this by making dask a requirement only if you want load netCDF files and/or load netCDF files lazily.

Potentially chunks=False in open_dataset could indicate that you're OK loading everything into memory with NumPy. We would then have to choose between making the default use Dask or NumPy.

This comment has the full context: https://github.com/pydata/xarray/issues/1372#issuecomment-293748654. To repeat myself:

You might ask why this separate lazy compute machinery exists. The answer is that dask fails to optimize element-wise operations like (scale * array)[subset] -> scale * array[subset], which is a critical optimization for lazy decoding of large datasets.

See https://github.com/dask/dask/issues/746 for discussion and links to PRs about this. jcrist had a solution that worked, but it slowed down every dask array operations by 20%, which wasn't a great win.

I wonder if this is worth revisiting with a simpler, less general optimization pass that doesn't bother with broadcasting. See the subclasses of NDArrayMixin in xarray/conventions.py for examples of the sorts of functionality we need: - Casting (e.g., array.astype(bool)). - Chained arithmetic with scalars (e.g., 0.5 + 0.5 * array). - Custom element-wise operations (e.g., map_blocks(convert_to_datetime64, array, dtype=np.datetime64)) - Custom aggregations that drop a dimension (e.g., map_blocks(characters_to_string, array, drop_axis=-1))

If we could optimize all these operations (and ideally chain them), then we could drop all the lazy loading stuff from xarray in favor of dask, which would be a real win.

The downside of this switch is that lazy loading of data from disk would now require dask, which would be at least slightly annoying to some users. But it's probably worth the tradeoff from a maintainability perspective, and also to fix issues like #1372.