I was working on some sparse matrix datastructures which used dask.arrays internally. Then I realized that dask.array was basically to restrictive for my approach. Notes on this are here. They might be useful if y'all consider doing sparse matrix stuff in the future.