I am using map_overlap and noticed the graph contains floating point keys. I saw this while troubleshooting performance, and I'm not sure if it's related. I thought perhaps the floating point might induce some sort lookup or cache miss. (I think I'm seeing a lot more tasks being computed in my scheduler status window than are reported being done.)

It's simple to see in the following with a small array. And if I make the array dimensionality larger, things get big very fast. But it's the floating points that worry me more than graph size (I understand overlaps are expensive).

# Using dask==2.9.2
import dask.array as da
da.zeros((10,)).map_overlap(lambda x: x, depth=1).visualize()

With shape 10:

With shape 10,10