A note to future folks who find this thread: I was running into issues with one of my datasets where my map_blocks code was really slow, and after a lot of debugging, I wasn’t seeing obvious improvements over running in serial. But then I applied the same code to a dataset where (1) each chunk size was smaller (first dataset had chunksizes ~100 MB, second dataset had chunksizes ~2 MB, (2) each chunk had a size of 1 in the time dimension (as opposed to a size of 2 in the time dimension for the first dataset), and (3) I was just plotting a lot less data (<1 MB of data in each figure vs. ~100 MB of data in each figure). If your code is unexpectedly slow, consider exploring these factors in greater detail.