A way to give scheduler hints on the size of unmanged memory?

[apiwat-chantawibul]

I'm a new user who just try to wrap some custom computation and map_blocks over some dask array.
The scheduler seems unable to take into account how much larger temporary memory is needed for the custom computation relative to its output size.
This results in scheduler attempting start too many tasks in parallel instead of finishing pending tasks which would reduce the memory foot print to the smaller size of the output.

To solve this I'm thinking of a couple of ways:

1. There must be a way to hint scheduler about the resources a task would require, right? (which is the title of this post)
2. Is there a way to set task priority to solve this problem?
3. Factoring the custom computation into smaller sub dask objects in hope that this will give scheduler more accurate view of the task. (this is a going to be hard for a new user like me + parts of the computation is scipy, it's not like I can really factor it.)

Please help