Thinking about this from a diagnostics perspective, I wonder if the thing to track might instead be the expected size in bytes of each key prefix rather than the delta. My guess is that this will be more generally useful for a variety of applications.

We might then derive the delta information from this information on the fly for the specific use case of memory-aware scheduling.

Sometimes when thinking about scheduler state I find myself also being motivated by what would look good in a dashboard. This is a rather shallow objective on its own, but the information for good automated task scheduling can look surprisingly like the information for good human consumption.

Thoughts @TomAugspurger ?

I think tracking prefix-level memory usage, and then deriving the delta, was my initial path. It's been a little while, but IIRC that ran a problem with knowing whether completing a task actually freed the memory of a tasks dependencies.

        a1
       / |
      b1 |
       \ |
        c1

it's not clear (to me) how we would know that completing b isn't responsible for freeing a. We could probably track some kind of task_frees : Dict[Tuple[prefix_1, prefix_2]]: int that keeps a count of how often completing task with prefix_1 caused a release of a task with prefix_2? I'm not sure it's clearer (though it would likely be more useful for diagnostics)

I think I mentioned this earlier, but have since forgotten what the response was, but for any particular task we could probably include a fractional hit for every dependency based the number of its dependents.

So in the case above b would be half responsible for freeing a (assuming top-to-bottom execution) and so when we compute if a task is memory-freeing or not it would get half of the weight of a.

I'm not actively working on this at the moment. Closing to clear the backlog.