Currently the Scheduler state is aware of a hierarchy of computation elements

• TaskState
• TaskGroup (often equivalent to a layer in a client-side graph)
• TaskPrefix (a collection of layers that we think have similar computational characteristics

All of these elements are useful in scheduling heuristics and visual diagnostics.

We might consider adding a fourth class to this system, the Computation. This would be the result of a single compute/persist/map call, and would include a collection of the groups involved, as well as the context of the call (perhaps using the same frame-grabbing tricks that we use in the performance reports to show the code in the first tab).

This would give us additional context that we could use in constructing diagnostics to give back to the user. For example, we might be able to generate a performance_report on the fly for every compute/persist call that they've run during their session.