@neilsong yup -- one thing i didn't explicitly mention but is actually pivotal to the design is that we switched sandbox scheduling to be synchronous w.r.t. sandbox creation (was previously async), which then allows us to remove state which could have been a scaling bottleneck otherwise