What should happen when local compute isn't enough?

IN PROGRESS

Local infrastructure has a ceiling, whether that ceiling is memory, accelerator availability, or just raw throughput. The interesting part isn’t that the ceiling exists — everyone building local-first systems knows it does — it’s what the system should do in the moment it’s reached.

The naive answer is “fail and tell the user to try something smaller.” The better answer, and the one we’re actually working through, is some form of graceful overflow: recognizing that a task has exceeded what local resources can handle and routing it somewhere else without the user having to notice or intervene. That requires the system to know, ahead of time, what’s safe to send externally and what isn’t — which loops directly back into the policy boundary problem rather than being a separate concern.

We’re currently working through what that decision point actually needs to know: is it purely a resource check (not enough memory, not enough throughput), or does it also need to understand something about the task itself (is this task allowed to leave local infrastructure at all, regardless of resource pressure)? Early indications are that it’s both, and that treating it as only a resource question will eventually send something external that should have stayed local.

No conclusions yet. This note exists to mark that the question is being worked, not to report an answer.