Expand Requirements Table

[mmccool]

Enumerate some additional potential requirements. These still need to be correlated with use cases that are still in flight. We may or may not want to adopt all of these, I am just trying to tease out some specifics.

[mmccool]

One thing currently missing that we probably want to discuss is whether the "work" is stateless. FaaS workloads are usually assumed to be stateless (pure functions) since this makes replicating and migrating them much easier. This makes sense for some tasks that have little or no state anyway, such as inferencing (as long as the neural net definition itself is not considered state, anyway, and as long as it's not actively being trained...). Other workloads, like game physics, may however need to retain a lot of state (e.g. the world model). Also tools that convert existing workloads (e.g. OpenMP -> Workers) will generate workloads that have state, usually. So we may want to add another requirement for "Stateful Workloads" that may only tie into some use cases.

Also some requirements may influence others. A requirement for stateful workload may affect how migration is done, or vice-versa. It's a lot easier to do live migration with little or no latency impact if the workload is stateless.

Finally, I think we probably want to eventually discuss priorities on use cases and requirements, but this should be driven by stakeholder interests. I want to reiterate that I think these are potential requirements, and we don't necessarily have to target all the use cases or the requirements, depending.

[mmccool]

Also upon re-reading the text I think there is an assumption that there is an cloud-like service providing edge computing, by a company. But I think the case of a single "edge computer" is also interesting, such as a lightweight laptop (or phone) offloading work to a personal desktop computer. That makes sense for something like VR (latency), or for an IoT device in the home doing video analytics (privacy).