Consider building the fs layers on the broker instead of having the worker do it.

[nfachan]

We're in a bit of a weird place right now.

The broker holds a reference count on all digests used by a job. This means it needs to read manifests to see what digests they use. It does this work, and caches the result.

Then, it sends jobs to workers. The workers then read the manifests and build their own fs layers, hopefully finding cached versions. This seems like a kind of weird and arbitrary separation of the two.

Instead, one could imagine the broker building the fs layers and sharing them with the workers. This way the fs layers would only be built once (not on every broker), and could be cached across broker invocations.

Alternatively, if we don't want the broker to actually be doing this work, we could have a mechanism where it could ask a single worker to read the manifests and build the fs layers, then share the result back. Then, the worker could re-distribute that data with the other workers.