2-way failures

[dougbtv]

How do we handle 2-way failures, e.g. side A can exit 0 but, what if the agent doesn't complete? Doug's idea is to run as an init container & then if there's an exit 1 then the pod doesn't come up. (Just a brainstorm)

[Levovar]

I think we should first discuss whether we really need an agent, or not! What tasks exactly we want to push to the agent?
If we treat "making resources consumable to the application" as part of the CNI's responsibility, then maybe the CNI itself can do whatever is required.
I think at the end we gonna see that we only need to a handful of tasks, like mounting some PCI devices, and volumes containing a file.

In any case, if we end-up with an agent, I agree with the idea that Pod instantiation should fail. At the beginning of our DANM work we also made interface provisionin asynchronous, but most of the applications are not coded in a way that they a: check during startup if all resources they asked were really provisioned by the CaaS layer b: exit if something is missing
So IMO it is better to not even let the application start-up in case something is missing

[dougbtv]

I like the idea of it being agentless -- this was a quick note that I took during a (very very initial) demo that Billy gave for me.