Yeah, SDI doesn't do a very good job of actually solving that chicken and egg problem. It tends to give a "solution" without motivating the design decisions.

After reading about 80% of SDI and not much else, I failed a system design interview when I provided the "right" answer of putting an async task on a messaging queue, but couldn't explain how at-least-once delivery could be used to deal with a flaky task.

I think I've filled a lot of these gaps now (passed L5 system design at Google for example) but there wasn't really a single efficient resource that helped me do so. DDIA is awesome though.