I was thinking of things like: thread-per-client vs C10K. Just go C10K all the way. That means write services to be async, async, async. Just do it. It will not be the case that thread-per-client is better than C10K except when you know you'll only have a handful of clients. Write code to be thread-safe / reentrant. Write code to be separable into libraries.

I'd also urge you to use higher-level languages from the get-go. If you can use Rust instead of C, do it. If Haskell is appropriate, use it. These are all choices that are fantastically easier to make early on than later.