They don't. You can have fewer stores than workers. In the benchmark, we could in fact do with very few stores because they are not really used. I'm sure you could fiddle with the worker, load balancer and store count to get better performance (it depends on the system's requirements).

node is single threaded

I'm aware, and what's the relevance? I didn't even mention threads.

You can't use more than 1 core with a unique node process. So they spawn n cores for n processes. With a multithreaded runtime, this would not be required.

Yes exactly. My suggestion was the application to fork (CPUs - 1) workers, ie. 1 master 7 workers, instead of 1 master 3 workers and 3 stores, and have workers manage their key-value stores. Apparently each worker don't need a store, see author's comment (https://news.ycombinator.com/item?id=7713561) so it looks good.