My impression is that the secret sauce is in the allocation between machines - managing the allocation of high-CPU vs. high-IO vs. high-memory jobs, while still

My impression is that the secret sauce is in the allocation between machines - managing the allocation of high-CPU vs. high-IO vs. high-memory jobs, while still preserving things like rack locality and latency guarantees. A multidimensional knapsack problem, that changes over time.