Finding a compromise between reliability and scalability is tough. Especially since everyone will be running the same software but have different specs, internet connect and other factors. There is no easy answer to building a decentralized systems with so many varying factors so trial and error will be required. I do think there should some way to redirect traffic if a node has issues or goes down. For example if periodically we check how well a node can perform a certain set of actions and if it fails, you flag the node for review.