the problem is that they want to apply a number of stateful/lookaside load balancing strategies, which become more difficult to do in a fully decentralized syst

the problem is that they want to apply a number of stateful/lookaside load balancing strategies, which become more difficult to do in a fully decentralized system. it’s generally easier to asynchronously aggregate information and either decide routing updates centrally or redistribute that aggregate to inform local decisions.