Another place to look for similar workflows is the HPC world. HPC schedulers have been doing resource requirement based scheduling for forever. Many of those those keep the scheduling state centralized though. Instead of asking the workers “who can fulfill these requirements”, the coordinator process keeps track of the state of what resources are still available for each worker. If you have a task that needs 12GB of RAM, 4 processors, and is in us-west, the coordinator already knows the answer. So instead of N^2 messages, you just have two (submit work and ack).

The major reason why HPC keeps this centrally is that it makes time based scheduling easier too. In general, you have more work than nodes and each job only has a finite run time. This lets the scheduler do fun things like backfilling jobs and more coordinated scheduling (such as keeping an entire rack idle for power savings unless necessary).