I had a lucky break in the day and was able to read this blog post which popped up in my social feed. In essence it talks about what Fly.io did to rebuild their scheduler to better match what they’re trying to accomplish. Orchestration and scheduling are topics I like to geek out on, going back many years as part of the Jenkins project. But this quote in particular caught my eye:

flyd has a radically different model from Kubernetes and Nomad. Mainstream orchestrators are like sophisticated memory allocators, operating from a reliable global picture of all capacity everywhere in the cluster. Not flyd.

Instead, flyd operates like a market. Requests to schedule jobs are bids for resources; workers are suppliers. Our orchestrator sits in the middle like an exchange. ratemysandwich.com asks for a Fly Machine with 4 dedicated CPU cores in Chennai (sandwich: bun kebab?). Some worker in MAA offers room; a match is made, the order is filled.

I love this idea for a lot of reasons, not the least of which is that it’s a real-world incantation of an unoriginal idea that I had for Otto, an overly ambitious CI/CD side-project.

In my work I referred to it as resource allocation by auction and had only just begun to experiment with the concept. I once read a computer science paper which described this concept more in detail, but I cannot seem to find it again.

Suffice it to say, there’s a lot of good efficiency to be gained by resource auctioning in this manner, especially in a multi-tenant system. The Fly.io blog post is an interesting read either way, but efficient resource scheduling in this way I hope makes it into a lot of other systems.