Ubers Journey To Ray On Kubernetes Resource Management

Uber Engineering lets teams borrow unused computing capacity while preserving the owners' ability to reclaim it. Its extensions to Kubernetes divide cluster resources into team pools and recalculate each pool's allowance using demand, usage and available capacity. Borrowed capacity supports work that may be interrupted when another pool needs its share.

That tradeoff is explicit: workloads that cannot be interrupted must stay within their reserved capacity. The resource-management account describes how this lets teams run less critical experimental work on spare capacity alongside production work. It complements the job-placement account, which describes choosing a cluster for each job.