Koordinator is a Kubernetes scheduling system for colocating latency-sensitive online services with batch and AI workloads on the same nodes. It was open-sourced by Alibaba out of their internal experience running mixed workloads at very high cluster utilization.
The core idea is QoS-aware overcommit. Koordinator labels pods with priority classes (product, mid, batch, free) and uses its koordlet node agent to continuously measure real CPU/memory headroom — not requests — and publish a “reclaimed” resource back to the scheduler. Batch pods then bin-pack into reclaimed capacity and get throttled or evicted under CPU/memory pressure before they can hurt the online tier. Isolation is enforced through cgroup tuning, CPU suppression, and memory QoS features in the kernel. It ships a scheduler built as a set of kube-scheduler plugins plus descheduler components that rebalance hot nodes.
It is the closest open-source analog to Google’s Borg-style colocation, and overlaps with Volcano (batch scheduling) and Crane (cost/efficiency). Its niche is specifically wringing 60-70% cluster utilization out of hardware that would otherwise sit at 20% because the online services refuse to share nodes.
