AI
Dynamic Scoring: Metric-Driven Placement Across Clusters
Dynamic scoring deploys agents to managed clusters as add-ons. Those agents pull metrics from the cluster or from an external source and forward the data to a scoring API. OCM supplies the scaffolding for that API; operators supply the logic. The API evaluates metrics continuously and writes placement decisions back into OCM. A workload owner submits a workload to an orchestrator, which reads the current placement and dispatches to whichever clusters score highest at that moment.
Federated Learning Without High Egress Costs
The OCM hub acts as the aggregator for federated learning, cutting the egress cost of running training logic at the edge. A Flower add-on handles super-node lifecycle on each managed cluster. A federated learning CRD reconciles state into a manifest work, which means all the federated learning logic translates directly into OCM placement. The approach is framework-agnostic, supporting Flower, OpenFL, or others. Cluster claim-based scheduling keeps workload placement within cluster boundaries, removing the extra coordination overhead that normally comes with edge training.
Project Updates: Kueue, CAP-5313, and Incubation
OCM is aligning with SIGMC’s cluster profile API to become a credential source, making it easier for other projects to plug in. A new Kueue add-on simplifies batch learning and job management across multiple clusters. Add-ons hit v1 beta 1, marking them stable for production use. CAP-5313 introduces a placement decision API that decouples scheduling selection from workload dispatch, so operators interact with a clean interface rather than OCM’s internal scheduling logic. OCM has also applied for CNCF incubation and shipped version 1.3, up from the 1.0 release.
Notable Quotes
We’re using cluster claimbased scheduling. That is very helpful because it makes sure that the the um the workplacement is always kept uh inside the behind the cluster boundary. August Simonelli · ▶ 3:02
add-ons are now v1 beta 1. So, you know what that means? Basically, they’re stable. They’re ready to go. August Simonelli · ▶ 3:56
it helps decouple scheduling selection from workload dispatch, meaning it keeps a cleaner uh system, allowing you not to have to worry about the internal logic of scheduling in something like open cluster management, but instead just throw your work at it. August Simonelli · ▶ 4:13
Key Takeaways
- Dynamic scoring continuously rewrites placement decisions from live cluster metrics.
- The federated learning add-on uses OCM placement to keep training data within cluster boundaries.
- Add-ons reached v1 beta 1 and OCM version 1.3 is now available.
About the Speaker(s)
August Simonelli is a Principal Product Manager at Red Hat. He has worked with customers around the world to help them adopt, use, improve, and implement open source technologies. Raised in Boulder, Colorado, August now lives in Sydney, Australia and is a strong advocate for using open source software.