Skip to content

Latest commit

 

History

History
36 lines (29 loc) · 1.9 KB

ROADMAP.md

File metadata and controls

36 lines (29 loc) · 1.9 KB

KubeFleet Roadmap

Project Website

  • Setup the project website

Support more cluster properties so that user can pick the right cluster for their workload

  • Support node level SKU as properties, e.g. CPU, GPU, Memory, etc
    • The application admin can choose clusters that have nodes with H100 GPU.
    • The application admin can choose clusters that have nodes with 128GB memory.
  • Support network topology
    • The application admin can choose the clusters with requires infiniband, or 100Gbps network.

Support scheduling for namespaced resources (heterogeneous namespace)

  • Support independent scheduling policy for namespaced resources
    • e.g. The application admin can pick one workload in a namespace to cluster A while the other workload in the same namespace to cluster B.

Dynamic scheduling

  • De-scheduler for the fleet
    • The de-scheduler would move the workload to the right cluster if the cluster is not the best fit for the workload anymore.
  • Cordon a cluster
    • The fleet admin can cordon a cluster to move all the workloads off the cluster.
  • Rebalance the workload
    • The application admin can rebalance the workload to make sure the workload is spread evenly across the clusters.

Support anti-affinity for workload

  • Support affinity/anti-affinity for their workload.
    • The application admin can specify that their workload A needs to be placed on the same clusters that workload B runs.
    • The application admin can specify that their workload A cannot be placed on the same clusters that workload B runs.

Support Customized health check for workload

  • Support user specified health check for their workload.
    • The application admin can provide a customized health check for their workload.

Support Spread mode for workload

  • The application admin can specify a spread mode for their workload.
    • The move between clusters would follow the max-unavailable/min-available pods rule.