KubeCon North America 2022 Product Ideas

[kannon92]

Batch Community Ideas

• Improve scalability of ETCD

  • Engage with ETCD groups about improving storage for large number of objects

• Start discussions about standardizing on a queue api definition for various CNCF projects

Armada Feature Ideas

• Convert Armada Queue Objects to use Kubernetes CRD
  • Allow use of Kubernetes API to manage Queue objects
  • Provide easier way for admins to manage queue objects
  • Queue currently does not have full CRUD ability
  • Design needed with GR team before development
• Convert Armada Job Definition to use Kubernetes Job Objects
  • Rational is to keep aligned with the momentum from K8-Batch group
• Investigate Yunikorn
  • https://yunikorn.apache.org/
  • Yunikorn seems way more lightweight than Volcano.
• Open Source armada namespaces
  • Internally, armada supports creation of queues with namespaces.
  • This work would be necessary for any production usage of Armada as users will probably not submit Jobs/Pods to the default namespace
  • Design is needed as we will not take the armada-namespaces code in GR and migrate to OSS.
• Support CRDs
  • Request from Bloomberg to support CRD (any general kubernetes object)
  • Initial ideas could be PyTorchJob, MPIJob, SparkJob, etc
• General design to incorporate arbitrary objects (MCAD actually provides a way to do this)
• Cloud aware and availability zone aware
  • One request was to only allow certain jobs to run in clusters in a certain AZ
  • A Banking customer said they had interest in running CronJobs accros multiple clusters but would only want certain jobs to run on cluster if it was legally allowed to run in that AZ
• Argo workflow as a workflow API
  • Albin suggested the idea of using the dag syntax for argo
  • We would build our own runner but we could support a CNCF dag language in our API so we can start supporting dags more directly
  • Currently GR has a myriad of dags spread across the entire repo
  • Argo has a proto api that definies their API so we could import that in our code.
• Google engineers mentioned the idea of providing manpower for support Kueue in Armada