PodCliqueSet CRD and Controller for Network Topology Aware Gang Scheduling & Autoscaling

👷 This project site is currently under active construction

Grove's PodClique Scaling Groups provide sophisticated gang scheduling and termination protection through two key configuration parameters: replicas and minAvailable.

Scaling Groups allow you to group multiple PodCliques together and scale them as a unit while maintaining gang scheduling semantics. This is particularly useful for distributed workloads that require coordinated scheduling and graceful scaling behavior.

• Purpose: Sets the desired number of replicas for the scaling group
• Default: 1 if not specified
• Behavior: Controls how many instances of the scaling group are created

• Purpose: Defines the minimum number of ready replicas required for operational stability
• Default: 1 if not specified
• Behavior: Enables gang scheduling and controls termination policies

Grove implements a sophisticated two-tier gang scheduling system based on the minAvailable setting:

• Replicas: 0 through (minAvailable - 1)
• Scheduling: All pods scheduled together as a single gang
• Purpose: Establishes the minimum viable cluster
• Gates Removed: Immediately when pods are assigned to the PodGang

• Replicas: minAvailable and above
• Scheduling: Each replica gets its own scaled PodGang
• Purpose: Provides additional capacity once core functionality is established
• Gates Removed: Only after the base PodGang is ready and running

apiVersion: grove.io/v1alpha1
kind: PodCliqueSet
spec:
  template:
    podCliqueScalingGroupConfigs:
    - name: "database-cluster"
      replicas: 5
      minAvailable: 3
      cliqueNames: ["db-primary", "db-secondary"]

Behavior:

• Replicas 0, 1, 2: Form base PodGang, scheduled together (minimum viable cluster)
• Replicas 3, 4: Scaled PodGangs, wait for base cluster to be ready
• Result: Ensures core 3-node cluster is operational before adding scale-out nodes

apiVersion: grove.io/v1alpha1  
kind: PodCliqueSet
spec:
  template:
    podCliqueScalingGroupConfigs:
    - name: "ml-training"
      replicas: 8
      minAvailable: 4
      cliqueNames: ["parameter-server", "worker"]

Behavior:

• Replicas 0, 1, 2, 3: Base PodGang for core training infrastructure
• Replicas 4, 5, 6, 7: Scaled PodGangs for additional training capacity
• Result: Core training cluster established before scaling out workers

A scaling group replica is considered "ready" when:

• Its associated PodClique has sufficient ready Pods: PodClique.Status.ReadyReplicas >= PodClique.Spec.MinReplicas
• A Pod is considered ready when its PodReady condition is True

If the number of ready replicas falls below minAvailable:

• Gang termination is triggered for the affected scaling group replica
• Purpose: Prevents resource waste and maintains workload integrity
• Behavior: Ensures workloads fail fast rather than running in degraded states

• Core functionality established first before scaling out
• Prevents wasteful scheduling of non-essential replicas

• Gang scheduling ensures all-or-nothing scheduling semantics
• Termination protection maintains minimum viable cluster size

• Base cluster provides stable foundation
• Scale-out replicas add capacity without disrupting core functionality