Karmada Documentation

The karmada-aggregated-apiserver is a component that starts an aggregated server. Its primary responsibilities are:

• Registering the Cluster API.
• Providing API aggregation capabilities, which allows users to access member clusters directly from the Karmada control plane.

The Thirdparty Resource Interpreter allows Karmada to handle custom resources from third-party applications and operators. Customizations are defined via YAML files and organized by API group, version, and kind.

Directory Structure:

resourcecustomizations/
├── <group>/
│   └── <version>/
│       └── <kind>/
│           ├── customizations.yaml                # Resource interpreter customization rules
│           └── testdata/
│               ├── <operation>-test.yaml          # Test cases for specific operations

Karmada supports priority and preemption for bindings (ResourceBinding and ClusterResourceBinding). This allows the scheduler to distinguish between high-priority critical tasks and non-urgent tasks during resource contention.

• Priority: A value assigned to a binding that determines its scheduling order. Priorities are independent of workload type and follow a total order.
• Binding Preemption: An action where the scheduler clears the scheduling results of lower-priority bindings to satisfy the requirements of a higher-priority pending binding when member clusters lack sufficient resources.

Note: Binding preemption is distinct from policy preemption, which involves high-priority policies preempting low-priority policies.

The Karmada Control Plane consists of three primary components:

• Karmada API Server: The REST endpoint for all components.
• Karmada Controller Manager: Performs operations based on API objects. It runs several specialized controllers:
  • Cluster Controller: Manages the lifecycle of member clusters.
  • Policy Controller: Watches PropagationPolicy objects to select resources and create ResourceBinding objects.
  • Binding Controller: Watches ResourceBinding objects to create Work objects.
  • Execution Controller: Watches Work objects to distribute resources to member clusters.
• Karmada Scheduler: Handles scheduling logic.

ETCD is used to store all Karmada API objects.

Version 1.13.0 introduced several significant features and optimizations:

• Application Priority Scheduling: Enhanced scheduling capabilities based on application priority.
• Resource Scheduling Suspend and Resume: Support for suspending and resuming resource scheduling.
• Karmada Operator Enhancements: Continuous improvements to the Karmada Operator.
• Controller Performance Optimization: Remarkable performance optimizations for Karmada controllers.
• Karmada Dashboard: Release of the first version of the official Karmada Dashboard.

The accurate estimator in Karmada calculates the maximum number of replicas (maxReplica) that can be scheduled on a cluster by combining multiple plugin results.

To account for resource fragmentation, the estimator follows these steps:

1. Plugin Execution: Runs maxReplicas calculation for each enabled plugin (e.g., quota-aware plugins).
2. Node Summation: Loops through all nodes and sums up the amount of replicas (sumReplicas) that can fit in each node.
3. Final Result: Returns Math.min(maxReplicas, sumReplicas).

This ensures that the estimation respects both cluster-wide quotas and the physical capacity of individual nodes.