版本：v1.7 🚧

Custom Resource Definitions

文档说明

This document is generated with assistance from Qoder AI.

Introduction

Koordinator extends Kubernetes functionality through Custom Resource Definitions (CRDs) that enable advanced scheduling, resource management, and performance optimization. These CRDs provide declarative interfaces for configuring and managing various aspects of cluster behavior, including resource recommendations, colocation policies, quota management, reservations, and system-level objectives. This documentation provides comprehensive details on each CRD, including schema definitions, field descriptions, validation rules, and usage patterns.

API Groups and Versions

Koordinator's CRDs are organized into several API groups, each serving a specific purpose within the system architecture:

analysis.koordinator.sh: Provides resource optimization recommendations
config.koordinator.sh: Manages colocation profiles for workload placement
quota.koordinator.sh: Handles elastic quota profiles for resource allocation
scheduling.koordinator.sh: Manages resource reservations and scheduling policies
slo.koordinator.sh: Defines Service Level Objectives for node performance and resource quality

All CRDs are currently at version v1alpha1, indicating they are in early development and subject to change. The CRDs follow Kubernetes API conventions with proper metadata, spec, and status fields, and leverage controller patterns for reconciliation and lifecycle management.

Recommendation CRD

The Recommendation CRD in the analysis.koordinator.sh API group provides resource optimization suggestions for workloads based on historical usage patterns.

Schema Documentation

type Recommendation struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec   RecommendationSpec   `json:"spec,omitempty"`
    Status RecommendationStatus `json:"status,omitempty"`
}

Spec Fields

target: Defines the analysis target with the following subfields:
- type: Specifies the target type (workload or podSelector)
- workload: References a Kubernetes workload object when type is "workload"
- podSelector: Label selector for targeting specific pods when type is "podSelector"

Status Fields

podStatus: Contains resource recommendations with:
- containerStatuses: Array of container-specific recommendations
  - containerName: Name of the container
  - resources: Recommended resource amounts (CPU, memory)
updateTime: Timestamp of the last recommendation update
conditions: Lifecycle and status conditions

Validation Rules

The target field is required
Type must be either "workload" or "podSelector"
When type is "workload", the workload field must be specified
When type is "podSelector", the podSelector field must be specified

Default Values

No default values are explicitly defined in the schema.

Section sources

ClusterColocationProfile CRD

The ClusterColocationProfile CRD in the config.koordinator.sh API group defines policies for workload colocation and QoS configuration.

Schema Documentation

type ClusterColocationProfile struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec              ClusterColocationProfileSpec   `json:"spec,omitempty"`
    Status            ClusterColocationProfileStatus `json:"status,omitempty"`
}

Spec Fields

namespaceSelector: Label selector to filter namespaces (optional, defaults to all)
selector: Label selector to filter pods (optional, defaults to all)
probability: Probability (0-100%) that the profile will be applied
qosClass: QoS class to assign (LSE, LSR, LS, BE, SYSTEM)
priorityClassName: Priority class to inject into pods
koordinatorPriority: Sub-priority value for Koordinator scheduling
labels: Key-value pairs to inject as pod labels
annotations: Key-value pairs to inject as pod annotations
labelKeysMapping: Maps existing label keys to new keys
annotationKeysMapping: Maps existing annotation keys to new keys
labelSuffixes: Appends suffixes to existing label values
schedulerName: Scheduler to dispatch pods
patch: Raw extension for patching pod templates

Validation Rules

qosClass must be one of: LSE, LSR, LS, BE, SYSTEM
namespaceSelector and selector follow standard Kubernetes label selector validation
probability must be a valid integer or string percentage

Default Values

namespaceSelector and selector default to empty (matches all)
probability has no default (optional)

Section sources

ElasticQuotaProfile CRD

The ElasticQuotaProfile CRD in the quota.koordinator.sh API group manages quota profiles for flexible resource allocation.

Schema Documentation

type ElasticQuotaProfile struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec   ElasticQuotaProfileSpec   `json:"spec,omitempty"`
    Status ElasticQuotaProfileStatus `json:"status,omitempty"`
}

Spec Fields

quotaName: Required name of the associated quota
quotaLabels: Labels to apply to the quota
resourceRatio: Ratio to address resource fragmentation (e.g., "0.9")
nodeSelector: Required node selector to target specific nodes

Validation Rules

quotaName is required
nodeSelector is required
resourceRatio must be a valid string representation of a ratio

Default Values

No default values are explicitly defined.

Section sources

Reservation CRD

The Reservation CRD in the scheduling.koordinator.sh API group enables resource reservation for future workload placement.

Schema Documentation

type Reservation struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec   ReservationSpec   `json:"spec,omitempty"`
    Status ReservationStatus `json:"status,omitempty"`
}

Spec Fields

template: Pod template specifying resource requirements and scheduling constraints
owners: Array of reservation owners (object reference, controller reference, or label selector)
ttl: Time-to-live duration (default: 24h)
expires: Absolute expiration timestamp
preAllocation: Allows overcommitment when set
allocateOnce: Limits allocation to first successful owner (default: true)
allocatePolicy: Allocation policy (Aligned, Restricted)
unschedulable: Controls reservation schedulability
taints: Taints applied to the reservation

Status Fields

phase: Current phase (Pending, Available, Succeeded, Waiting, Failed)
conditions: Status conditions with type, status, reason, and message
currentOwners: Currently allocated owners
nodeName: Node where resources are reserved
allocatable: Total allocatable resources
allocated: Currently allocated resources

Validation Rules

template and owners are required
ttl and expires are mutually exclusive
allocatePolicy must be Aligned or Restricted
preAllocation and allocateOnce have specific interaction rules

Default Values

ttl defaults to "24h"
allocateOnce defaults to true
phase defaults to "Pending"

Section sources

NodeMetric CRD

The NodeMetric CRD in the slo.koordinator.sh API group collects and reports node-level resource metrics.

Schema Documentation

type NodeMetric struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec   NodeMetricSpec   `json:"spec,omitempty"`
    Status NodeMetricStatus `json:"status,omitempty"`
}

Spec Fields

metricCollectPolicy: Collection policy with:
- aggregateDurationSeconds: Aggregation period in seconds
- reportIntervalSeconds: Reporting interval in seconds
- nodeAggregatePolicy: Target grain for node aggregation
- nodeMemoryCollectPolicy: Memory collection method (usageWithHotPageCache, usageWithoutPageCache, usageWithPageCache)

Status Fields

updateTime: Last update timestamp
nodeMetric: Node resource usage metrics
- nodeUsage: Total node resource usage
- aggregatedNodeUsages: Historical usage aggregates
- systemUsage: System-level resource usage
- aggregatedSystemUsages: Historical system usage aggregates
podsMetric: Array of pod-level metrics
hostApplicationMetric: Metrics for host applications
prodReclaimableMetric: Reclaimable resource indicators

Validation Rules

nodeMemoryCollectPolicy must be one of the defined enum values
Duration fields must be valid time durations

Default Values

nodeMemoryCollectPolicy defaults to usageWithoutPageCache

Section sources

NodeSLO CRD

The NodeSLO CRD in the slo.koordinator.sh API group defines Service Level Objectives for node performance and resource quality.

Schema Documentation

type NodeSLO struct {
    metav1.TypeMeta   `json:",inline"`
    metav1.ObjectMeta `json:"metadata,omitempty"`
    Spec   NodeSLOSpec   `json:"spec,omitempty"`
    Status NodeSLOStatus `json:"status,omitempty"`
}

Spec Fields

resourceUsedThresholdWithBE: Resource threshold strategy for BE workloads
resourceQOSStrategy: QoS configuration strategy for different workload classes
cpuBurstStrategy: CPU burst configuration
systemStrategy: Node-level system configuration
extensions: Third-party extensions
hostApplications: QoS management for host applications

The resourceQOSStrategy supports detailed configuration for CPU, memory, block I/O, network, and resctrl QoS across different workload classes (LSR, LS, BE, SYSTEM).

Validation Rules

Various percentage fields have minimum and maximum constraints (0-100)
CPU and memory burst policies have specific value ranges
Network bandwidth values must be non-negative

Default Values

Many fields have sensible defaults for production environments, such as:

cpuSuppressThresholdPercent: 65
memoryEvictThresholdPercent: 70
cpuEvictTimeWindowSeconds: 300

Section sources

Usage Examples

Creating Resources with kubectl

# Create a Recommendation
kubectl apply -f recommendation.yaml

# Create a ClusterColocationProfile
kubectl apply -f cluster-colocation-profile.yaml

# Create an ElasticQuotaProfile
kubectl apply -f elastic-quota-profile.yaml

# Create a Reservation
kubectl apply -f reservation.yaml

# Create a NodeMetric
kubectl apply -f nodemetric.yaml

# Create a NodeSLO
kubectl apply -f nodeslo.yaml

Programmatic Creation with Go Client

// Example of creating a Reservation programmatically
reservation := &schedulingv1alpha1.Reservation{
    ObjectMeta: metav1.ObjectMeta{
        Name: "example-reservation",
    },
    Spec: schedulingv1alpha1.ReservationSpec{
        Template: &corev1.PodTemplateSpec{
            // Template specification
        },
        Owners: []schedulingv1alpha1.ReservationOwner{
            {
                LabelSelector: &metav1.LabelSelector{
                    MatchLabels: map[string]string{"app": "example"},
                },
            },
        },
        TTL: &metav1.Duration{Duration: 24 * time.Hour},
    },
}

createdReservation, err := client.SchedulingV1alpha1().Reservations().Create(context.TODO(), reservation, metav1.CreateOptions{})

Sample YAML Manifests

Recommendation

apiVersion: analysis.koordinator.sh/v1alpha1
kind: Recommendation
metadata:
  name: example-recommendation
spec:
  target:
    type: workload
    workload:
      kind: Deployment
      name: example-app
      apiVersion: apps/v1

ClusterColocationProfile

apiVersion: config.koordinator.sh/v1alpha1
kind: ClusterColocationProfile
metadata:
  name: high-priority
spec:
  selector:
    matchLabels:
      priority: high
  qosClass: LSE
  priorityClassName: system-cluster-critical

ElasticQuotaProfile

apiVersion: quota.koordinator.sh/v1alpha1
kind: ElasticQuotaProfile
metadata:
  name: gpu-quota-profile
spec:
  quotaName: gpu-quota
  nodeSelector:
    matchLabels:
      accelerator: gpu
  resourceRatio: "0.8"

Reservation

apiVersion: scheduling.koordinator.sh/v1alpha1
kind: Reservation
metadata:
  name: batch-job-reservation
spec:
  template:
    spec:
      containers:
      - name: placeholder
        resources:
          requests:
            cpu: "8"
            memory: "32Gi"
  owners:
  - labelSelector:
      matchLabels:
        job-type: batch
  ttl: "12h"

NodeMetric

apiVersion: slo.koordinator.sh/v1alpha1
kind: NodeMetric
metadata:
  name: node-sample
spec:
  metricCollectPolicy:
    aggregateDurationSeconds: 60
    reportIntervalSeconds: 15
    nodeMemoryCollectPolicy: usageWithoutPageCache

NodeSLO

apiVersion: slo.koordinator.sh/v1alpha1
kind: NodeSLO
metadata:
  name: production-node-slo
spec:
  resourceUsedThresholdWithBE:
    enable: true
    cpuSuppressThresholdPercent: 70
    memoryEvictThresholdPercent: 75
  resourceQOSStrategy:
    beClass:
      cpuQOS:
        enable: true
        groupIdentity: 1
      memoryQOS:
        enable: true
        minLimitPercent: 20
        lowLimitPercent: 40
        throttlingPercent: 80

Section sources

Controller Integration

Koordinator's CRDs integrate with the Kubernetes controller pattern through dedicated controllers that reconcile the desired state defined in the spec with the actual cluster state.

Controller Responsibilities

RecommendationController: Analyzes workload resource usage and generates optimization recommendations
ColocationProfileController: Applies colocation policies to matching pods during admission
QuotaProfileController: Manages elastic quota profiles and their application to workloads
ReservationController: Handles resource reservation lifecycle, scheduling, and allocation
NodeMetricController: Collects and aggregates node resource metrics
NodeSLOController: Enforces node-level SLOs and QoS policies

Reconciliation Process

Each controller follows the standard reconciliation loop:

Watches for CRD events (create, update, delete)
Fetches the current resource state
Compares spec (desired state) with status (current state)
Takes actions to reconcile differences
Updates status to reflect current state
Emits events and conditions for observability

Controllers use Kubernetes client-go libraries for API interactions and leverage informers for efficient event handling. They also implement proper error handling and retry mechanisms to ensure reliability.

Section sources

Common Configuration Pitfalls

Invalid Field Combinations

Setting both ttl and expires in Reservation (mutually exclusive)
Specifying conflicting QoS policies in NodeSLO
Using invalid enum values for qosClass or allocationPolicy

Missing Required Fields

Omitting quotaName or nodeSelector in ElasticQuotaProfile
Not specifying template or owners in Reservation
Forgetting to set required selectors in ClusterColocationProfile

Validation Errors

Exceeding percentage limits (e.g., cpuSuppressThresholdPercent > 100)
Using invalid duration formats in TTL fields
Providing malformed label selectors or match expressions

Best Practices

Use descriptive names and labels for easy identification
Set appropriate TTL values to avoid resource leaks
Test colocation profiles in staging environments before production
Monitor reservation utilization to optimize resource allocation
Regularly review and update NodeSLO configurations based on workload patterns

Section sources

Custom Resource Definitions

Introduction​

API Groups and Versions​

Recommendation CRD​

Schema Documentation​

Spec Fields​

Status Fields​

Validation Rules​

Default Values​

ClusterColocationProfile CRD​

Schema Documentation​

Spec Fields​

Validation Rules​

Default Values​

ElasticQuotaProfile CRD​

Schema Documentation​

Spec Fields​

Validation Rules​

Default Values​

Reservation CRD​

Schema Documentation​

Spec Fields​

Status Fields​

Validation Rules​

Default Values​

NodeMetric CRD​

Schema Documentation​

Spec Fields​

Status Fields​

Validation Rules​

Default Values​

NodeSLO CRD​

Schema Documentation​

Spec Fields​

Validation Rules​

Default Values​

Usage Examples​

Creating Resources with kubectl​

Programmatic Creation with Go Client​

Sample YAML Manifests​

Recommendation​

ClusterColocationProfile​

ElasticQuotaProfile​

Reservation​

NodeMetric​

NodeSLO​

Controller Integration​

Controller Responsibilities​

Reconciliation Process​

Common Configuration Pitfalls​

Invalid Field Combinations​

Missing Required Fields​

Validation Errors​

Best Practices​

Introduction

API Groups and Versions

Recommendation CRD

Schema Documentation

Spec Fields

Status Fields

Validation Rules

Default Values

ClusterColocationProfile CRD

Schema Documentation

Spec Fields

Validation Rules

Default Values

ElasticQuotaProfile CRD

Schema Documentation

Spec Fields

Validation Rules

Default Values

Reservation CRD

Schema Documentation

Spec Fields

Status Fields

Validation Rules

Default Values

NodeMetric CRD

Schema Documentation

Spec Fields

Status Fields

Validation Rules

Default Values

NodeSLO CRD

Schema Documentation

Spec Fields

Validation Rules

Default Values

Usage Examples

Creating Resources with kubectl

Programmatic Creation with Go Client

Sample YAML Manifests

Recommendation

ClusterColocationProfile

ElasticQuotaProfile

Reservation

NodeMetric

NodeSLO

Controller Integration

Controller Responsibilities

Reconciliation Process

Common Configuration Pitfalls

Invalid Field Combinations

Missing Required Fields

Validation Errors

Best Practices