Karmada Scheduler Edge Connect

Combined Karmada scheduler that replaces the stock karmada-scheduler with three additional plugins compiled in:

Plugin	Type	Source	What it does
CarbonScore	Filter + Score	karmada-carbon-scheduler	Filters/scores clusters by carbon intensity and renewable energy share via ClusterCarbonMetrics CRD
PredictionFilter	Filter	karmada-prediction-scheduler	Hard-rejects clusters with catastrophic predicted pressure (>95%) or extreme hotspots
PredictionScore	Score	karmada-prediction-scheduler	Ranks clusters by predicted future capacity with optional estimator fusion

All stock Karmada scheduler plugins remain active — these three are additive.

Full documentation lives under docs/. Start with docs/dependencies.md for what each plugin needs to actually do useful work (CRDs, collectors, prediction pipeline), then docs/getting-started/quickstart.md for installation and docs/how-to/enable-plugins-staged.md for the safe rollout sequence.

Why?

Karmada does not support chaining multiple scheduler binaries. A ResourceBinding is handled by exactly one scheduler. Running carbon and prediction as separate schedulers means workloads get one or the other, never both.

This repo compiles all plugins into a single binary that replaces the stock karmada-scheduler deployment so every workload benefits from all scheduling signals.

Architecture

┌─────────────────────────────────────────────────────┐
│              karmada-scheduler-edge-connect          │
│                                                     │
│  ┌─────────────┐  ┌──────────────────┐              │
│  │ Filter phase │  │ Score phase      │              │
│  │              │  │                  │              │
│  │ • built-ins  │  │ • built-ins      │              │
│  │ • CarbonScore│  │ • CarbonScore    │              │
│  │ • Prediction │  │ • PredictionScore│              │
│  │   Filter     │  │                  │              │
│  └──────┬───────┘  └────────┬─────────┘              │
│         │ AND               │ SUM                    │
│         ▼                   ▼                        │
│    candidates          ranked list                   │
└─────────────────────────────────────────────────────┘
         ▲                    ▲
         │                    │
  ClusterCarbonMetrics  ClusterPredictionState
  (carbon-data-collector)  (prediction-service)

Prerequisites

Both CRDs must be installed in the Karmada control plane:

• ClusterCarbonMetrics (API group carbon.edge-connect.eu)
• ClusterPredictionState (API group prediction.edge-connect.eu)

The plugins degrade gracefully if their CRDs are absent — they log a warning and allow all clusters through (fail-open).

Development

# Build locally
make build-local

# Run tests
make test

# Build Docker image (run from repo root — Dockerfile uses sibling dirs)
make docker-build

The go.mod uses replace directives pointing to sibling directories:

replace (
    github.com/telekom-8ra/karmada-carbon-scheduler => ../karmada-carbon-scheduler
    karmada-prediction-scheduler => ../karmada-prediction-scheduler
)

Ensure both repos are checked out alongside this one under ../.

Deployment

helm install karmada-scheduler chart/ \
  --namespace karmada-system \
  --set image.repository=<registry>/karmada-scheduler-edge-connect \
  --set image.tag=v0.1.0 \
  --set kubeconfig.existingSecret=karmada-kubeconfig

This deployment replaces the stock karmada-scheduler. Delete or scale down the original deployment first.

Configuration

CarbonScore plugin (env vars)

Variable	Default	Description
CARBON_WEIGHT	0.7	Weight for carbon intensity in score calculation
RENEWABLE_WEIGHT	0.3	Weight for renewable share in score calculation
MAX_CARBON_INTENSITY	200	Max gCO₂/kWh threshold for filter (0=disabled)
MIN_RENEWABLE_SHARE	0.0	Min renewable % threshold for filter (0=disabled)

PredictionScore plugin (env vars)

Variable	Default	Description
PREDICTION_SCORE_MODE	prediction_only	prediction_only or estimator_fusion
PREDICTION_SCORE_WEIGHT	30	Weight for prediction score
PREDICTION_HORIZON	10m	Prediction horizon (10m, 30m, 60m)
PREDICTION_ESTIMATOR_TIMEOUT	500ms	RPC timeout for estimator calls

Scheduler flags

Flag	Default	Description
--plugins	*,CarbonScore,PredictionFilter,PredictionScore	Enable all defaults + custom plugins
--enable-scheduler-estimator	true	Enable Karmada scheduler estimator
--leader-elect	true	HA leader election

License

Apache-2.0 — Copyright 2026 Deutsche Telekom AG / IPCEI-CIS