K8s Optimizer

Service: k8s-optimizer · Port: 8097 · Stack: Python, scikit-learn, FastAPI

k8s-optimizer uses machine learning to right-size workloads. It learns each workload's CPU and memory patterns, predicts where utilization is heading, and recommends — or, if you allow it, applies — replica changes before a workload becomes over- or under-provisioned.

How it works

1. Collect — gather per-workload CPU and memory usage from Prometheus and the Kubernetes API.
2. Learn — train a RandomForest model per workload and resource once it has enough history (100 samples by default), and retrain as new data arrives. Models are saved, so they survive restarts.
3. Predict — forecast each workload's utilization.
4. Recommend — propose a scale-up or scale-down when predicted utilization crosses your thresholds and the model is confident enough (0.7 by default).
5. Apply (optional) — outside dry-run mode, apply the change within your replica bounds.
6. Learn from outcomes — record whether each change brought utilization back into range, and use that feedback in future predictions.

Safety

• Dry-run mode (DRY_RUN=true) records recommendations without changing anything — start here to build trust.
• Hard bounds (MIN_REPLICAS, MAX_REPLICAS) cap every change, whatever the model says.
• Confidence gating — low-confidence predictions never produce recommendations.
• Applied changes go through the action agent, so they respect PodDisruptionBudgets and are audited.

REST API

All routes require a valid KubeOpera token.

Method	Path	Description
GET	/api/v1/status	Optimizer status and model readiness per workload.
GET	/api/v1/recommendations	Current recommendations with confidence.
GET	/api/v1/feedback	Outcomes of previously applied recommendations.
GET	/metrics	Prometheus metrics.
GET	/healthz	Health check.

Configuration

Variable	Default	Description
PROMETHEUS_URL	http://kube-prometheus-stack-prometheus.monitoring:9090	Prometheus.
NAMESPACES	all	Namespaces to optimize (comma-separated).
OPTIMIZATION_INTERVAL	300	Seconds between cycles.
CPU_SCALE_UP_THRESHOLD / CPU_SCALE_DOWN_THRESHOLD	0.8 / 0.3	CPU utilization that triggers a recommendation.
MEMORY_SCALE_UP_THRESHOLD / MEMORY_SCALE_DOWN_THRESHOLD	0.8 / 0.3	Memory utilization that triggers a recommendation.
MIN_CONFIDENCE_SCORE	0.7	Minimum model confidence.
MIN_TRAINING_SAMPLES	100	History needed before predictions start.
MIN_REPLICAS / MAX_REPLICAS	1 / 10	Replica bounds.
MODEL_STORE_PATH	/data/models	Where trained models are saved.
DRY_RUN	true	Recommend without applying.
AUTH_JWT_ACCESS_SECRET	—	Validates tokens.
PORT	8097	HTTP port.