K8s Optimizer

Repo: python/k8s-optimizer · Stack: Python 3.11+, scikit-learn, FastAPI

The only Python service in the platform. A RandomForest-based resource-rightsizing agent: it collects CPU/memory metrics from Prometheus and the Kubernetes API, trains an in-memory model per cluster, and recommends (or, outside dry-run, applies) scale-up/scale-down changes.

Key Capabilities

• Metrics collection — MetricsCollector polls Prometheus and the Kubernetes API for per-workload resource usage
• AI-driven predictions — AIOptimizer trains a RandomForest model per resource dimension (CPU, memory) once MIN_TRAINING_SAMPLES (default 100) samples are collected; retrains every 1,000 new samples. Models are not persisted — they re-warm from scratch on every process restart
• Threshold-gated recommendations — a recommendation is only generated when predicted utilization crosses the configured scale-up/scale-down thresholds and the model's confidence exceeds MIN_CONFIDENCE_SCORE (default 0.7)
• Dry-run mode — DRY_RUN=true (or --dry-run) logs the patch body instead of calling the Kubernetes API; useful for validating recommendations before trusting the agent to act
• Hard replica bounds — MIN_REPLICAS/MAX_REPLICAS cap what the optimizer will ever set, regardless of model output

Auth

None. Unlike the Go services in o-apps, this service has no auth middleware of any kind on its FastAPI routes — confirmed by reading api/server.py directly. Anyone who can reach it can read status/recommendations/feedback. Not yet part of the platform's auth-triage effort, since it lives outside the o-apps monorepo.

REST API

Method	Path	Description
GET	/api/v1/status	Optimizer run status
GET	/api/v1/recommendations	Current recommendations
GET	/api/v1/feedback	Outcome feedback on previously-applied recommendations
GET	/api/v1/metrics	Prometheus-format metrics export
GET	/health	Health check

Environment Variables

Variable	Default	Purpose
PROMETHEUS_URL	http://prometheus-server:9090	Prometheus endpoint
NAMESPACE	default	Target namespace
OPTIMIZATION_INTERVAL	300	Seconds between optimization cycles
CPU_SCALE_UP_THRESHOLD / CPU_SCALE_DOWN_THRESHOLD	0.8 / 0.3	CPU utilization ratios that trigger a recommendation
MEMORY_SCALE_UP_THRESHOLD / MEMORY_SCALE_DOWN_THRESHOLD	0.8 / 0.3	Memory utilization ratios
MIN_CONFIDENCE_SCORE	0.7	Minimum model confidence before a recommendation is generated
MIN_TRAINING_SAMPLES	100	Samples required before the AI model activates
MAX_REPLICAS / MIN_REPLICAS	10 / 1	Hard scaling bounds
KUBECONFIG	—	Path to kubeconfig (omit for in-cluster)
DRY_RUN	false	Log recommendations without applying them