Crane-scheduler: Kubernetes Scheduling Optimization Based on Real Workload

Crane-scheduler addresses the gap between Kubernetes' resource requests and actual usage, which causes uneven cluster load. The article explains how this leads to resource waste and instability, then introduces Crane-scheduler as a solution that uses real-time node metrics for scheduling.

Architecture: Crane-scheduler relies on Node-exporter and Prometheus to collect and aggregate node metrics. It consists of Scheduler-Controller (periodically pulls metrics and annotates nodes) and Scheduler (uses annotations for filtering and scoring).

Scheduling Strategies: Users can configure predicates (threshold-based filtering) and priorities (weighted scoring based on metrics). The system also includes a hot-value mechanism to prevent scheduling hotspots.

Scheduler-Controller periodically pulls metrics from Prometheus and annotates nodes with CPU/memory usage averages and maximums.

Predicates filter nodes where metrics exceed thresholds (e.g., CPU usage >65%).

Priorities assign weights to metrics (e.g., CPU usage max avg 1h: weight 0.5) and calculate scores.

Hot-value tracks recent scheduling frequency to adjust scores.

Benefits: Balances load, reduces waste, and maintains stability. Case studies show improved resource utilization in public and internal clusters.