Deep Dive into Kubernetes Scheduler: Principles, Algorithms, and Code Walkthrough

Introduction

This article examines the kube‑scheduler component deployed on the Kubernetes master node, explaining its operating mechanism and analyzing core source code (Kubernetes 1.9, commit 925c127ec). It covers the scheduler’s watch process, binding workflow, two‑stage algorithm (Predicates and Priorities), and detailed code paths.

Scheduler Basics

The scheduler runs as an independent process, watches the API server for Pods without a NodeName, and binds them to suitable Nodes. It receives Pods and Nodes as input, outputs the selected Node, and performs binding.

Algorithm Stages

Scheduling consists of Predicates (filtering unsuitable Nodes) and Priorities (scoring remaining Nodes from 0‑10, weighted sum determines the best Node). The highest‑scoring Node wins; ties are broken randomly.

Source Code Structure

Entry point: k8s.io/kubernetes/plugin/cmd/scheduler.go. Configuration in k8s.io/kubernetes/plugin/cmd/kube-scheduler/app/server.go. Core logic resides in k8s.io/kubernetes/plugin/pkg/scheduler. Key functions include NewSchedulerCommand, command.Execute, loadConfigFromFile, NewSchedulerServer, createClients, makeLeaderElectionConfig, makeHealthzServer, makeMetricsServer, SchedulerServer.Run.

Scheduler Execution Flow

Each scheduling cycle calls Scheduler.Run(), which syncs caches (WaitForCacheSync), fetches the next Pod (NextPod), and runs the algorithm (schedule). The schedule step uses nodeLister.List, findNodesThatFit, PrioritizeNodes, selectHost, then assumes the Pod and finally binds it via the API server.

Conclusion

The kube‑scheduler provides pod‑to‑node assignment on the master, maintains a FIFO (or optional PriorityQueue) PodQueue, and relies on the DefaultProvider’s Predicates and Priorities to select the optimal Node, updating the SchedulerCache and invoking the API server’s Bind operation.