Kubernetes Cluster:
- Nodes (also called workers) are the physical or virtual machines that make up the Kubernetes cluster. These are provisioned in CDK with
cluster.addNodeCapacity().- runs the container runtime (Docker)
- runs kubelet (agent that communicates with the control plane)
- runs kube-proxy (network proxy)
- Pods are the smallest deployable unit. Single instance of a running process in your cluster. These are created using
kubectl apply(kubectldoes not create nodes, but it can create other custom resources like Ray.)- pods run on nodes (a Ray cluster “head-node” is a pod)
- pods can contain one or more containers, which share the same network namespace
- pods are ephemeral (created, destroyed, and rescheduled based on the cluster’s needs)
- pods are scheduled onto nodes by the Kubernetes scheduler
- Namespaces are cluster-wide resources.
kubectl get namespace- do not directly affect or partition nodes. All namespaces can potentially use any node in the cluster
- Pods must be created within a namespace. Pods from different namespaces can coexist on the same node.
- for organising and isolating resources in multi-tenant environments
'default','kube-system','kube-ray'
- scaling
- horizontal pod autoscaling: increasing or decreasing the number of pod replicas
- cluster autoscaling: adding or removing nodes based on resource demands
- blueprints/templates/manifests are yaml/json configuration files that describe how a Kubernetes application and its associated resources should be deployed and managed
- Helm charts are packages of pre-configured Kubernetes resources (package management)
- Control Plane are components that manage the cluster: kube-api-server, kube-scheduler, kube-controller-manager, etcd (the cluster’s database, key value store for all API server data)