Pods are not assigned to any node when they are first created; it is the job of the scheduler to assign the pod to a node. A scheduler watches the apiServer for pods that have no value in the
spec.nodeName field. The scheduler finds the most suitable node for a pod-based on the podSpec and node statistics. After finding a suitable node, the scheduler sends an event to the apiServer, assigning the pod to the node.
The kubelet watches the apiServer for pods are assigned to their node but not running. The kubelet detects a pod assign to its node, obtains the podSpec and executes the containers using its container runtime.
The scheduler sends an error to the apiServer if it is unable to find a suitable node.
No scheduling occurs if
spec.nodeName has a value specified in a pod manifest. This behaviour occurs because the manifest has performed the assignment already.
No scheduling occurs for static pods; those where the manifest file is located on a node (
/etc/kubenetes/manifests by default). The local kubelet process manages static pods; not the apiServer.
The default Kubernetes scheduler uses a two-step process to assign pods to nodes.
Filtering: Filter out nodes which are not suitable, nodes that remain are
feasible node. Filtering criteria examples:
- Taints and tolerances
- Available resources
Scoring: Feasible nodes are scored against scoring rules, the pod is assigned to the node with the highest score.
Configuring the default scheduler
The default scheduler has many arguments available to fine-tune its configuration; these options can are found here. Options are configured as flags pushed to the scheduler pod or in the service file.
A Kubernetes cluster can have multiple schedulers; additional schedules are implemented in a similar fas
You need to specify which scheduler a Kubernetes is to use for a pod using the
spec.schedulerName field. If this is omitted, the default scheduler is used.
I haven’t included examples in this post because I don’t think I have anything to examples in the links below.
Viewing scheduler events
The cluster troubleshooting guide states that the default log file location is
/var/log/kube-scheduler.log, which is partly true. The Kubernetes installation method used influences where log files are stored if they are stored at all.
You should always ensure that deployment scripts specify log methods and paths. You can find information about configuring master service logs here.