The Kubernetes Scheduler operator manages and updates the Kubernetes Scheduler deployed on top of OpenShift. The operator is based on OpenShift library-go framework and it is installed via Cluster Version Operator (CVO).

It contains the following components:

• Operator
• Bootstrap manifest renderer
• Installer based on static pods
• Configuration observer

By default, the operator exposes Prometheus metrics via metrics service. The metrics are collected from following components:

• Kubernetes Scheduler operator

The configuration for the Kubernetes Scheduler is the result of merging:

• a default config
• an observed config (compare observed values above) from the spec schedulers.config.openshift.io.

All of these are sparse configurations, i.e. unvalidated json snippets which are merged in order to form a valid configuration at the end.

The following profiles are currently provided:

• HighNodeUtilization
• LowNodeUtilization
• NoScoring

Each of these enables cluster-wide scheduling. Configured via Scheduler custom resource:

$ oc get scheduler cluster -o yaml

apiVersion: config.openshift.io/v1
kind: Scheduler
metadata:
  name: cluster
spec:
  mastersSchedulable: false
  policy:
    name: ""
  profile: LowNodeUtilization
  ...

This profile disables NodeResourcesBalancedAllocation and NodeResourcesFit plugin with (LeastAllocated type) and enables NodeResourcesFit plugin (with MostAllocated type). Favoring nodes that have a high allocation of resources. In the past the profile corresponded to disabling NodeResourcesLeastAllocated and NodeResourcesBalancedAllocation plugins and enabling NodeResourcesMostAllocated plugin.

The default list of scheduling profiles as provided by the kube-scheduler.

This profiles disabled all scoring plugins.

Customizations of existing profiles are available under the .spec.profileCustomizations field:

E.g.

apiVersion: config.openshift.io/v1
kind: Scheduler
metadata:
  name: cluster
spec:
  mastersSchedulable: false
  policy:
    name: ""
  profile: HighNodeUtilization
  profileCustomizations:
    dynamicResourceAllocation: Enabled
  ...

Operator also expose events that can help debugging issues. To get operator events, run following command:

$ oc get events -n  openshift-cluster-kube-scheduler-operator

This operator is configured via KubeScheduler custom resource:

$ oc describe kubescheduler

apiVersion: operator.openshift.io/v1
kind: KubeScheduler
metadata:
  name: cluster
spec:
  managementState: Managed
  ...

The log level of individual kube-scheduler instances can be increased by setting .spec.logLevel field:

$ oc explain kubescheduler.spec.logLevel
KIND:     KubeScheduler
VERSION:  operator.openshift.io/v1

FIELD:    logLevel <string>

DESCRIPTION:
     logLevel is an intent based logging for an overall component. It does not
     give fine grained control, but it is a simple way to manage coarse grained
     logging choices that operators have to interpret for their operands. Valid
     values are: "Normal", "Debug", "Trace", "TraceAll". Defaults to "Normal".

For example:

apiVersion: operator.openshift.io/v1
kind: KubeScheduler
metadata:
  name: cluster
spec:
  logLevel: Debug
  ...

Currently the log levels correspond to:

More about the individual configuration options can be learnt by invoking oc explain:

$ oc explain kubescheduler

The current operator status is reported using the ClusterOperator resource. To get the current status you can run follow command:

$ oc get clusteroperator/kube-scheduler

In the running cluster cluster-version-operator is responsible for maintaining functioning and non-altered elements. In that case to be able to use custom operator image one has to perform one of these operations:

1. Set your operator in umanaged state, see here for details, in short:

oc patch clusterversion/version --type='merge' -p "$(cat <<- EOF
spec:
  overrides:
  - group: apps
    kind: Deployment
    name: kube-scheduler-operator
    namespace: openshift-kube-scheduler-operator
    unmanaged: true
EOF
)"

2. Scale down cluster-version-operator:

oc scale --replicas=0 deploy/cluster-version-operator -n openshift-cluster-version

IMPORTANT: This apprach disables cluster-version-operator completly, whereas previous only tells it to not manage a kube-scheduler-operator!

After doing this you can now change the image of the operator to the desired one:

oc patch pod/openshift-kube-scheduler-operator-<rand_digits> -n openshift-kube-scheduler-operator -p '{"spec":{"containers":[{"name":"kube-scheduler-operator-container","image":"<user>/cluster-kube-scheduler-operator"}]}}'

The operator image version used by the installer bootstrap phase can be overridden by creating a custom origin-release image pointing to the developer's operator :latest image:

$ IMAGE_ORG=<user> make images
$ docker push <user>/origin-cluster-kube-scheduler-operator

$ cd ../cluster-kube-apiserver-operator
$ IMAGES=cluster-kube-scheduler-operator IMAGE_ORG=<user> make origin-release
$ docker push <user>/origin-release:latest

$ cd ../installer
$ OPENSHIFT_INSTALL_RELEASE_IMAGE_OVERRIDE=docker.io/<user>/origin-release:latest bin/openshift-install cluster ...

By default the kube-scheduler profiling is disabled. The profiling can be enabled manually by editing config.yaml files under each master node.

Warning: the configuration gets undone after the new revision gets performed and the steps need to be repeated.

Steps:

1. access every master node (e.g. ssh or with oc debug)

   1. edit /etc/kubernetes/static-pod-resources/kube-scheduler-pod-$REV/configmaps/config/config.yaml (where $REV corresponds to the latest revision) and set enableProfiling field to True.
   2. make a benign change to /etc/kubernetes/manifests/kube-scheduler-pod.yaml, e.g. updating "Waiting for port" to "Waiting for port" (adding one blank space to the string). Wait for the updated pod manifest to be picked up and a new kube-scheduler instance running and ready.

2. oc port-forward pod/$KUBE_SCHEDULER_POD_NAME 10259:10259 in a separate terminal/window (where $KUBE_SCHEDULER_POD_NAME corresponds to a running kube-scheduler pod instance)

3. apply the following manifests to allow anonymous access:

   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRole
   metadata:
    name: kubescheduler-anonymous-access
   rules:
   - nonResourceURLs: ["/debug", "/debug/*"]
     verbs:
     - get
     - list
   ---
   apiVersion: rbac.authorization.k8s.io/v1
   kind: ClusterRoleBinding
   metadata:
    name: kubescheduler-anonymous-access
   roleRef:
    apiGroup: rbac.authorization.k8s.io
    kind: ClusterRole
    name: kubescheduler-anonymous-access
   subjects:
   - apiGroup: rbac.authorization.k8s.io
     kind: User
     name: system:anonymous
   

4. access https://localhost:10259/debug/pprof/

The tool requires to pull the heap file and the kube-scheduler binary.

Steps:

1. Pull the heap data by accessing https://localhost:10259/debug/pprof/heap
2. Extract the kube-scheduler binary from the corresponding image (by checking the kube-scheduler pod manifest):

   $ podman pull --authfile $AUTHFILE $KUBE_SCHEDULER_IMAGE
   $ podman cp $(podman create --name kube-scheduler $KUBE_SCHEDULER_IMAGE):/usr/bin/kube-scheduler ./kube-scheduler

   • $AUTHFILE corresponds to your authentication file if not already located in the known paths
   • $KUBE_SCHEDULER_IMAGE corresponds to the kube-scheduler image found in a kube-scheduler pod manifest
3. Run go tool pprof kube-scheduler heap

From https://github.com/kubernetes/kubernetes/blob/be77b0b82b01a3fc810118f095594ec8bdd3c3aa/pkg/scheduler/internal/cache/debugger/debugger.go#L58:

// CacheDebugger provides ways to check and write cache information for debugging.
// ListenForSignal starts a goroutine that will trigger the CacheDebugger's
// behavior when the process receives SIGINT (Windows) or SIGUSER2 (non-Windows).

When a kube-scheduler process receives SIGUSER2 the node cache gets dumped into the logs. E.g.:

I0105 03:32:31.936642       1 dumper.go:52] "Dump of cached NodeInfo" nodes=<
    Node name: NODENAME1
    Deleted: false
    Requested Resources: ...
    Scheduled Pods(number: 41):
    name: POD_NAME, namespace: POD_NAMESPACE, uid: 23c63c58-cc36-48be-97d9-f4f6088a709d, phase: Running, nominated node:
    name: POD_NAME, namespace: POD_NAMESPACE, uid: 04b3b3b4-52a3-46d0-b7ff-aa748eecd404, phase: Running, nominated node:
    ...


    Node name: NODENAME2
    Deleted: false
    Requested Resources: ...
    Scheduled Pods(number: 53):
    name: POD_NAME, namespace: POD_NAMESPACE, uid: 7cbce63f-3fb9-404a-a69b-6728592e6b2, phase: Running, nominated node:
    name: POD_NAME, namespace: POD_NAMESPACE, uid: 50bc7d7e-bd30-4c47-82ce-a9d3eb737434, phase: Running, nominated node:
    ...