Managing control plane: FAQ

How do I add a master node to a static or hybrid cluster?

It is important to have an odd number of masters to ensure a quorum.

Adding a master node to a static or hybrid cluster has no difference from adding a regular node to a cluster. To do this, use the corresponding instruction. All the necessary actions to configure a cluster control plane components on the new master nodes are performed automatically. Wait until the master nodes appear in Ready status.

How do I add a master nodes to a cloud cluster (single-master to a multi-master)?

Before adding nodes, ensure you have the required quotas in the cloud provider.

It is important to have an odd number of masters to ensure a quorum.

1. Make a backup of etcd and the /etc/kubernetes directory.
2. Transfer the archive to a server outside the cluster (e.g., on a local machine).
3. Ensure there are no alerts in the cluster that can prevent the creation of new master nodes.

4. Run the appropriate edition and version of the Deckhouse installer container on the local machine (change the container registry address if necessary):

   DH_VERSION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/version}') \
   DH_EDITION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/edition}' | tr '[:upper:]' '[:lower:]' ) \
   docker run --pull=always -it -v "$HOME/.ssh/:/tmp/.ssh/" \
     registry.deckhouse.io/deckhouse/${DH_EDITION}/install:${DH_VERSION} bash
   

5. In the installer container, run the following command to check the state before working:

   dhctl terraform check --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> --ssh-host <MASTER-NODE-0-HOST>
   

   The response should tell you that Terraform does not want to change anything.

6. In the installer container, run the following command and specify the required number of replicas using the masterNodeGroup.replicas parameter:

   dhctl config edit provider-cluster-configuration --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> \
     --ssh-host <MASTER-NODE-0-HOST>
   

7. In the installer container, run the following command to start scaling:

   dhctl converge --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> --ssh-host <MASTER-NODE-0-HOST>
   

8. Wait until the required number of master nodes are Ready and all control-plane-manager instances are up and running:

   kubectl -n kube-system wait pod --timeout=10m --for=condition=ContainersReady -l app=d8-control-plane-manager
   

How do I reduce the number of master nodes in a cloud cluster (multi-master to single-master)?

1. Make a backup of etcd and the /etc/kubernetes directory.
2. Transfer the archive to a server outside the cluster (e.g., on a local machine).
3. Ensure there are no alerts in the cluster that can prevent the update of the master nodes.

4. Run the appropriate edition and version of the Deckhouse installer container on the local machine (change the container registry address if necessary):

   DH_VERSION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/version}') \
   DH_EDITION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/edition}' | tr '[:upper:]' '[:lower:]' ) \
   docker run --pull=always -it -v "$HOME/.ssh/:/tmp/.ssh/" \
     registry.deckhouse.io/deckhouse/${DH_EDITION}/install:${DH_VERSION} bash
   

5. In the installer container, run the following command to check the state before working:

   dhctl terraform check --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> --ssh-host <MASTER-NODE-0-HOST>
   

   The response should tell you that Terraform does not want to change anything.

6. Run the following command in the installer container and set masterNodeGroup.replicas to 1:

   dhctl config edit provider-cluster-configuration --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> \
     --ssh-user=<USERNAME> --ssh-host <MASTER-NODE-0-HOST>
   

7. Remove the following labels from the master nodes to be deleted:
   • node-role.kubernetes.io/control-plane
   • node-role.kubernetes.io/master
   • node.deckhouse.io/group

   Use the following command to remove labels:

   kubectl label node <MASTER-NODE-N-NAME> node-role.kubernetes.io/control-plane- node-role.kubernetes.io/master- node.deckhouse.io/group-
   

8. Make sure that the master nodes to be deleted are no longer listed as etcd cluster members:

   kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o name | head -n1) -- sh -c \
   "ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt \
   --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key \
   --endpoints https://127.0.0.1:2379/ member list -w table"
   

9. Drain the nodes being deleted:

   kubectl drain <MASTER-NODE-N-NAME> --ignore-daemonsets --delete-emptydir-data
   

10. Shut down the virtual machines corresponding to the nodes to be deleted, remove the instances of those nodes from the cloud and the disks connected to them (kubernetes-data-master-<N>).

11. In the cluster, delete the Pods running on the nodes being deleted:

    kubectl delete pods --all-namespaces --field-selector spec.nodeName=<MASTER-NODE-N-NAME> --force
    

12. In the cluster, delete the Nore objects associated with the nodes being deleted:

    kubectl delete node <MASTER-NODE-N-NAME>
    

13. In the installer container, run the following command to start scaling:

    dhctl converge --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> --ssh-host <MASTER-NODE-0-HOST>
    

How do I dismiss the master role while keeping the node?

1. Make a backup of etcd and the /etc/kubernetes directory.
2. Transfer the archive to a server outside the cluster (e.g., on a local machine).
3. Ensure there are no alerts in the cluster that can prevent the update of the master nodes.
4. Remove the node.deckhouse.io/group: master and node-role.kubernetes.io/control-plane: "" labels.

5. Make sure that the master node to be deleted is no longer listed as a member of the etcd cluster:

   kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o name | head -n1) -- sh -c \
   "ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt \
   --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key \
   --endpoints https://127.0.0.1:2379/ member list -w table"
   

6. Exec to the node and run the following commands:

   rm -f /etc/kubernetes/manifests/{etcd,kube-apiserver,kube-scheduler,kube-controller-manager}.yaml
   rm -f /etc/kubernetes/{scheduler,controller-manager}.conf
   rm -f /etc/kubernetes/authorization-webhook-config.yaml
   rm -f /etc/kubernetes/admin.conf /root/.kube/config
   rm -rf /etc/kubernetes/deckhouse
   rm -rf /etc/kubernetes/pki/{ca.key,apiserver*,etcd/,front-proxy*,sa.*}
   rm -rf /var/lib/etcd/member/
   

How do I switch to a different OS image in a multi-master cluster?

1. Make a backup of etcd and the /etc/kubernetes directory.
2. Transfer the archive to a server outside the cluster (e.g., on a local machine).
3. Ensure there are no alerts in the cluster that can prevent the update of the master nodes.

4. Run the appropriate edition and version of the Deckhouse installer container on the local machine (change the container registry address if necessary):

   DH_VERSION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/version}') \
   DH_EDITION=$(kubectl -n d8-system get deployment deckhouse -o jsonpath='{.metadata.annotations.core\.deckhouse\.io\/edition}' | tr '[:upper:]' '[:lower:]' ) \
   docker run --pull=always -it -v "$HOME/.ssh/:/tmp/.ssh/" \
     registry.deckhouse.io/deckhouse/${DH_EDITION}/install:${DH_VERSION} bash
   

5. In the installer container, run the following command to check the state before working:

   dhctl terraform check --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> \
     --ssh-host <MASTER-NODE-0-HOST> --ssh-host <MASTER-NODE-1-HOST> --ssh-host <MASTER-NODE-2-HOST>
   

   The response should tell you that Terraform does not want to change anything.

6. In the installer container, run the following command and specify the required OS image using the masterNodeGroup.instanceClass parameter (specify the addresses of all master nodes using the -ssh-host parameter):

   dhctl config edit provider-cluster-configuration --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> \
     --ssh-host <MASTER-NODE-0-HOST> --ssh-host <MASTER-NODE-1-HOST> --ssh-host <MASTER-NODE-2-HOST>
   

Repeat the steps below for each master node one by one, starting with the node with the highest number (suffix 2) and ending with the node with the lowest number (suffix 0).

1. Select the master node to update (enter its name):

   NODE="<MASTER-NODE-N-NAME>"
   

2. Run the following command to remove the node-role.kubernetes.io/control-plane, node-role.kubernetes.io/master, and node.deckhouse.io/group labels from the node:

   kubectl label node ${NODE} \
     node-role.kubernetes.io/control-plane- node-role.kubernetes.io/master- node.deckhouse.io/group-
   

3. Make sure that the node is no longer listed as an etcd cluster member:

   kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o json | jq -r '.items[] | select( .status.conditions[] | select(.type == "ContainersReady" and .status == "True")) | .metadata.name') -- sh -c \
   "ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt \
   --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key \
   --endpoints https://127.0.0.1:2379/ member list -w table"
   

4. Drain the node:

   kubectl drain ${NODE} --ignore-daemonsets --delete-emptydir-data
   

5. Shut down the virtual machine associated with the node, remove the node instance from the cloud and the disks connected to it (kubernetes-data).

6. In the cluster, delete the Pods remaining on the node being deleted:

   kubectl delete pods --all-namespaces --field-selector spec.nodeName=${NODE} --force
   

7. In the cluster, delete the Node object for the node being deleted:

   kubectl delete node ${NODE}
   

8. In the installer container, run the following command to create the updated node:

   You should read carefully what converge is going to do when it asks for approval.

   If converge requests approval for another master node, it should be skipped by saying “no”.

   dhctl converge --ssh-agent-private-keys=/tmp/.ssh/<SSH_KEY_FILENAME> --ssh-user=<USERNAME> \
     --ssh-host <MASTER-NODE-0-HOST> --ssh-host <MASTER-NODE-1-HOST> --ssh-host <MASTER-NODE-2-HOST>
   

9. On the newly created node, check the systemd-unit log for the bashible.service. Wait until the node configuration is complete (you will see a message nothing to do in the log):

   journalctl -fu bashible.service
   

10. Make sure the node is listed as an etcd cluster member:

    kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o json | jq -r '.items[] | select( .status.conditions[] | select(.type == "ContainersReady" and .status == "True")) | .metadata.name') -- sh -c \
    "ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt \
    --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key \
    --endpoints https://127.0.0.1:2379/ member list -w table"
    

11. Make sure control-plane-manager is running on the node:

    kubectl -n kube-system wait pod --timeout=10m --for=condition=ContainersReady \
      -l app=d8-control-plane-manager --field-selector spec.nodeName=${NODE}
    

12. Proceed to update the next node (repeat the steps above).

How do I switch to a different OS image in a single-master cluster?

1. Convert your single-master cluster to a multi-master one, as described in the guide on adding master nodes to a cluster.

   In addition to increasing the number of replicas, you can also specify the image with the required OS version using the masterNode.instanceClass parameter.

2. Update the master nodes following the instructions.
3. Convert your multi-master cluster to a single-master one according to the guide on excluding master nodes from the cluster.

How do I view the list of etcd members?

Option 1

1. Exec to the etcd Pod:

   kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o name | head -n1) sh
   

2. Execute the command:

   ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt --cert /etc/kubernetes/pki/etcd/ca.crt \
   --key /etc/kubernetes/pki/etcd/ca.key --endpoints https://127.0.0.1:2379/ member list -w table
   

Warning! The last parameter in the output table shows etcd member is in learner state, is not in leader state.

Option 2

Use the etcdctl endpoint status command. The fith parameter in the output table will be true for the leader.

Example:

$ ETCDCTL_API=3 etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt --cert /etc/kubernetes/pki/etcd/ca.crt \  
  --key /etc/kubernetes/pki/etcd/ca.key --endpoints https://127.0.0.1:2379/ endpoint status -w table

https://10.2.1.101:2379, ade526d28b1f92f7, 3.5.3, 177 MB, false, false, 42007, 406566258, 406566258,
https://10.2.1.102:2379, d282ac2ce600c1ce, 3.5.3, 182 MB, true, false, 42007, 406566258, 406566258,

What if something went wrong?

The control-plane-manager saves backups to /etc/kubernetes/deckhouse/backup. They can be useful in diagnosing the issue.

What if the etcd cluster fails?

1. Stop (delete the /etc/kubernetes/manifests/etcd.yaml file) etcd on all nodes except one. This last node will serve as a starting point for the new multi-master cluster.
2. On the last node, edit etcd manifest /etc/kubernetes/manifests/etcd.yaml and add the parameter --force-new-cluster to spec.containers.command.
3. After the new cluster is ready, remove the --force-new-cluster parameter.

Caution! This operation is unsafe and breaks the guarantees given by the consensus protocol. Note that it brings the cluster to the state that was saved on the node. Any pending entries will be lost.

How do I configure additional audit policies?

1. Enable the auditPolicyEnabled flag in the module configuration:

   apiVersion: deckhouse.io/v1alpha1
   kind: ModuleConfig
   metadata:
     name: control-plane-manager
   spec:
     version: 1
     settings:
       apiserver:
         auditPolicyEnabled: true
   

2. Create the kube-system/audit-policy Secret containing a base64-encoded yaml file:

   apiVersion: v1
   kind: Secret
   metadata:
     name: audit-policy
     namespace: kube-system
   data:
     audit-policy.yaml: <base64>
   

   The minimum viable example of the audit-policy.yaml file looks as follows:

   apiVersion: audit.k8s.io/v1
   kind: Policy
   rules:
   - level: Metadata
     omitStages:
     - RequestReceived
   

   You can find the detailed information about configuring the audit-policy.yaml file at the following links:

   • The official Kubernetes documentation.
   • The code of the generator script used in GCE.

How to omit Deckhouse built-in policy rules?

Set the apiserver.basicAuditPolicyEnabled module parameter to false.

An example:

apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
  name: control-plane-manager
spec:
  version: 1
  settings:
    apiserver:
      auditPolicyEnabled: true
      basicAuditPolicyEnabled: false

How stream audit log to stdout instead of files?

Set the apiserver.auditLog.output parameter to Stdout.

An example:

apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
  name: control-plane-manager
spec:
  version: 1
  settings:
    apiserver:
      auditPolicyEnabled: true
      auditLog:
        output: Stdout

How to deal with the audit log?

There must be some log scraper on master nodes (log-shipper, promtail, filebeat) that will monitor the log file:

/var/log/kube-audit/audit.log

The following fixed parameters of log rotation are in use:

• The maximum disk space is limited to 1000 Mb.
• Logs older than 7 days will be deleted.

Depending on the Policy settings and the number of requests to the apiserver, the amount of logs collected may be high. Thus, in some cases, logs can only be kept for less than 30 minutes.

Cautionary note

*Note! The current implementation of this feature isn’t safe and may lead to a temporary failure of the control plane.

The apiserver will not be able to start if there are unsupported options or a typo in the Secret.

If apiserver is unable to start, you have to manually disable the --audit-log-* parameters in the /etc/kubernetes/manifests/kube-apiserver.yaml manifest and restart apiserver using the following command:

docker stop $(docker ps | grep kube-apiserver- | awk '{print $1}')
# Or (depending on your CRI).
crictl stopp $(crictl pods --name=kube-apiserver -q)

After the restart, you will be able to fix the Secret or delete it:

kubectl -n kube-system delete secret audit-policy

How do I speed up the restart of Pods if the connection to the node has been lost?

By default, a node is marked as unavailable if it does not report its state for 40 seconds. After another 5 minutes, its Pods will be rescheduled to other nodes. Thus, the overall application unavailability lasts approximately 6 minutes.

In specific cases, if an application cannot run in multiple instances, there is a way to lower its unavailability time:

1. Reduce the period required for the node to become Unreachable if the connection to it is lost by setting the nodeMonitorGracePeriodSeconds parameter.
2. Set a lower timeout for evicting Pods on a failed node using the failedNodePodEvictionTimeoutSeconds parameter.

An example

apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
  name: control-plane-manager
spec:
  version: 1
  settings:
    nodeMonitorGracePeriodSeconds: 10
    failedNodePodEvictionTimeoutSeconds: 50

In this case, if the connection to the node is lost, the applications will be restarted in about 1 minute.

Cautionary note

Both these parameters directly impact the CPU and memory resources consumed by the control plane. By lowering timeouts, we force system components to send statuses more frequently and check the resource state more often.

When deciding on the appropriate threshold values, consider resources consumed by the control nodes (graphs can help you with this). Note that the lower parameters are, the more resources you may need to allocate to these nodes.

etc backup and restore

How do make etcd backup?

Login into any control-plane node with root user and use next script:

#!/usr/bin/env bash

for pod in $(kubectl get pod -n kube-system -l component=etcd,tier=control-plane -o name); do
  if kubectl -n kube-system exec "$pod" -- sh -c "ETCDCTL_API=3 /usr/bin/etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key --endpoints https://127.0.0.1:2379/ snapshot save /tmp/${pod##*/}.snapshot" && \
  kubectl -n kube-system exec "$pod" -- gzip -c /tmp/${pod##*/}.snapshot | zcat > "${pod##*/}.snapshot" && \
  kubectl -n kube-system exec "$pod" -- sh -c "cd /tmp && sha256sum ${pod##*/}.snapshot" | sha256sum -c && \
  kubectl -n kube-system exec "$pod" -- rm "/tmp/${pod##*/}.snapshot"; then
    mv "${pod##*/}.snapshot" etcd-backup.snapshot
    break
  fi
done
cp -r /etc/kubernetes/ ./
tar -cvzf kube-backup.tar.gz ./etcd-backup.snapshot ./kubernetes/
rm -r ./kubernetes

In the current directory etcd snapshot file etcd-backup.snapshot will be created from one of an etcd cluster members. From this file, you can restore the previous etcd cluster state in the future.

Also, we recommend making a backup of the /etc/kubernetes directory, which contains:

• manifests and configurations of control-plane components;
• Kubernetes cluster PKI. This directory will help to quickly restore a cluster in case of complete loss of control-plane nodes without creating a new cluster and without rejoin the remaining nodes into the new cluster.

We recommend encrypting etcd snapshot backups as well as backup of the directory /etc/kubernetes/ and saving them outside the Deckhouse cluster. You can use one of third-party files backup tools, for example: Restic, Borg, Duplicity, etc.

You can see here for learn about etcd disaster recovery procedures from snapshots.

How do I restore a Kubernetes object from an etcd backup?

To get cluster objects data from an etcd backup, you need:

1. Start an temporary instance of etcd.
2. Fill it with data from the backup.
3. Get desired objects using etcdhelper.

Example of steps to restore objects from an etcd backup

In the example below, etcd-snapshot.bin is a etcd shapshot, infra-production is the namespace in which objects need to be restored.

1. Start the Pod, with a temporary instance of etcd.

   • Prepare the file etcd.pod.yaml of the Pod template by executing the following commands:

     cat <<EOF >etcd.pod.yaml 
     apiVersion: v1
     kind: Pod
     metadata:
       name: etcdrestore
       namespace: default
     spec:
       containers:
       - command:
         - /bin/sh
         - -c
         - "sleep 96h"
         image: IMAGE
         imagePullPolicy: IfNotPresent
         name: etcd
         volumeMounts:
         - name: etcddir
           mountPath: /default.etcd
       volumes:
       - name: etcddir
         emptyDir: {}
     EOF
     IMG=`kubectl -n kube-system get pod -l component=etcd -o jsonpath="{.items[*].spec.containers[*].image}" | cut -f 1 -d ' '`
     sed -i -e "s#IMAGE#$IMG#" etcd.pod.yaml
     

   • Create the Pod:

     kubectl create -f etcd.pod.yaml
     

2. Copy the etcdhelper and the etc snapshot into the Pod container.

   You can build etcdhelper from the source code or copy from another image (for example, from the image of etcdhelper on Docker Hub).

   Example:

   kubectl cp etcd-snapshot.bin default/etcdrestore:/tmp/etcd-snapshot.bin
   kubectl cp etcdhelper default/etcdrestore:/usr/bin/etcdhelper
   

3. Set the rights to run etcdhelper in the container, restore the data from the backup and run etcd.

   Example:

   ~ # kubectl -n default exec -it etcdrestore -- sh
   / # chmod +x /usr/bin/etcdhelper
   / # etcdctl snapshot restore /tmp/etcd-snapshot.bin
   / # etcd &
   

4. Get necessary cluster objects by filtering them using `grep’.

   Example:

   ~ # kubectl -n default exec -it etcdrestore -- sh
   / # mkdir /tmp/restored_yaml
   / # cd /tmp/restored_yaml
   /tmp/restored_yaml # for o in `etcdhelper -endpoint 127.0.0.1:2379 ls /registry/ | grep infra-production` ; do etcdhelper -endpoint 127.0.0.1:2379 get $o > `echo $o | sed -e "s#/registry/##g;s#/#_#g"`.yaml ; done
   

   Replacing characters with sed in the example allows you to save descriptions of objects in files named similar to the etcd registry structure. For example: /registry/deployments/infra-production/supercronic.yaml → deployments_infra-production_supercronic.yaml.

5. Copy the received object descriptions to the master node:

   kubectl cp default/etcdrestore:/tmp/restored_yaml restored_yaml
   

6. Delete information about the creation time, UID, status, and other operational data from the received object descriptions, and then restore the objects:

   kubectl create -f restored_yaml/deployments_infra-production_supercronic.yaml
   

7. Delete the Pod with a temporary instance of etcd:

   kubectl -n default delete pod etcdrestore
   

How the node to run the Pod on is selected

The Kubernetes scheduler component selects the node to run the Pod on. The selection process involves two phases, namely Filtering and Scoring. They are supposed to efficiently distribute the Pods between the nodes. Although there are some additional phases, such as pre-filtering, post-filtering, and so on, you can safely narrow them down to the global phases mentioned above, as they merely increase flexibility and help to optimize things.

The structure of the Kubernetes scheduler

The Scheduler comprises plugins that function in either or both phases.

Example of plugins:

• ImageLocality — favors nodes that already have the container images that the Pod runs. Phase: Scoring.
• TaintToleration — implements taints and tolerations. Phases: Filtering, Scoring.
• NodePorts - checks whether the ports required for the Pod to run are available on the node. Phase: Filtering.

The full list of plugins is available in the Kubernetes documentation.

Working logic

The selection process starts with the Filtering phase. During it, filter plugins select nodes that satisfy filter conditions such as taints, nodePorts, nodeName, unschedulable, etc. If the nodes are in different zones, the scheduler alternates zones when selecting to ensure that all Pods will not end up in the same zone.

Suppose there are two zones with the following nodes:

Zone 1: Node 1, Node 2, Node 3, Node 4
Zone 2: Node 5, Node 6

In this case, the nodes will be selected in the following order:

Node 1, Node 5, Node 2, Node 6, Node 3, Node 4.

Note that Kubernetes limits the number of nodes to calculate their scores during scheduling. This optimizes the selection process and prevents unnecessary scoring. By default, the threshold is linear. For clusters with less than or equal to 50 nodes, 100% of nodes are considered for scheduling; for clusters with 100 nodes, a 50%-threshold is used; and for clusters with 5000 nodes, a 10%-threshold is used. The minimum threshold value is 5% for clusters with more than 5000 nodes. Therefore, even if all the conditions are met, a node may not be included in the list of candidates for scheduling if the default settings are used. This logic can be changed (read more about the parameter percentage Of Nodes To Score in the Kubernetes documentation), but Deckhouse does not provide such an option.

The Scoring phase follows once the nodes that meet the conditions are selected. Each plugin evaluates the filtered node list and assigns a score to each node based on available resources, Pod capacity, affinity, volume provisioning, and other factors. The scores from the different plugins are then summed up and the node with the highest score is selected. If several nodes have the same score, the node is selected at random.

Finally, the scheduler assigns the Pod to the node with the highest ranking.

Documentation

• General description of the scheduler
• Plugin system
• Node Filtering Details
• Scheduler source code