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 examples. 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?

The following describes the conversion of a single-master cluster into 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. Make sure that Deckhouse queue is empty.

  5. 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

  6. 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 command output should indicate that Terraform found no inconsistencies and no changes are required.

  7. 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>

    For Yandex Cloud, when using external addresses on master nodes, the number of array elements in the masterNodeGroup.instanceClass.externalIPAddresses parameter must equal the number of master nodes. If Auto is used (public IP addresses are provisioned automatically), the number of array elements must still equal the number of master nodes.

    To illustrate, with three master nodes (masterNodeGroup.replicas: 3) and automatic address reservation, the masterNodeGroup.instanceClass.externalIPAddresses parameter would look as follows:

    externalIPAddresses:
- "Auto"
- "Auto"
- "Auto"

  8. 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>

  9. 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?

The following describes the conversion of a multi-master cluster into a single-master.

The steps described below must be performed from the first in order of the master node of the cluster (master-0). This is because the cluster is always scaled in order: for example, it is impossible to delete nodes master-0 and master-1, leaving master-2.

  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. Make sure that Deckhouse queue is empty.

  5. 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

  6. 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 command output should indicate that Terraform found no inconsistencies and no changes are required.

  7. 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>

    For Yandex Cloud, when using external addresses on master nodes, the number of array elements in the masterNodeGroup.instanceClass.externalIPAddresses parameter must equal the number of master nodes. If Auto is used (public IP addresses are provisioned automatically), the number of array elements must still equal the number of master nodes.

    To illustrate, with three master nodes (masterNodeGroup.replicas: 1) and automatic address reservation, the masterNodeGroup.instanceClass.externalIPAddresses parameter would look as follows:

    externalIPAddresses:
- "Auto"
  8. 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-

  9. 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) -- \
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

  10. drain the nodes being deleted:

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

  11. 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>).

  12. 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

  13. In the cluster, delete the Node objects associated with the nodes being deleted:

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

  14. 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. Make sure that Deckhouse queue is empty.
  5. Remove the node.deckhouse.io/group: master and node-role.kubernetes.io/control-plane: "" labels.

  6. 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) -- \
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

  7. 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. Make sure that Deckhouse queue is empty.

  5. 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

  6. 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 command output should indicate that Terraform found no inconsistencies and no changes are required.

  7. 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 name | head -n1) -- \
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 name | head -n1) -- \
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.
  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

Use the etcdctl member list command.

Example:

kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod -l component=etcd,tier=control-plane -o name | head -n1) -- \
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. For this command, every control-plane address must be passed after --endpoints flag. The fifth parameter in the output table will be true for the leader.

Example of a script that automatically passes all control-plane nodes to the command:

MASTER_NODE_IPS=($(kubectl get nodes -l \
node-role.kubernetes.io/control-plane="" \
-o 'custom-columns=IP:.status.addresses[?(@.type=="InternalIP")].address' \
--no-headers))
unset ENDPOINTS_STRING
for master_node_ip in ${MASTER_NODE_IPS[@]}
do ENDPOINTS_STRING+="--endpoints https://${master_node_ip}:2379 "
done
kubectl -n kube-system exec -ti $(kubectl -n kube-system get pod \
-l component=etcd,tier=control-plane -o name | head -n1) \
-- etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt  --cert /etc/kubernetes/pki/etcd/ca.crt \
--key /etc/kubernetes/pki/etcd/ca.key \
$(echo -n $ENDPOINTS_STRING) endpoint status -w table

What if something went wrong?

During its operation, control-plane-manager automatically creates backups of configurations and data that may be useful in case of issues. These backups are stored in the /etc/kubernetes/deckhouse/backup directory. If errors or unforeseen situations occur during operation, you can use these backups to restore to the previous stable state.

What if the etcd cluster fails?

If the etcd cluster is not functioning and it cannot be restored from a backup, you can attempt to rebuild it from scratch by following the steps below.

  1. First, on all nodes that are part of your etcd cluster, except for one, remove the etcd.yaml manifest located in the /etc/kubernetes/manifests/ directory. 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.

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.

What if etcd restarts with an error?

This method may be necessary if the --force-new-cluster option doesn’t restore etcd work. Such a scenario can occur during an unsuccessful converge of master nodes, where a new master node was created with an old etcd disk, changed its internal address, and other master nodes are absent. Symptoms indicating the need for this method include: the etcd container being stuck in an endless restart with the log showing the error: panic: unexpected removal of unknown remote peer.

  1. Install the etcdutl utility.

  2. Create a new etcd database snapshot from the current local snapshot (/var/lib/etcd/member/snap/db):

    ./etcdutl snapshot restore /var/lib/etcd/member/snap/db --name <HOSTNAME> \
--initial-cluster=HOSTNAME=https://<ADDRESS>:2380 --initial-advertise-peer-urls=https://ADDRESS:2380 \
--skip-hash-check=true --data-dir /var/lib/etcdtest
    • <HOSTNAME> — the name of the master node;
    • <ADDRESS> — the address of the master node.

  3. Execute the following commands to use the new snapshot:

    cp -r /var/lib/etcd /tmp/etcd-backup
rm -rf /var/lib/etcd
mv /var/lib/etcdtest /var/lib/etcd

  4. Locate the etcd and api-server containers:

    crictl ps -a | egrep "etcd|apiserver"

  5. Remove the etcd and api-server containers:

    crictl rm <CONTAINER-ID>
  6. Restart the master node.

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 detailed information on how to configure audit-policy.yaml file here:

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.

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.

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.

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.

etcd backup and restore

What is done automatically

CronJob kube-system/d8-etcd-backup-* is automatically started at 00:00 UTC+0. The result is saved in /var/lib/etcd/etcd-backup.tar.gz on all nodes with control-plane in the cluster (master nodes).

How to manually backup etcd

Using Deckhouse CLI (Deckhouse Kubernetes Platform v1.65+)

Starting with Deckhouse Kubernetes Platform v1.65, a new d8 backup etcd tool is available for taking snapshots of etcd state.

d8 backup etcd --kubeconfig $KUBECONFIG ./etcd-backup.snapshot

Using bash (Deckhouse Kubernetes Platform v1.64 and older)

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

#!/usr/bin/env bash
set -e

pod=etcd-`hostname`
kubectl -n kube-system exec "$pod" -- /usr/bin/etcdctl --cacert /etc/kubernetes/pki/etcd/ca.crt --cert /etc/kubernetes/pki/etcd/ca.crt --key /etc/kubernetes/pki/etcd/ca.key --endpoints https://127.0.0.1:2379/ snapshot save /var/lib/etcd/${pod##*/}.snapshot && \
mv /var/lib/etcd/"${pod##*/}.snapshot" etcd-backup.snapshot && \
cp -r /etc/kubernetes/ ./ && \
tar -cvzf kube-backup.tar.gz ./etcd-backup.snapshot ./kubernetes/
rm -r ./kubernetes ./etcd-backup.snapshot

In the current directory etcd snapshot file kube-backup.tar.gz 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:

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 documentation for learn about etcd disaster recovery procedures from snapshots.

How do perform a full recovery of the cluster state from an etcd backup?

The following steps will be described to restore to the previous state of the cluster from a backup in case of complete data loss.

Restoring a single-master cluster

Follow these steps to restore a single-master cluster on master node:

  1. Find etcdctl utility on the master-node and copy the executable to /usr/local/bin/:

    cp $(find /var/lib/containerd/io.containerd.snapshotter.v1.overlayfs/snapshots/ \
-name etcdctl -print | tail -n 1) /usr/local/bin/etcdctl
etcdctl version

    The result must be a correct output of etcdctl version command without errors.

    Alternatively, you can download etcdctl executable to the node (preferably its version is the same as the etcd version in the cluster):

    wget "https://github.com/etcd-io/etcd/releases/download/v3.5.16/etcd-v3.5.16-linux-amd64.tar.gz"
tar -xzvf etcd-v3.5.16-linux-amd64.tar.gz && mv etcd-v3.5.16-linux-amd64/etcdctl /usr/local/bin/etcdctl

    You can check current etcd version using following command (might not work, if etcd and Kubernetes API are already unavailable):

    kubectl -n kube-system exec -ti etcd-$(hostname) -- etcdctl version

  2. Stop the etcd.

    mv /etc/kubernetes/manifests/etcd.yaml ~/etcd.yaml

  3. Save the current etcd data.

    cp -r /var/lib/etcd/member/ /var/lib/deckhouse-etcd-backup

  4. Clean the etcd directory.

    rm -rf /var/lib/etcd/member/

  5. Put the etcd backup to ~/etcd-backup.snapshot file.

  6. Restore the etcd database.

    ETCDCTL_API=3 etcdctl snapshot restore ~/etcd-backup.snapshot --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/ --data-dir=/var/lib/etcd

  7. Run etcd. The process may take some time.

    mv ~/etcd.yaml /etc/kubernetes/manifests/etcd.yaml
crictl ps --label io.kubernetes.pod.name=etcd-$HOSTNAME

Restorige a multi-master cluster

Follow these steps to restore a multi-master cluster:

  1. Explicitly set the High Availability (HA) mode by specifying the highAvailability parameter. This is necessary, for example, in order not to lose one Prometheus replica and its PVC, since HA is disabled by default in single-master mode.

  2. Switch the cluster to single-master mode according to instruction for cloud clusters or independently remove static master-node from the cluster.

  3. On a single master-node, perform the steps to restore etcd from backup in accordance with the instructions for a single-master cluster.

  4. When etcd operation is restored, delete the information about the master nodes already deleted in step 1 from the cluster:

    kubectl delete node MASTER_NODE_I

  5. Restart all nodes of the cluster.

  6. Wait for the deckhouse queue to complete:

    kubectl -n d8-system exec svc/deckhouse-leader -c deckhouse -- deckhouse-controller queue main

  7. Switch the cluster back to multi-master mode according to instructions for cloud clusters or instructions for static or hybrid clusters.

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 auger.

Example of steps to restore objects from an etcd backup

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

  • To decode objects from etcd you would need auger. It can be built from source on any machine that has Docker installed (it cannot be done on cluster nodes).

    git clone -b v1.0.1 --depth 1 https://github.com/etcd-io/auger
cd auger
make release
build/auger -h

  • Resulting executable build/auger, and also the snapshot from the backup copy of etcd must be uploaded on master-node, on which following actions would be performed.

Following actions are performed on a master node, to which etcd snapshot file and auger tool were copied:

  1. Set full path for snapshot file and for the tool into environmental variables:

    SNAPSHOT=/root/etcd-restore/etcd-backup.snapshot
AUGER_BIN=/root/auger 
chmod +x $AUGER_BIN
  2. Run a Pod with temporary instance of etcd.

    • Create Pod manifest. It should schedule on current master node by $HOSTNAME variable, and mounts snapshot file by $SNAPSHOT variable, which it then restores in temporary etcd instance:

      cat <<EOF >etcd.pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: etcdrestore
  namespace: default
spec:
  nodeName: $HOSTNAME
  tolerations:
  - operator: Exists
  initContainers:
  - command:
    - etcdctl
    - snapshot
    - restore
    - "/tmp/etcd-snapshot"
    image: $(kubectl -n kube-system get pod -l component=etcd -o jsonpath="{.items[*].spec.containers[*].image}" | cut -f 1 -d ' ')
    imagePullPolicy: IfNotPresent
    name: etcd-snapshot-restore
    volumeMounts:
    - name: etcddir
      mountPath: /default.etcd
    - name: etcd-snapshot
      mountPath: /tmp/etcd-snapshot
      readOnly: true
  containers:
  - command:
    - etcd
    image: $(kubectl -n kube-system get pod -l component=etcd -o jsonpath="{.items[*].spec.containers[*].image}" | cut -f 1 -d ' ')
    imagePullPolicy: IfNotPresent
    name: etcd-temp
    volumeMounts:
    - name: etcddir
      mountPath: /default.etcd
  volumes:
  - name: etcddir
    emptyDir: {}
  - name: etcd-snapshot
    hostPath:
      path: $SNAPSHOT
      type: File
EOF

    • Create Pod from the resulting manifest:

      kubectl create -f etcd.pod.yaml

  3. Set environment variables. In this example:

    • infra-production - namespace which we will search resources in.

    • /root/etcd-restore/output - path for outputting recovered resource manifests.

    • /root/auger - path to auger executable.

      FILTER=infra-production
BACKUP_OUTPUT_DIR=/root/etcd-restore/output
mkdir -p $BACKUP_OUTPUT_DIR && cd $BACKUP_OUTPUT_DIR

  4. Commands below will filter needed resources by $FILTER and output them into $BACKUP_OUTPUT_DIR directory:

    files=($(kubectl -n default exec etcdrestore -c etcd-temp -- etcdctl  --endpoints=localhost:2379 get / --prefix --keys-only | grep "$FILTER"))
for file in "${files[@]}"
do
  OBJECT=$(kubectl -n default exec etcdrestore -c etcd-temp -- etcdctl  --endpoints=localhost:2379 get "$file" --print-value-only | $AUGER_BIN decode)
  FILENAME=$(echo $file | sed -e "s#/registry/##g;s#/#_#g")
  echo "$OBJECT" > "$BACKUP_OUTPUT_DIR/$FILENAME.yaml"
  echo $BACKUP_OUTPUT_DIR/$FILENAME.yaml
done

  5. From resulting yaml files, delete creationTimestamp, UID, status and other operational fields, and then restore the objects:

    kubectl create -f deployments_infra-production_supercronic.yaml

  6. 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

Scheduler profiles

There are two predefined scheduler profiles:

  • default-scheduler: The default profile that distributes pods to nodes with the lowest load;
  • high-node-utilization: A profile that places pods on nodes with the highest load.

To specify a scheduler profile, use the spec.schedulerName parameter in the pod manifest.

Example of using a profile:

apiVersion: v1
kind: Pod
metadata:
  name: scheduler-example
  labels:
    name: scheduler-example
spec:
  schedulerName: high-node-utilization
  containers:
  - name: example-pod
    image: registry.k8s.io/pause:2.0

Pod scheduling stages

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

How to change or extend the scheduler logic

To change the logic of the scheduler it is possible to use the extension mechanism Extenders.

Each plugin is a webhook that must satisfy the following requirements:

  • Use of TLS.
  • Accessibility through a service within the cluster.
  • Support for standard Verbs (filterVerb = filter, prioritizeVerb = prioritize).
  • It is also assumed that all plugins can cache node information (nodeCacheCapable: true).

You can connect an extender using KubeSchedulerWebhookConfiguration resource.

When using the failurePolicy: Fail option, in case of an error in the webhook’s operation, the scheduler will stop working and new pods will not be able to start.