×

As a day 2 operation, you can resolve and prevent performance issues of your OpenStack installation by moving etcd from a root volume (provided by OpenStack Cinder) to a dedicated ephemeral local disk.

Deploying OpenStack on local disk

If you have an existing OpenStack cloud, you can move etcd from that cloud to a dedicated ephemeral local disk.

Prerequisites
  • You have an OpenStack cloud with a working Cinder.

  • Your OpenStack cloud has at least 75 GB of available storage to accommodate 3 root volumes for the OpenShift control plane.

  • The OpenStack cloud is deployed with Nova ephemeral storage that uses a local storage backend and not rbd.

Procedure
  1. Create a Nova flavor for the control plane with at least 10 GB of ephemeral disk by running the following command, replacing the values for --ram, --disk, and <flavor_name> based on your environment:

    $ openstack flavor create --<ram 16384> --<disk 0> --ephemeral 10 --vcpus 4 <flavor_name>
  2. Deploy a cluster with root volumes for the control plane; for example:

    Example YAML file
    # ...
    controlPlane:
      name: master
      platform:
        openstack:
          type: ${CONTROL_PLANE_FLAVOR}
          rootVolume:
            size: 25
            types:
            - ${CINDER_TYPE}
      replicas: 3
    # ...
  3. Deploy the cluster you created by running the following command:

    $ openshift-install create cluster --dir <installation_directory> (1)
    1 For <installation_directory>, specify the location of the customized ./install-config.yaml file that you previously created.
  4. Verify that the cluster you deployed is healthy before proceeding to the next step by running the following command:

    $ oc wait clusteroperators --all --for=condition=Progressing=false (1)
    1 Ensures that the cluster operators are finished progressing and that the cluster is not deploying or updating.
  5. Create a file named 98-var-lib-etcd.yaml by using the following YAML file:

    Details
    apiVersion: machineconfiguration.openshift.io/v1
    kind: MachineConfig
    metadata:
      labels:
        machineconfiguration.openshift.io/role: master
      name: 98-var-lib-etcd
    spec:
      config:
        ignition:
          version: 3.4.0
        systemd:
          units:
          - contents: |
              [Unit]
              Description=Mount local-etcd to /var/lib/etcd
    
              [Mount]
              What=/dev/disk/by-label/local-etcd (1)
              Where=/var/lib/etcd
              Type=xfs
              Options=defaults,prjquota
    
              [Install]
              WantedBy=local-fs.target
            enabled: true
            name: var-lib-etcd.mount
          - contents: |
              [Unit]
              Description=Create local-etcd filesystem
              DefaultDependencies=no
              After=local-fs-pre.target
              ConditionPathIsSymbolicLink=!/dev/disk/by-label/local-etcd (2)
    
              [Service]
              Type=oneshot
              RemainAfterExit=yes
              ExecStart=/bin/bash -c "[ -L /dev/disk/by-label/ephemeral0 ] || ( >&2 echo Ephemeral disk does not exist; /usr/bin/false )"
              ExecStart=/usr/sbin/mkfs.xfs -f -L local-etcd /dev/disk/by-label/ephemeral0 (3)
    
              [Install]
              RequiredBy=dev-disk-by\x2dlabel-local\x2detcd.device
            enabled: true
            name: create-local-etcd.service
          - contents: |
              [Unit]
              Description=Migrate existing data to local etcd
              After=var-lib-etcd.mount
              Before=crio.service (4)
    
              Requisite=var-lib-etcd.mount
              ConditionPathExists=!/var/lib/etcd/member
              ConditionPathIsDirectory=/sysroot/ostree/deploy/rhcos/var/lib/etcd/member (5)
    
              [Service]
              Type=oneshot
              RemainAfterExit=yes
    
              ExecStart=/bin/bash -c "if [ -d /var/lib/etcd/member.migrate ]; then rm -rf /var/lib/etcd/member.migrate; fi" (6)
    
              ExecStart=/usr/bin/cp -aZ /sysroot/ostree/deploy/rhcos/var/lib/etcd/member/ /var/lib/etcd/member.migrate
              ExecStart=/usr/bin/mv /var/lib/etcd/member.migrate /var/lib/etcd/member (7)
    
              [Install]
              RequiredBy=var-lib-etcd.mount
            enabled: true
            name: migrate-to-local-etcd.service
          - contents: |
              [Unit]
              Description=Relabel /var/lib/etcd
    
              After=migrate-to-local-etcd.service
              Before=crio.service
              Requisite=var-lib-etcd.mount
    
              [Service]
              Type=oneshot
              RemainAfterExit=yes
    
              ExecCondition=/bin/bash -c "[ -n \"$(restorecon -nv /var/lib/etcd)\" ]" (8)
    
              ExecStart=/usr/sbin/restorecon -R /var/lib/etcd
    
              [Install]
              RequiredBy=var-lib-etcd.mount
            enabled: true
            name: relabel-var-lib-etcd.service
    1 The etcd database must be mounted by the device, not a label, to ensure that systemd generates the device dependency used in this config to trigger filesystem creation.
    2 Do not run if the file system dev/disk/by-label/local-etcd already exists.
    3 Fails with an alert message if /dev/disk/by-label/ephemeral0 does not exist.
    4 Migrates existing data to local etcd database. This config does so after /var/lib/etcd is mounted, but before CRI-O starts so etcd is not running yet.
    5 Requires that etcd is mounted and does not contain a member directory, but the ostree does.
    6 Cleans up any previous migration state.
    7 Copies and moves in separate steps to ensure atomic creation of a complete member directory.
    8 Performs a quick check of the mount point directory before performing a full recursive relabel. If restorecon in the file path /var/lib/etcd cannot rename the directory, the recursive rename is not performed.

    After you apply the 98-var-lib-etcd.yaml file to the system, do not remove it. Removing this file will break etcd members and lead to system instability.

    If a rollback is necessary, modify the ControlPlaneMachineSet object to use a flavor that does not include ephemeral disks. This change regenerates the control plane nodes without using ephemeral disks for the etcd partition, which avoids issues related to the 98-var-lib-etcd.yaml file. It is safe to remove the 98-var-lib-etcd.yaml file only after the update to the ControlPlaneMachineSet object is complete and no control plane nodes are using ephemeral disks.

  6. Create the new MachineConfig object by running the following command:

    $ oc create -f 98-var-lib-etcd.yaml

    Moving the etcd database onto the local disk of each control plane machine takes time.

  7. Verify that the etcd databases has been transferred to the local disk of each control plane by running the following commands:

    1. Verify that the cluster is still updating by running the following command:

      $ oc wait --timeout=45m --for=condition=Updating=false machineconfigpool/master
    2. Verify that the cluster is ready by running the following command:

      $ oc wait node --selector='node-role.kubernetes.io/master' --for condition=Ready --timeout=30s
    3. Verify that the cluster Operators are running in the cluster by running the following command:

      $ oc wait clusteroperators --timeout=30m --all --for=condition=Progressing=false