mirrors/ceph-csi
1
0
Fork 0
mirror of https://github.com/ceph/ceph-csi.git synced 2024-11-23 23:00:19 +00:00
Code Issues Packages Projects Releases Wiki Activity
4ad4cb162b
ceph-csi/docs/ceph-csi-upgrade.md
Madhu Rajanna 2974ad3464 Update upgrade doc for node hang issue 
This PR updates the upgrade doc to handle the
node drain issue what we have seen in
https://github.com/ceph/ceph-csi/issues/756

Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
(cherry picked from commit 7f8c535c42)
2020-03-19 15:08:54 +00:00

12 KiB
Raw Blame History

Ceph-csi Upgrade

Pre-upgrade considerations

In some scenarios there is an issue in the CSI driver that will cause application pods to be disconnected from their mounts when the CSI driver is restarted. Since the upgrade would cause the CSI driver to restart if it is updated, you need to be aware of whether this affects your applications. This issue will happen when using the Ceph fuse client or rbd-nbd:

CephFS: If you are provision volumes for CephFS and have a kernel less than version 4.17, The CSI driver will fall back to use the FUSE client.

RBD: If you have set the mounter: rbd-nbd option in the RBD storage class, the NBD mounter will have this issue.

If you are affected by this issue, you will need to proceed carefully during the upgrade to restart your application pods. The recommended step is to modify the update strategy of the CSI nodeplugin daemonsets to OnDelete so that you can control when the CSI driver pods are restarted on each node.

To avoid this issue in future upgrades, we recommend that you do not use the fuse client or rbd-nbd as of now.

This guide will walk you through the steps to upgrade the software in a cluster from v1.2.x to v2.0.0

Upgrading from v1.2.x to v2.0.0

Ceph-csi releases from master are expressly unsupported. It is strongly recommended that you use official releases of Ceph-csi. Unreleased versions from the master branch are subject to changes and incompatibilities that will not be supported in the official releases. Builds from the master branch can have functionality changed and even removed at any time without compatibility support and without prior notice.

git checkout release v2.0.0 branch

[$] git clone https://github.com/ceph/ceph-csi.git
[$] git check v2.0.0
[$] cd ./ceph-csi

Note: While upgrading please Ignore warning messages from kubectl output

Warning: kubectl apply should be used on resource created by either kubectl create --save-config or kubectl apply

Upgrading CephFS

Upgrading cephfs csi includes upgrade of cephfs driver and as well as kubernetes sidecar containers and also the permissions required for the kubernetes sidecar containers, lets upgrade the things one by one

1. Upgrade CephFS Provisioner resources

Upgrade provisioner resources include updating the provisioner RBAC and Provisioner deployment

1.1 Update the CephFS Provisioner RBAC
[$] kubectl apply -f deploy/cephfs/kubernetes/csi-provisioner-rbac.yaml
serviceaccount/cephfs-csi-provisioner configured
clusterrole.rbac.authorization.k8s.io/cephfs-external-provisioner-runner configured
clusterrole.rbac.authorization.k8s.io/cephfs-external-provisioner-runner-rules configured
clusterrolebinding.rbac.authorization.k8s.io/cephfs-csi-provisioner-role configured
role.rbac.authorization.k8s.io/cephfs-external-provisioner-cfg configured
rolebinding.rbac.authorization.k8s.io/cephfs-csi-provisioner-role-cfg configured
1.2 Update the CephFS Provisioner deployment
[$]kubectl apply -f deploy/cephfs/kubernetes/csi-cephfsplugin-provisioner.yaml
service/csi-cephfsplugin-provisioner configured
deployment.apps/csi-cephfsplugin-provisioner configured

wait for the deployment to complete

[$]kubectl get deployment
NAME                           READY   UP-TO-DATE   AVAILABLE   AGE
csi-cephfsplugin-provisioner   3/3     1            3           104m

deployment UP-TO-DATE value must be same as READY

2. Upgrade CephFS Nodeplugin resources

Upgrading nodeplugin resources include updating the nodeplugin RBAC and nodeplugin daemonset

2.1 Update the CephFS Nodeplugin RBAC
[$]kubectl apply -f deploy/cephfs/kubernetes/csi-nodeplugin-rbac.yaml
serviceaccount/cephfs-csi-nodeplugin configured
clusterrole.rbac.authorization.k8s.io/cephfs-csi-nodeplugin configured
clusterrole.rbac.authorization.k8s.io/cephfs-csi-nodeplugin-rules configured
clusterrolebinding.rbac.authorization.k8s.io/cephfs-csi-nodeplugin configured

If you determined in Pre-upgrade considerations that you were affected by the CSI driver restart issue that disconnects the application pods from their mounts, continue with this section. Otherwise, you can skip to step 2.2

vi deploy/cephfs/kubernetes/csi-cephfsplugin.yaml

kind: DaemonSet
apiVersion: apps/v1
metadata:
  name: csi-cephfsplugin
spec:
  selector:
    matchLabels:
      app: csi-cephfsplugin
  updateStrategy:
    type: OnDelete
  template:
    metadata:
      labels:
        app: csi-cephfsplugin
    spec:
      serviceAccount: cephfs-csi-nodeplugin

in the above template we have added updateStrategy and its type to the daemonset spec

2.2 Update the CephFS Nodeplugin daemonset
[$]kubectl apply -f deploy/cephfs/kubernetes/csi-cephfsplugin.yaml
daemonset.apps/csi-cephfsplugin configured
service/csi-metrics-cephfsplugin configured
2.3 Manual deletion of CephFS Nodeplugin daemonset pods

If you determined in Pre-upgrade considerations that you were affected by the CSI driver restart issue that disconnects the application pods from their mounts, continue with this section. Otherwise, you can skip this section.

As we have set the updateStrategy to OnDelete the CSI driver pods will not be updated until you delete them manually. This allows you to control when your application pods will be affected by the CSI driver restart.

For each node:

  • Drain your application pods from the node
  • Delete the CSI driver pods on the node
    • The pods to delete will be named with a csi-cephfsplugin prefix and have a random suffix on each node. However, no need to delete the provisioner pods: csi-cephfsplugin-provisioner-* .
    • The pod deletion causes the pods to be restarted and updated automatically on the node.

we have successfully upgraded cephfs csi from v1.2.2 to v2.0.0

Upgrading RBD

Upgrading rbd csi includes upgrade of rbd driver and as well as kubernetes sidecar containers and also the permissions required for the kubernetes sidecar containers, lets upgrade the things one by one

3. Upgrade RBD Provisioner resources

Upgrading provisioner resources include updating the provisioner RBAC and Provisioner deployment

3.1 Update the RBD Provisioner RBAC
[$]kubectl apply -f deploy/rbd/kubernetes/csi-provisioner-rbac.yaml
serviceaccount/rbd-csi-provisioner configured
clusterrole.rbac.authorization.k8s.io/rbd-external-provisioner-runner configured
clusterrole.rbac.authorization.k8s.io/rbd-external-provisioner-runner-rules configured
clusterrolebinding.rbac.authorization.k8s.io/rbd-csi-provisioner-role configured
role.rbac.authorization.k8s.io/rbd-external-provisioner-cfg configured
rolebinding.rbac.authorization.k8s.io/rbd-csi-provisioner-role-cfg configured
3.2 Update the RBD Provisioner deployment
[$]kubectl apply -f deploy/rbd/kubernetes/csi-rbdplugin-provisioner.yaml
service/csi-rbdplugin-provisioner configured
deployment.apps/csi-rbdplugin-provisioner configured

wait for the deployment to complete

[$]kubectl get deployments
NAME                           READY   UP-TO-DATE   AVAILABLE   AGE
csi-rbdplugin-provisioner      3/3     3            3           139m

deployment UP-TO-DATE value must be same as READY

4. Upgrade RBD Nodeplugin resources

Upgrading nodeplugin resources include updating the nodeplugin RBAC and nodeplugin daemonset

4.1 Update the RBD Nodeplugin RBAC
[$]kubectl apply -f deploy/rbd/kubernetes/csi-nodeplugin-rbac.yaml
serviceaccount/rbd-csi-nodeplugin configured
clusterrole.rbac.authorization.k8s.io/rbd-csi-nodeplugin configured
clusterrole.rbac.authorization.k8s.io/rbd-csi-nodeplugin-rules configured
clusterrolebinding.rbac.authorization.k8s.io/rbd-csi-nodeplugin configured

If you determined in Pre-upgrade considerations that you were affected by the CSI driver restart issue that disconnects the application pods from their mounts, continue with this section. Otherwise, you can skip to step 4.2

vi deploy/rbd/kubernetes/csi-rbdplugin.yaml

kind: DaemonSet
apiVersion: apps/v1
metadata:
  name: csi-rbdplugin
spec:
  selector:
    matchLabels:
      app: csi-rbdplugin
  updateStrategy:
    type: OnDelete
  template:
    metadata:
      labels:
        app: csi-rbdplugin
    spec:
      serviceAccount: rbd-csi-nodeplugin

in the above template we have added updateStrategy and its type to the daemonset spec

4.2 Update the RBD Nodeplugin daemonset
[$]kubectl apply -f deploy/rbd/kubernetes/csi-rbdplugin.yaml
daemonset.apps/csi-rbdplugin configured
service/csi-metrics-rbdplugin configured
4.3 Manual deletion of RBD Nodeplugin daemonset pods

If you determined in Pre-upgrade considerations that you were affected by the CSI driver restart issue that disconnects the application pods from their mounts, continue with this section. Otherwise, you can skip this section.

As we have set the updateStrategy to OnDelete the CSI driver pods will not be updated until you delete them manually. This allows you to control when your application pods will be affected by the CSI driver restart.

For each node:

  • Drain your application pods from the node
  • Delete the CSI driver pods on the node
    • The pods to delete will be named with a csi-rbdplugin prefix and have a random suffix on each node. However, no need to delete the provisioner pods csi-rbdplugin-provisioner-* .
    • The pod deletion causes the pods to be restarted and updated automatically on the node.

we have successfully upgraded RBD csi from v1.2.2 to v2.0.0

Handling node reboot hangs due to existing network mounts

With prior versions of ceph-csi a node reboot may hang the reboot process. This is due to existing cephfs or rbd network mounts not carrying the _netdev mount option. The missing mount option is fixed, and future mounts using ceph-csi would automatically carry the required flags needed for a clean reboot.

It is suggested to upgrade to the latest ceph-csi release and drain application pods on all the nodes so that new mount option _netdev can be added to all the mountpoints.