Filesystem Storage Overview
A filesystem storage (also named shared filesystem) can be mounted with read/write permission from multiple pods. This may be useful for applications which can be clustered using a shared filesystem.
This example runs a shared filesystem for the kube-registry.
This guide assumes you have created a Rook cluster as explained in the main quickstart guide
Multiple Filesystems Support¶
Multiple filesystems are supported as of the Ceph Pacific release.
Create the Filesystem¶
Create the filesystem by specifying the desired settings for the metadata pool, data pools, and metadata server in the
CephFilesystem CRD. In this example we create the metadata pool with replication of three and a single data pool with replication of three. For more options, see the documentation on creating shared filesystems.
Save this shared filesystem definition as
The Rook operator will create all the pools and other resources necessary to start the service. This may take a minute to complete.
To confirm the filesystem is configured, wait for the mds pods to start:
To see detailed status of the filesystem, start and connect to the Rook toolbox. A new line will be shown with
ceph status for the
mds service. In this example, there is one active instance of MDS which is up, with one MDS instance in
standby-replay mode in case of failover.
Before Rook can start provisioning storage, a StorageClass needs to be created based on the filesystem. This is needed for Kubernetes to interoperate with the CSI driver to create persistent volumes.
Save this storage class definition as
If you've deployed the Rook operator in a namespace other than "rook-ceph" as is common change the prefix in the provisioner to match the namespace you used. For example, if the Rook operator is running in "rook-op" the provisioner value should be "rook-op.rbd.csi.ceph.com".
Create the storage class.
The CephFS CSI driver uses quotas to enforce the PVC size requested.
Only newer kernels support CephFS quotas (kernel version of at least 4.17). If you require quotas to be enforced and the kernel driver does not support it, you can disable the kernel driver and use the FUSE client. This can be done by setting
CSI_FORCE_CEPHFS_KERNEL_CLIENT: false in the operator deployment (
operator.yaml). However, it is important to know that when the FUSE client is enabled, there is an issue that during upgrade the application pods will be disconnected from the mount and will need to be restarted. See the upgrade guide for more details.
Consume the Shared Filesystem: K8s Registry Sample¶
As an example, we will start the kube-registry pod with the shared filesystem as the backing store. Save the following spec as
Create the Kube registry deployment:
You now have a docker registry which is HA with persistent storage.
Kernel Version Requirement¶
If the Rook cluster has more than one filesystem and the application pod is scheduled to a node with kernel version older than 4.7, inconsistent results may arise since kernels older than 4.7 do not support specifying filesystem namespaces.
Consume the Shared Filesystem: Toolbox¶
Once you have pushed an image to the registry (see the instructions to expose and use the kube-registry), verify that kube-registry is using the filesystem that was configured above by mounting the shared filesystem in the toolbox pod. See the Direct Filesystem topic for more details.
Consume the Shared Filesystem across namespaces¶
A PVC that you create using the
rook-cephfs storageClass can be shared between different Pods simultaneously, either read-write or read-only, but is restricted to a single namespace (a PVC is a namespace-scoped resource, so you cannot use it in another one).
However there are some use cases where you want to share the content from a CephFS-based PVC among different Pods in different namespaces, for a shared library for example, or a collaboration workspace between applications running in different namespaces.
You can do that using the following recipe.
Shared volume creation¶
- In the
rooknamespace, create a copy of the secret
rook-csi-cephfs-node, name it
- Edit your new secret, changing the name of the keys (keep the value as it is):
- Create the PVC you want to share, for example:
- The corresponding PV that is created will have all the necessary info to connect to the CephFS volume (all non-necessary information are removed here):
On this PV, change the
Retainto avoid it from being deleted when you will delete PVCs. Don't forget to change it back to
Deletewhen you want to remove the shared volume (see full procedure in the next section).
Copy the YAML content of the PV, and create a new static PV with the same information and some modifications. From the original YAML, you must:
- Modify the original name. To keep track, the best solution is to append to the original name the namespace name where you want your new PV. In this example
- Modify the volumeHandle. Again append the targeted namespace.
- Add the
staticVolume: "true"entry to the volumeAttributes.
- Add the rootPath entry to the volumeAttributes, with the same content as
- In the
nodeStageSecretRefsection, change the name to point to the secret you created earlier,
- Remove the unnecessary information before applying the YAML (claimRef, managedFields,...):
Your YAML should look like this:
- In a new or other namespace, create a new PVC that will use this new PV you created. You simply have to point to it in the
volumeNameparameter. Make sure you enter the same size as the original PVC!:
You have now access to the same CephFS subvolume from different PVCs in different namespaces. Redo the previous steps (copy PV with a new name, create a PVC pointing to it) in each namespace you want to use this subvolume.
Note: the new PVCs/PVs we have created are static. Therefore CephCSI does not support snapshots, clones, resizing or delete operations for them. If those operations are required, you must make them on the original PVC.
Shared volume removal¶
As the same CephFS volume is used by different PVCs/PVs, you must proceed very orderly to remove it properly.
- Delete the static PVCs in the different namespaces, but keep the original one!
- Delete the corresponding static PVs that should now have been marked as "Released". Again, don't delete the original one yet!
- Edit the original PV, changing back the
- Delete the original PVC. It will now properly delete the original PV, as well as the subvolume in CephFS.
Due to this bug, the global mount for a Volume that is mounted multiple times on the same node will not be unmounted. This does not result in any particular problem, apart from polluting the logs with unmount error messages, or having many different mounts hanging if you create and delete many shared PVCs, or you don't really use them.
Until this issue is solved, either on the Rook or Kubelet side, you can always manually unmount the unwanted hanging global mounts on the nodes:
- Log onto each node where the volume has been mounted.
- Check for hanging mounts using their
- Unmount the unwanted volumes.
To clean up all the artifacts created by the filesystem demo:
To delete the filesystem components and backing data, delete the Filesystem CRD.
Data will be deleted if preserveFilesystemOnDelete=false**.
Note: If the "preserveFilesystemOnDelete" filesystem attribute is set to true, the above command won't delete the filesystem. Recreating the same CRD will reuse the existing filesystem.
Advanced Example: Erasure Coded Filesystem¶
The Ceph filesystem example can be found here: Ceph Shared Filesystem - Samples - Erasure Coded.