Ceph
PLEASE NOTE: This document applies to v0.7 version and not to the latest stable release v1.8
Documentation for other releases can be found by using the version selector in the top right of any doc page.Shared File System
A shared file system can be mounted read-write from multiple pods. This may be useful for applications which can be clustered using a shared filesystem.
This example runs a shared file system for the kube-registry.
Prerequisites
This guide assumes you have created a Rook cluster as explained in the main Kubernetes guide
Create the File System
Create the file system by specifying the desired settings for the metadata pool, data pools, and metadata server in the Filesystem
CRD. In this example we create the metadata pool with replication of three and a single data pool with erasure coding. For more options, see the documentation on creating shared file systems.
Save this shared file system definition as rook-filesystem.yaml
:
apiVersion: rook.io/v1alpha1
kind: Filesystem
metadata:
name: myfs
namespace: rook
spec:
metadataPool:
replicated:
size: 3
dataPools:
- erasureCoded:
dataChunks: 2
codingChunks: 1
metadataServer:
activeCount: 1
activeStandby: true
Now let’s create the file system. The Rook operator will create all the pools and other resources necessary to start the service. This may take a minute to complete.
# Create the file system
kubectl create -f rook-filesystem.yaml
# To confirm the file system is configured, wait for the mds pods to start
kubectl -n rook get pod -l app=rook-ceph-mds
To see detailed status of the file system, 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.
$ ceph status
...
services:
mds: myfs-1/1/1 up {[myfs:0]=mzw58b=up:active}, 1 up:standby-replay
Consume the file system
As an example, we will start the kube-registry pod with the shared file system as the backing store.
Save the following spec as kube-registry.yaml
:
apiVersion: v1
kind: ReplicationController
metadata:
name: kube-registry-v0
namespace: kube-system
labels:
k8s-app: kube-registry
version: v0
kubernetes.io/cluster-service: "true"
spec:
replicas: 3
selector:
k8s-app: kube-registry
version: v0
template:
metadata:
labels:
k8s-app: kube-registry
version: v0
kubernetes.io/cluster-service: "true"
spec:
containers:
- name: registry
image: registry:2
resources:
limits:
cpu: 100m
memory: 100Mi
env:
- name: REGISTRY_HTTP_ADDR
value: :5000
- name: REGISTRY_STORAGE_FILESYSTEM_ROOTDIRECTORY
value: /var/lib/registry
volumeMounts:
- name: image-store
mountPath: /var/lib/registry
ports:
- containerPort: 5000
name: registry
protocol: TCP
volumes:
- name: image-store
flexVolume:
driver: rook.io/rook
fsType: ceph
options:
fsName: myfs # name of the filesystem specified in the filesystem CRD.
clusterName: rook # namespace where the Rook cluster is deployed
# by default the path is /, but you can override and mount a specific path of the filesystem by using the path attribute
# path: /some/path/inside/cephfs
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.
Test the storage
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 file system in the toolbox pod.
Start and connect to the Rook toolbox.
# Mount the same filesystem that the kube-registry is using
mkdir /tmp/registry
rookctl filesystem mount --name myfs --path /tmp/registry
# If you have pushed images to the registry you will see a directory called docker
ls /tmp/registry
# Cleanup the filesystem mount
rookctl filesystem unmount --path /tmp/registry
rmdir /tmp/registry
Teardown
To clean up all the artifacts created by the file system demo:
kubectl -n kube-system delete secret rook-admin
kubectl delete -f kube-registry.yaml