Kubernetes集群跨节点挂载CephFS
在Kubernetes集群中运行有状态服务或应用总是不那么容易的。比如,之前我在项目中使用了CephRBD,虽然遇到过几次问题,但总体算是运行良好。但最近发现CephRBD无法满足跨节点挂载的需求,我只好另辟蹊径。由于CephFS和CephRBD师出同门,它自然成为了这次我首要考察的目标。这里将跨节点挂载CephFS的考察过程记录一下,一是备忘,二则也可以为其他有相似需求的朋友提供些资料。
一、CephRBD的问题
这里先提一嘴CephRBD的问题。最近项目中有这样的需求:让集群中的Pod共享外部分布式存储,即多个Pod共同挂载一份存储,实现存储共享,这样可大大简化系统设计和复杂性。之前CephRBD都是挂载到一个Pod中运行的,CephRBD是否支持多Pod同时挂载呢?官方文档中给出了否定的答案: 基于CephRBD的Persistent Volume仅支持两种accessmode:
ReadWriteOnce和ReadOnlyMany,不支持ReadWriteMany。这样对于有读写需求的Pod来说,一个CephRBD pv仅能被一个node挂载一次。
我们来验证一下这个“不幸的”事实。
我们首先创建一个测试用的image:foo1。这里我利用了项目里写的CephRBD API服务,也可通过ceph命令手工创建:
# curl -v -H "Content-type: application/json" -X POST -d '{"kind": "Images","apiVersion": "v1", "metadata": {"name": "foo1", "capacity": 512} ' http://192.168.3.22:8080/api/v1/pools/rbd/images
... ...
{
"errcode": 0,
"errmsg": "ok"
}
# curl http://192.168.3.22:8080/api/v1/pools/rbd/images
{
"Kind": "ImagesList",
"APIVersion": "v1",
"Items": [
{
"name": "foo1"
}
]
}
利用下面文件创建pv和pvc:
//ceph-pv.yaml
apiVersion: v1
kind: PersistentVolume
metadata:
name: foo-pv
spec:
capacity:
storage: 512Mi
accessModes:
- ReadWriteMany
rbd:
monitors:
- ceph_monitor_ip:port
pool: rbd
image: foo1
user: admin
secretRef:
name: ceph-secret
fsType: ext4
readOnly: false
persistentVolumeReclaimPolicy: Recycle
//ceph-pvc.yaml
kind: PersistentVolumeClaim
apiVersion: v1
metadata:
name: foo-claim
spec:
accessModes:
- ReadWriteMany
resources:
requests:
storage: 512Mi
创建后:
# kubectl get pv
[NAME CAPACITY ACCESSMODES RECLAIMPOLICY STATUS CLAIM REASON AGE
foo-pv 512Mi RWO Recycle Bound default/foo-claim 20h
# kubectl get pvc
NAME STATUS VOLUME CAPACITY ACCESSMODES AGE
foo-claim Bound foo-pv 512Mi RWO 20h
创建挂载上述image的Pod:
// ceph-pod2.yaml
apiVersion: v1
kind: Pod
metadata:
name: ceph-pod2
spec:
containers:
- name: ceph-ubuntu2
image: ubuntu:14.04
command: ["tail", "-f", "/var/log/bootstrap.log"]
volumeMounts:
- name: ceph-vol2
mountPath: /mnt/cephrbd/data
readOnly: false
volumes:
- name: ceph-vol2
persistentVolumeClaim:
claimName: foo-claim
创建成功后,我们可以查看挂载目录的数据:
# kubectl exec ceph-pod2 ls /mnt/cephrbd/data
1.txt
lost+found
我们在同一个kubernetes node上再启动一个pod(可以把上面的ceph-pod2.yaml的pod name改为ceph-pod3),挂载同样的pv:
NAMESPACE NAME READY STATUS RESTARTS AGE IP NODE
default ceph-pod2 1/1 Running 0 3m 172.16.57.9 xx.xx.xx.xx
default ceph-pod3 1/1 Running 0 0s 172.16.57.10 xx.xx.xx.xx
# kubectl exec ceph-pod3 ls /mnt/cephrbd/data
1.txt
lost+found
我们通过ceph-pod2写一个文件,在ceph-pod3中将其读出:
# kubectl exec ceph-pod2 -- bash -c "for i in {1..10}; do sleep 1; echo 'pod2: Hello, World'>> /mnt/cephrbd/data/foo.txt ; done "
root@node1:~/k8stest/k8s-cephrbd/footest# kubectl exec ceph-pod3 cat /mnt/cephrbd/data/foo.txt
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
pod2: Hello, World
到目前为止,在一个node上多个Pod是可以以ReadWrite模式挂载同一个CephRBD的。
我们在另外一个节点启动一个试图挂载该pv的Pod,该Pod启动后一直处于pending状态,通过kubectl describe查看其详细信息,可以看到:
Events:
FirstSeen LastSeen Count From SubobjectPath Type Reason Message
--------- -------- ----- ---- ------------- -------- ------ -------
.. ...
2m 37s 2 {kubelet yy.yy.yy.yy} Warning FailedMount Unable to mount volumes for pod "ceph-pod2-master_default(a45f62aa-2bc3-11e7-9baa-00163e1625a9)": timeout expired waiting for volumes to attach/mount for pod "ceph-pod2-master"/"default". list of unattached/unmounted volumes=[ceph-vol2]
2m 37s 2 {kubelet yy.yy.yy.yy} Warning FailedSync Error syncing pod, skipping: timeout expired waiting for volumes to attach/mount for pod "ceph-pod2-master"/"default". list of unattached/unmounted volumes=[ceph-vol2]
查看kubelet.log中的错误日志:
I0428 11:39:15.737729 1241 reconciler.go:294] MountVolume operation started for volume "kubernetes.io/rbd/a45f62aa-2bc3-11e7-9baa-00163e1625a9-foo-pv" (spec.Name: "foo-pv") to pod "a45f62aa-2bc3-11e7-9baa-00163e1625a9" (UID: "a45f62aa-2bc3-11e7-9baa-00163e1625a9").
I0428 11:39:15.939183 1241 operation_executor.go:768] MountVolume.SetUp succeeded for volume "kubernetes.io/secret/923700ff-12c2-11e7-9baa-00163e1625a9-default-token-40z0x" (spec.Name: "default-token-40z0x") pod "923700ff-12c2-11e7-9baa-00163e1625a9" (UID: "923700ff-12c2-11e7-9baa-00163e1625a9").
E0428 11:39:17.039656 1241 disk_manager.go:56] failed to attach disk
E0428 11:39:17.039722 1241 rbd.go:228] rbd: failed to setup mount /var/lib/kubelet/pods/a45f62aa-2bc3-11e7-9baa-00163e1625a9/volumes/kubernetes.io~rbd/foo-pv rbd: image foo1 is locked by other nodes
E0428 11:39:17.039857 1241 nestedpendingoperations.go:254] Operation for "\"kubernetes.io/rbd/a45f62aa-2bc3-11e7-9baa-00163e1625a9-foo-pv\" (\"a45f62aa-2bc3-11e7-9baa-00163e1625a9\")" failed. No retries permitted until 2017-04-28 11:41:17.039803969 +0800 CST (durationBeforeRetry 2m0s). Error: MountVolume.SetUp failed for volume "kubernetes.io/rbd/a45f62aa-2bc3-11e7-9baa-00163e1625a9-foo-pv" (spec.Name: "foo-pv") pod "a45f62aa-2bc3-11e7-9baa-00163e1625a9" (UID: "a45f62aa-2bc3-11e7-9baa-00163e1625a9") with: rbd: image foo1 is locked by other nodes
可以看到“rbd: image foo1 is locked by other nodes”的日志。我们用试验证明了目前CephRBD仅能被k8s中的一个node挂载的事实。
二、Ceph集群安装mds以支持CephFS
这次我在两个Ubuntu 16.04的vm上新部署了一套Ceph,过程与之前第一次部署Ceph时大同小异,这里就不赘述了。要让Ceph支持CephFS,我们需要安装mds组件,有了前面的基础,通过ceph-deploy工具安装mds十分简单:
# ceph-deploy mds create yypdmaster yypdnode
[ceph_deploy.conf][DEBUG ] found configuration file at: /root/.cephdeploy.conf
[ceph_deploy.cli][INFO ] Invoked (1.5.37): /usr/bin/ceph-deploy mds create yypdmaster yypdnode
[ceph_deploy.cli][INFO ] ceph-deploy options:
[ceph_deploy.cli][INFO ] username : None
[ceph_deploy.cli][INFO ] verbose : False
[ceph_deploy.cli][INFO ] overwrite_conf : False
[ceph_deploy.cli][INFO ] subcommand : create
[ceph_deploy.cli][INFO ] quiet : False
[ceph_deploy.cli][INFO ] cd_conf : <ceph_deploy.conf.cephdeploy.Conf instance at 0x7f60fb5e71b8>
[ceph_deploy.cli][INFO ] cluster : ceph
[ceph_deploy.cli][INFO ] func : <function mds at 0x7f60fba4e140>
[ceph_deploy.cli][INFO ] ceph_conf : None
[ceph_deploy.cli][INFO ] mds : [('yypdmaster', 'yypdmaster'), ('yypdnode', 'yypdnode')]
[ceph_deploy.cli][INFO ] default_release : False
[ceph_deploy.mds][DEBUG ] Deploying mds, cluster ceph hosts yypdmaster:yypdmaster yypdnode:yypdnode
[yypdmaster][DEBUG ] connected to host: yypdmaster
[yypdmaster][DEBUG ] detect platform information from remote host
[yypdmaster][DEBUG ] detect machine type
[ceph_deploy.mds][INFO ] Distro info: Ubuntu 16.04 xenial
[ceph_deploy.mds][DEBUG ] remote host will use systemd
[ceph_deploy.mds][DEBUG ] deploying mds bootstrap to yypdmaster
[yypdmaster][DEBUG ] write cluster configuration to /etc/ceph/{cluster}.conf
[yypdmaster][DEBUG ] create path if it doesn't exist
[yypdmaster][INFO ] Running command: ceph --cluster ceph --name client.bootstrap-mds --keyring /var/lib/ceph/bootstrap-mds/ceph.keyring auth get-or-create mds.yypdmaster osd allow rwx mds allow mon allow profile mds -o /var/lib/ceph/mds/ceph-yypdmaster/keyring
[yypdmaster][INFO ] Running command: systemctl enable ceph-mds@yypdmaster
[yypdmaster][WARNIN] Created symlink from /etc/systemd/system/ceph-mds.target.wants/ceph-mds@yypdmaster.service to /lib/systemd/system/ceph-mds@.service.
[yypdmaster][INFO ] Running command: systemctl start ceph-mds@yypdmaster
[yypdmaster][INFO ] Running command: systemctl enable ceph.target
[yypdnode][DEBUG ] connected to host: yypdnode
[yypdnode][DEBUG ] detect platform information from remote host
[yypdnode][DEBUG ] detect machine type
[ceph_deploy.mds][INFO ] Distro info: Ubuntu 16.04 xenial
[ceph_deploy.mds][DEBUG ] remote host will use systemd
[ceph_deploy.mds][DEBUG ] deploying mds bootstrap to yypdnode
[yypdnode][DEBUG ] write cluster configuration to /etc/ceph/{cluster}.conf
[yypdnode][DEBUG ] create path if it doesn't exist
[yypdnode][INFO ] Running command: ceph --cluster ceph --name client.bootstrap-mds --keyring /var/lib/ceph/bootstrap-mds/ceph.keyring auth get-or-create mds.yypdnode osd allow rwx mds allow mon allow profile mds -o /var/lib/ceph/mds/ceph-yypdnode/keyring
[yypdnode][INFO ] Running command: systemctl enable ceph-mds@yypdnode
[yypdnode][WARNIN] Created symlink from /etc/systemd/system/ceph-mds.target.wants/ceph-mds@yypdnode.service to /lib/systemd/system/ceph-mds@.service.
[yypdnode][INFO ] Running command: systemctl start ceph-mds@yypdnode
[yypdnode][INFO ] Running command: systemctl enable ceph.target
非常顺利。安装后,可以在任意一个node上看到mds在运行:
# ps -ef|grep ceph
ceph 7967 1 0 17:23 ? 00:00:00 /usr/bin/ceph-osd -f --cluster ceph --id 1 --setuser ceph --setgroup ceph
ceph 15674 1 0 17:32 ? 00:00:00 /usr/bin/ceph-mon -f --cluster ceph --id yypdnode --setuser ceph --setgroup ceph
ceph 18019 1 0 17:35 ? 00:00:00 /usr/bin/ceph-mds -f --cluster ceph --id yypdnode --setuser ceph --setgroup ceph
mds是存储cephfs的元信息的,我的ceph是10.2.7版本:
# ceph -v
ceph version 10.2.7 (50e863e0f4bc8f4b9e31156de690d765af245185)
虽然支持多 active mds并行运行,但官方文档建议保持一个active mds,其他mds作为standby(见下面ceph集群信息中的fsmap部分):
# ceph -s
cluster ffac3489-d678-4caf-ada2-3dd0743158b6
... ...
fsmap e6: 1/1/1 up {0=yypdnode=up:active}, 1 up:standby
osdmap e19: 2 osds: 2 up, 2 in
flags sortbitwise,require_jewel_osds
pgmap v192498: 576 pgs, 5 pools, 126 MB data, 238 objects
44365 MB used, 31881 MB / 80374 MB avail
576 active+clean
三、创建fs并测试挂载
我们在ceph上创建一个fs:
# ceph osd pool create cephfs_data 128
pool 'cephfs_data' created
# ceph osd pool create cephfs_metadata 128
pool 'cephfs_metadata' created
# ceph fs new test_fs cephfs_metadata cephfs_data
new fs with metadata pool 2 and data pool 1
# ceph fs ls
name: test_fs, metadata pool: cephfs_metadata, data pools: [cephfs_data ]
不过,ceph当前正式版功能中仅支持一个fs,对多个fs的支持仅存在于实验feature中:
# ceph osd pool create cephfs1_data 128
# ceph osd pool create cephfs1_metadata 128
# ceph fs new test_fs1 cephfs1_metadata cephfs1_data
Error EINVAL: Creation of multiple filesystems is disabled. To enable this experimental feature, use 'ceph fs flag set enable_multiple true'
在物理机上挂载cephfs可以使用mount命令、mount.ceph(apt-get install ceph-fs-common)或ceph-fuse(apt-get install ceph-fuse),我们先用mount命令挂载:
我们将上面创建的cephfs挂载到主机的/mnt下:
#mount -t ceph ceph_mon_host:6789:/ /mnt -o name=admin,secretfile=admin.secret
# cat admin.secret //ceph.client.admin.keyring中的key
AQDITghZD+c/DhAArOiWWQqyMAkMJbWmHaxjgQ==
查看cephfs信息:
# df -h
ceph_mon_host:6789:/ 79G 45G 35G 57% /mnt
可以看出:cephfs将两个物理节点上的磁盘全部空间作为了自己的空间。
通过ceph-fuse挂载,还可以限制对挂载路径的访问权限,我们来创建用户foo,让其仅仅拥有对/ceph-volume1-test路径具有只读访问权限:
# ceph auth get-or-create client.foo mon 'allow *' mds 'allow r path=/ceph-volume1-test' osd 'allow *'
# ceph-fuse -n client.foo -m 10.47.217.91:6789 /mnt -r /ceph-volume1-test
ceph-fuse[10565]: starting ceph client2017-05-03 16:07:25.958903 7f1a14fbff00 -1 init, newargv = 0x557e350defc0 newargc=11
ceph-fuse[10565]: starting fuse
查看挂载路径,并尝试创建文件:
# cd /mnt
root@yypdnode:/mnt# ls
1.txt
root@yypdnode:/mnt# touch 2.txt
touch: cannot touch '2.txt': Permission denied
由于foo用户只拥有对 /ceph-volume1-test的只读权限,因此创建文件失败了!
四、Kubernetes跨节点挂载CephFS
在K8s中,至少可以通过两种方式挂载CephFS,一种是通过Pod直接挂载;另外一种则是通过pv和pvc挂载。我们分别来看。
1、Pod直接挂载CephFS
//ceph-pod2-with-secret.yaml
apiVersion: v1
kind: Pod
metadata:
name: ceph-pod2-with-secret
spec:
containers:
- name: ceph-ubuntu2
image: ubuntu:14.04
command: ["tail", "-f", "/var/log/bootstrap.log"]
volumeMounts:
- name: ceph-vol2
mountPath: /mnt/cephfs/data
readOnly: false
volumes:
- name: ceph-vol2
cephfs:
monitors:
- ceph_mon_host:6789
user: admin
secretFile: "/etc/ceph/admin.secret"
readOnly: false
注意:保证每个节点上都存在/etc/ceph/admin.secret文件。
查看Pod挂载的内容:
# docker ps|grep pod
bc96431408c7 ubuntu:14.04 "tail -f /var/log/boo" About a minute ago Up About a minute k8s_ceph-ubuntu2.66c44128_ceph-pod2-with-secret_default_3d8a05f8-33c3-11e7-bcd9-6640d35a0e90_fc483b8a
bcc65ab82069 gcr.io/google_containers/pause-amd64:3.0 "/pause" About a minute ago Up About a minute k8s_POD.d8dbe16c_ceph-pod2-with-secret_default_3d8a05f8-33c3-11e7-bcd9-6640d35a0e90_02381204
root@yypdnode:~# docker exec bc96431408c7 ls /mnt/cephfs/data
1.txt
apps
ceph-volume1-test
test1.txt
我们再在另外一个node上启动挂载同一个cephfs的Pod,看是否可以跨节点挂载:
# kubectl get pods
NAMESPACE NAME READY STATUS RESTARTS AGE IP NODE
default ceph-pod2-with-secret 1/1 Running 0 3m 172.30.192.2 iz2ze39jeyizepdxhwqci6z
default ceph-pod2-with-secret-on-master 1/1 Running 0 3s 172.30.0.51 iz25beglnhtz
... ...
# kubectl exec ceph-pod2-with-secret-on-master ls /mnt/cephfs/data
1.txt
apps
ceph-volume1-test
test1.txt
可以看到不同节点可以挂载同一CephFS。我们在一个pod中操作一下挂载的cephfs:
# kubectl exec ceph-pod2-with-secret-on-master -- bash -c "for i in {1..10}; do sleep 1; echo 'pod2-with-secret-on-master: Hello, World'>> /mnt/cephfs/data/foo.txt ; done "
root@yypdmaster:~/k8stest/cephfstest/footest# kubectl exec ceph-pod2-with-secret-on-master cat /mnt/cephfs/data/foo.txt
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
2、通过PV和PVC挂载CephFS
挂载cephfs的pv和pvc在写法方面与上面挂载rbd的类似:
//ceph-pv.yaml
apiVersion: v1
kind: PersistentVolume
metadata:
name: foo-pv
spec:
capacity:
storage: 512Mi
accessModes:
- ReadWriteMany
cephfs:
monitors:
- ceph_mon_host:6789
path: /
user: admin
secretRef:
name: ceph-secret
readOnly: false
persistentVolumeReclaimPolicy: Recycle
//ceph-pvc.yaml
kind: PersistentVolumeClaim
apiVersion: v1
metadata:
name: foo-claim
spec:
accessModes:
- ReadWriteMany
resources:
requests:
storage: 512Mi
使用pvc的pod:
//ceph-pod2.yaml
apiVersion: v1
kind: Pod
metadata:
name: ceph-pod2
spec:
containers:
- name: ceph-ubuntu2
image: ubuntu:14.04
command: ["tail", "-f", "/var/log/bootstrap.log"]
volumeMounts:
- name: ceph-vol2
mountPath: /mnt/cephfs/data
readOnly: false
volumes:
- name: ceph-vol2
persistentVolumeClaim:
claimName: foo-claim
创建pv、pvc:
# kubectl create -f ceph-pv.yaml
persistentvolume "foo-pv" created
# kubectl create -f ceph-pvc.yaml
persistentvolumeclaim "foo-claim" created
# kubectl get pvc
NAME STATUS VOLUME CAPACITY ACCESSMODES AGE
foo-claim Bound foo-pv 512Mi RWX 4s
# kubectl get pv
NAME CAPACITY ACCESSMODES RECLAIMPOLICY STATUS CLAIM REASON AGE
foo-pv 512Mi RWX Recycle Bound default/foo-claim 24s
启动pod,通过exec命令查看挂载情况:
# docker ps|grep pod
a6895ec0274f ubuntu:14.04 "tail -f /var/log/boo" About a minute ago Up About a minute k8s_ceph-ubuntu2.66c44128_ceph-pod2_default_4e4fc8d4-33c6-11e7-bcd9-6640d35a0e90_1b37ed76
52b6811a6584 gcr.io/google_containers/pause-amd64:3.0 "/pause" About a minute ago Up About a minute k8s_POD.d8dbe16c_ceph-pod2_default_4e4fc8d4-33c6-11e7-bcd9-6640d35a0e90_27e5f988
55b96edbf4bf ubuntu:14.04 "tail -f /var/log/boo" 14 minutes ago Up 14 minutes k8s_ceph-ubuntu2.66c44128_ceph-pod2-with-secret_default_9d383b0c-33c4-11e7-bcd9-6640d35a0e90_1656e5e0
f8b699bc0459 gcr.io/google_containers/pause-amd64:3.0 "/pause" 14 minutes ago Up 14 minutes k8s_POD.d8dbe16c_ceph-pod2-with-secret_default_9d383b0c-33c4-11e7-bcd9-6640d35a0e90_effdfae7
root@yypdnode:~# docker exec a6895ec0274f ls /mnt/cephfs/data
1.txt
apps
ceph-volume1-test
foo.txt
test1.txt
# docker exec a6895ec0274f cat /mnt/cephfs/data/foo.txt
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
pod2-with-secret-on-master: Hello, World
五、pv的状态
如果你不删除pvc,一切都安然无事:
# kubectl get pv
NAME CAPACITY ACCESSMODES RECLAIMPOLICY STATUS CLAIM REASON AGE
foo-pv 512Mi RWX Recycle Bound default/foo-claim 1h
# kubectl get pvc
NAME STATUS VOLUME CAPACITY ACCESSMODES AGE
foo-claim Bound foo-pv 512Mi RWX 1h
但是如果删除pvc,pv的状态将变成failed:
删除pvc:
# kubectl get pv
NAME CAPACITY ACCESSMODES RECLAIMPOLICY STATUS CLAIM REASON AGE
foo-pv 512Mi RWX Recycle Failed default/foo-claim 2h
# kubectl describe pv/foo-pv
Name: foo-pv
Labels: <none>
Status: Failed
Claim: default/foo-claim
Reclaim Policy: Recycle
Access Modes: RWX
Capacity: 512Mi
Message: No recycler plugin found for the volume!
Source:
Type: RBD (a Rados Block Device mount on the host that shares a pod's lifetime)
CephMonitors: [xx.xx.xx.xx:6789]
RBDImage: foo1
FSType: ext4
RBDPool: rbd
RadosUser: admin
Keyring: /etc/ceph/keyring
SecretRef: &{ceph-secret}
ReadOnly: false
Events:
FirstSeen LastSeen Count From SubobjectPath Type Reason Message
--------- -------- ----- ---- ------------- -------- ------ -------
29s 29s 1 {persistentvolume-controller } Warning VolumeFailedRecycle No recycler plugin found for the volume!
我们在pv中指定的persistentVolumeReclaimPolicy是Recycle,但无论是cephrbd还是cephfs都不没有对应的recycler plugin,导致pv的status变成了failed,只能手工删除重建。
评论