Object Store Multisite

Multisite is a feature of Ceph that allows object stores to replicate their data over multiple Ceph clusters.

Multisite also allows object stores to be independent and isolated from other object stores in a cluster.

When a ceph-object-store is created without the zone section; a realm, zone group, and zone is created with the same name as the ceph-object-store.

Since it is the only ceph-object-store in the realm, the data in the ceph-object-store remain independent and isolated from others on the same cluster.

When a ceph-object-store is created with the zone section, the ceph-object-store will join a custom created zone, zone group, and realm each with a different names than its own.

This allows the ceph-object-store to replicate its data over multiple Ceph clusters.

To review core multisite concepts please read the ceph-multisite design overview.

Prerequisites¶

This guide assumes a Rook cluster as explained in the Quickstart.

Creating Object Multisite¶

If an admin wants to set up multisite on a Rook Ceph cluster, the following resources must be created:

A realm
A zonegroup
A zone
A ceph object store with the zone section

object-multisite.yaml in the examples directory can be used to create the multisite CRDs.

1	`kubectl create -f object-multisite.yaml`

The first zone group created in a realm is the master zone group. The first zone created in a zone group is the master zone.

When a non-master zone or non-master zone group is created, the zone group or zone is not in the Ceph Radosgw Multisite Period until an object-store is created in that zone (and zone group).

The zone will create the pools for the object-store(s) that are in the zone to use.

When one of the multisite CRs (realm, zone group, zone) is deleted the underlying ceph realm/zone group/zone is not deleted, neither are the pools created by the zone. See the "Multisite Cleanup" section for more information.

For more information on the multisite CRDs, see the related CRDs:

Pulling a Realm¶

If an admin wants to sync data from another cluster, the admin needs to pull a realm on a Rook Ceph cluster from another Rook Ceph (or Ceph) cluster.

To begin doing this, the admin needs 2 pieces of information:

An endpoint from the realm being pulled from
The access key and the system key of the system user from the realm being pulled from.

Getting the Pull Endpoint¶

To pull a Ceph realm from a remote Ceph cluster, an endpoint must be added to the CephObjectRealm's pull section in the spec. This endpoint must be from the master zone in the master zone group of that realm.

If an admin does not know of an endpoint that fits this criteria, the admin can find such an endpoint on the remote Ceph cluster (via the tool box if it is a Rook Ceph Cluster) by running:

$ radosgw-admin zonegroup get --rgw-realm=$REALM_NAME --rgw-zonegroup=$MASTER_ZONEGROUP_NAME
{
    ...
    "endpoints": [http://10.17.159.77:80],
    ...
}

A list of endpoints in the master zone group in the master zone is in the endpoints section of the JSON output of the zonegoup get command.

This endpoint must also be resolvable from the new Rook Ceph cluster. To test this run the curl command on the endpoint:

$ curl -L http://10.17.159.77:80
<?xml version="1.0" encoding="UTF-8"?><ListAllMyBucketsResult xmlns="http://s3.amazonaws.com/doc/2006-03-01/"><Owner><ID>anonymous</ID><DisplayName></DisplayName></Owner><Buckets></Buckets></ListAllMyBucketsResult>

Finally add the endpoint to the pull section of the CephObjectRealm's spec. The CephObjectRealm should have the same name as the CephObjectRealm/Ceph realm it is pulling from.

apiVersion: ceph.rook.io/v1
kind: CephObjectRealm
metadata:
  name: realm-a
  namespace: rook-ceph
spec:
  pull:
    endpoint: http://10.17.159.77:80

Getting Realm Access Key and Secret Key¶

The access key and secret key of the system user are keys that allow other Ceph clusters to pull the realm of the system user.

Getting the Realm Access Key and Secret Key from the Rook Ceph Cluster¶

System User for Multisite¶

When an admin creates a ceph-object-realm a system user automatically gets created for the realm with an access key and a secret key.

This system user has the name "$REALM_NAME-system-user". For the example if realm name is realm-a, then uid for the system user is "realm-a-system-user".

These keys for the user are exported as a kubernetes secret called "$REALM_NAME-keys" (ex: realm-a-keys).

This system user used by RGW internally for the data replication.

Getting keys from k8s secret¶

To get these keys from the cluster the realm was originally created on, run:

1	`kubectl -n $ORIGINAL_CLUSTER_NAMESPACE get secrets realm-a-keys -o yaml > realm-a-keys.yaml`

Edit the realm-a-keys.yaml file, and change the namespace with the namespace that the new Rook Ceph cluster exists in.

Then create a kubernetes secret on the pulling Rook Ceph cluster with the same secrets yaml file.

1	`kubectl create -f realm-a-keys.yaml`

Getting the Realm Access Key and Secret Key from a Non Rook Ceph Cluster¶

The access key and the secret key of the system user can be found in the output of running the following command on a non-rook ceph cluster:

1	`radosgw-admin user info --uid="realm-a-system-user"`

{
    ...
    "keys": [
        {
            "user": "realm-a-system-user"
            "access_key": "aSw4blZIKV9nKEU5VC0="
            "secret_key": "JSlDXFt5TlgjSV9QOE9XUndrLiI5JEo9YDBsJg==",
        }
    ],
    ...
}

Then base64 encode the each of the keys and create a .yaml file for the Kubernetes secret from the following template.

Only the access-key, secret-key, and namespace sections need to be replaced.

apiVersion: v1
data:
  access-key: YVN3NGJsWklLVjluS0VVNVZDMD0=
  secret-key: SlNsRFhGdDVUbGdqU1Y5UU9FOVhVbmRyTGlJNUpFbzlZREJzSmc9PQ==
kind: Secret
metadata:
  name: realm-a-keys
  namespace: $NEW_ROOK_CLUSTER_NAMESPACE
type: kubernetes.io/rook

Finally, create a kubernetes secret on the pulling Rook Ceph cluster with the new secrets yaml file.

1	`kubectl create -f realm-a-keys.yaml`

Pulling a Realm on a New Rook Ceph Cluster¶

Once the admin knows the endpoint and the secret for the keys has been created, the admin should create:

A CephObjectRealm matching to the realm on the other Ceph cluster, with an endpoint as described above.
A CephObjectZoneGroup matching the master zone group name or the master CephObjectZoneGroup from the cluster the realm was pulled from.
A CephObjectZone referring to the CephObjectZoneGroup created above.
A CephObjectStore referring to the new CephObjectZone resource.

object-multisite-pull-realm.yaml (with changes) in the examples directory can be used to create the multisite CRDs.

1	`kubectl create -f object-multisite-pull-realm.yaml`

Scaling a Multisite¶

Scaling the number of gateways that run the synchronization thread to 2 or more can increase the latency of the replication of each S3 object. The recommended way to scale a multisite configuration is to dissociate the gateway dedicated to the synchronization from gateways that serve clients.

The two types of gateways can be deployed by creating two CephObjectStores associated with the same CephObjectZone. The objectstore that deploys the gateway dedicated to the synchronization must have spec.gateway.instances set to 1, while the objectstore that deploys the client gateways have multiple replicas and should disable the synchronization thread on the gateways by setting spec.gateway.disableMultisiteSyncTraffic to true.

---
apiVersion: ceph.rook.io/v1
kind: CephObjectStore
metadata:
  name: replication
  namespace: rook-ceph
spec:
  gateway:
    port: 80
    instances: 1
    disableMultisiteSyncTraffic: false
  zone:
    name: zone-a
---
apiVersion: ceph.rook.io/v1
kind: CephObjectStore
metadata:
  name: clients
  namespace: rook-ceph
spec:
  gateway:
    port: 80
    instances: 5
    disableMultisiteSyncTraffic: true
  zone:
    name: zone-a

Multisite Cleanup¶

Multisite configuration must be cleaned up by hand. Deleting a realm/zone group/zone CR will not delete the underlying Ceph realm, zone group, zone, or the pools associated with a zone.

Deleting and Reconfiguring the Ceph Object Zone¶

Changes made to the resource's configuration or deletion of the resource are not reflected on the Ceph cluster.

When the ceph-object-zone resource is deleted or modified, the zone is not deleted from the Ceph cluster. Zone deletion must be done through the toolbox.

Changing the Master Zone¶

The Rook toolbox can change the master zone in a zone group.

radosgw-admin zone modify --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a --master
radosgw-admin zonegroup modify --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --master
radosgw-admin period update --commit --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a

Deleting Zone¶

The Rook toolbox can modify the Ceph Multisite state via the radosgw-admin command.

There are two scenarios possible when deleting a zone. The following commands, run via the toolbox, deletes the zone if there is only one zone in the zone group.

1 2	`radosgw-admin zone delete --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a radosgw-admin period update --commit --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a`

In the other scenario, there are more than one zones in a zone group.

Care must be taken when changing which zone is the master zone.

Please read the following documentation before running the below commands:

The following commands, run via toolboxes, remove the zone from the zone group first, then delete the zone.

radosgw-admin zonegroup rm --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a
radosgw-admin period update --commit --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a
radosgw-admin zone delete --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a
radosgw-admin period update --commit --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a --rgw-zone=zone-a

When a zone is deleted, the pools for that zone are not deleted.

Deleting Pools for a Zone¶

The Rook toolbox can delete pools. Deleting pools should be done with caution.

The following documentation on pools should be read before deleting any pools.

When a zone is created the following pools are created for each zone:

$ZONE_NAME.rgw.control
$ZONE_NAME.rgw.meta
$ZONE_NAME.rgw.log
$ZONE_NAME.rgw.buckets.index
$ZONE_NAME.rgw.buckets.non-ec
$ZONE_NAME.rgw.buckets.data

Here is an example command to delete the .rgw.buckets.data pool for zone-a.

1	`ceph osd pool rm zone-a.rgw.buckets.data zone-a.rgw.buckets.data --yes-i-really-really-mean-it`

In this command the pool name must be mentioned twice for the pool to be removed.

Removing an Object Store from a Zone¶

When an object-store (created in a zone) is deleted, the endpoint for that object store is removed from that zone, via

1	`kubectl delete -f object-store.yaml`

Removing object store(s) from the master zone of the master zone group should be done with caution. When all of these object-stores are deleted the period cannot be updated and that realm cannot be pulled.

Zone Group Deletion¶

Changes made to the resource's configuration or deletion of the resource are not reflected on the Ceph cluster.

When the ceph-object-zone group resource is deleted or modified, the zone group is not deleted from the Ceph cluster. Zone Group deletion must be done through the toolbox.

Deleting a Zone Group¶

The Rook toolbox can modify the Ceph Multisite state via the radosgw-admin command.

The following command, run via the toolbox, deletes the zone group.

1 2	`radosgw-admin zonegroup delete --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a radosgw-admin period update --commit --rgw-realm=realm-a --rgw-zonegroup=zonegroup-a`

Realm Deletion¶

Changes made to the resource's configuration or deletion of the resource are not reflected on the Ceph cluster.

When the ceph-object-realm resource is deleted or modified, the realm is not deleted from the Ceph cluster. Realm deletion must be done via the toolbox.

Deleting a Realm¶

The Rook toolbox can modify the Ceph Multisite state via the radosgw-admin command.

The following command, run via the toolbox, deletes the realm.

1	`radosgw-admin realm rm --rgw-realm=realm-a`

Configure an Existing Object Store for Multisite¶

When an object store is configured by Rook, it internally creates a zone, zone group, and realm with the same name as the object store. To enable multisite, you will need to create the corresponding zone, zone group, and realm CRs with the same name as the object store. For example, to create multisite CRs for an object store named my-store:

apiVersion: ceph.rook.io/v1
kind: CephObjectRealm
metadata:
  name: my-store
  namespace: rook-ceph # namespace:cluster
---
apiVersion: ceph.rook.io/v1
kind: CephObjectZoneGroup
metadata:
  name: my-store
  namespace: rook-ceph # namespace:cluster
spec:
  realm: my-store
---
apiVersion: ceph.rook.io/v1
kind: CephObjectZone
metadata:
  name: my-store
  namespace: rook-ceph # namespace:cluster
spec:
  zoneGroup: my-store
  metadataPool:
    replicated:
      size: 3
  dataPool:
    replicated:
      size: 3
  preservePoolsOnDelete: false
  # recommended to set this value if ingress used for exposing rgw endpoints
  # customEndpoints:
  #   - "http://rgw-a.fqdn"

Now modify the existing CephObjectStore CR to exclude pool settings and add a reference to the zone.

apiVersion: ceph.rook.io/v1
kind: CephObjectStore
metadata:
  name: my-store
  namespace: rook-ceph # namespace:cluster
spec:
  gateway:
    port: 80
    instances: 1
  zone:
    name: my-store

Using custom names¶

If names different from the object store need to be set for the realm, zone, or zone group, first rename them in the backend via toolbox pod, then following the procedure above.

radosgw-admin realm rename --rgw-realm=my-store --realm-new-name=<new-realm-name>
radosgw-admin zonegroup rename --rgw-zonegroup=my-store --zonegroup-new-name=<new-zonegroup-name> --rgw-realm=<new-realm-name>
radosgw-admin zone rename --rgw-zone=my-store --zone-new-name=<new-zone-name>  --rgw-zonegroup=<new-zonegroup-name> --rgw-realm=<new-realm-name>
radosgw-admin period update --commit

Important

Renaming in the toolbox must be performed before creating the multisite CRs