my-rook-config/design/ceph/object/store.md

Rook Object Store

Overview

An object store is a collection of resources and services that work together to serve HTTP requests to PUT and GET objects. Rook will automate the configuration of the Ceph resources and services that are necessary to start and maintain a highly available, durable, and performant object store.

The Ceph object store supports S3 and Swift APIs and a multitude of features such as replication of object stores between different zones. The Rook object store is designed to support all of these features, though will take some time to implement them. We welcome contributions! In the meantime, features that are not yet implemented can be configured by using the Rook toolbox to run the radosgw-admin and other tools for advanced configuration.

Prerequisites

A Rook storage cluster must be configured and running in Kubernetes. In this example, it is assumed the cluster is in the rook namespace.

Object Store Walkthrough

When the storage admin is ready to create an object storage, the admin will specify his desired configuration settings in a yaml file such as the following object-store.yaml. This example is a simple object store with metadata that is replicated across different hosts, and the data is erasure coded across multiple devices in the cluster.

apiVersion: ceph.rook.io/v1alpha1
kind: CephObjectStore
metadata:
  name: my-store
  namespace: rook-ceph
spec:
  metadataPool:
    failureDomain: host
    replicated:
      size: 3
  dataPool:
    failureDomain: device
    erasureCoded:
      dataChunks: 6
      codingChunks: 2
  gateway:
    port: 80
    securePort: 443
    instances: 3
    hosting:
      dnsNames:
      - "my-ingress.mydomain.com"
      - "rook-ceph-rgw-my-store.rook-ceph.svc"

Now create the object store.

kubectl create -f object-store.yaml

At this point the Rook operator recognizes that a new object store resource needs to be configured. The operator will create all of the resources to start the object store.

1. Metadata pools are created (.rgw.root, my-store.rgw.control, my-store.rgw.meta, my-store.rgw.log, my-store.rgw.buckets.index)
2. The data pool is created (my-store.rgw.buckets.data)
3. A Kubernetes service is created to provide load balancing for the RGW pod(s)
4. A Kubernetes deployment is created to start the RGW pod(s) with the settings for the new zone
5. The zone is modified to add the RGW pod endpoint(s) if zone is mentioned in the configuration

When the RGW pods start, the object store is ready to receive the http or https requests as configured.

Object Store CRD

The object store settings are exposed to Rook as a Custom Resource Definition (CRD). The CRD is the Kubernetes-native means by which the Rook operator can watch for new resources. The operator stays in a control loop to watch for a new object store, changes to an existing object store, or requests to delete an object store.

Pools

Pools created by CephObjectStore CRD

The pools are the backing data store for the object store and are created with specific names to be private to an object store. Pools can be configured with all of the settings that can be specified in the Pool CRD. The underlying schema for pools defined by a pool CRD is the same as the schema under the metadataPool and dataPool elements of the object store CRD. All metadata pools are created with the same settings, while the data pool can be created with independent settings. The metadata pools must use replication, while the data pool can use replication or erasure coding.

If preservePoolsOnDelete is set to 'true' the pools used to support the object store will remain when the object store will be deleted. This is a security measure to avoid accidental loss of data. It is set to 'false' by default. If not specified is also deemed as 'false'.

spec:
  metadataPool:
    failureDomain: host
    replicated:
      size: 3
  dataPool:
    failureDomain: device
    erasureCoded:
      dataChunks: 6
      codingChunks: 2
  preservePoolsOnDelete: true

Pools shared by multiple CephObjectStore

If user want to use existing pools for metadata and data, the pools must be created before the object store is created. This will be useful if multiple objectstore can share same pools. The detail of pools need to shared in sharedPools settings in object-store CRD. Now the object stores can consume same pool isolated with different namespaces. Usually RGW server itself create different namespaces on the pools. User can create via Pool CRD, this is need to present before the object store is created. Similar to preservePoolsOnDelete setting, preserveRadosNamespaceDataOnDelete is used to preserve the data in the rados namespace when the object store is deleted. It is set to 'false' by default.

spec:
  sharedPools:
    metadataPoolName: rgw-meta-pool
    dataPoolName: rgw-data-pool
    preserveRadosNamespaceDataOnDelete: true

To create the pools that will be shared by multiple object stores, create the following CephBlockPool CRs:

apiVersion: ceph.rook.io/v1
kind: CephBlockPool
metadata:
    name: rgw-meta-pool
  spec:
    failureDomain: host
    replicated:
      size: 3
    parameters:
      pg_num: 8
    application: rgw
---
apiVersion: ceph.rook.io/v1
kind: CephBlockPool
 metadata:
    name: rgw-data-pool
  spec:
    failureDomain: osd
      erasureCoded:
        dataChunks: 6
        codingChunks: 2
    application: rgw
---
apiVersion: ceph.rook.io/v1
kind: CephBlockPool
 metadata:
    name: .rgw.root
  spec:
    name: .rgw.root
    failureDomain: host
    replicated:
      size: 3
    parameters:
      pg_num: 8
    application: rgw

The pools for this configuration will be created as below:

# ceph osd pool ls
.rgw.root
rgw-meta-pool
rgw-data-pool

And the pool configuration in zone is as below:

# radosgw-admin zone get --rgw-zone=my-store
{
    "id": "2220eb5f-2751-4a51-9c7d-da4ce1b0e4e1",
    "name": "my-store",
    "domain_root": "rgw-meta-pool:my-store.meta.root",
    "control_pool": "rgw-meta-pool:my-store.control",
    "gc_pool": "rgw-meta-pool:my-store.log.gc",
    "lc_pool": "rgw-meta-pool:my-store.log.lc",
    "log_pool": "rgw-meta-pool:my-store.log",
    "intent_log_pool": "rgw-meta-pool:my-store.log.intent",
    "usage_log_pool": "rgw-meta-pool:my-store.log.usage",
    "roles_pool": "rgw-meta-pool:my-store.meta.roles",
    "reshard_pool": "rgw-meta-pool:my-store.log.reshard",
    "user_keys_pool": "rgw-meta-pool:my-store.meta.users.keys",
    "user_email_pool": "rgw-meta-pool:my-store.meta.users.email",
    "user_swift_pool": "rgw-meta-pool:my-store.meta.users.swift",
    "user_uid_pool": "rgw-meta-pool:my-store.meta.users.uid",
    "otp_pool": "rgw-meta-pool:my-store.otp",
    "system_key": {
        "access_key": "",
        "secret_key": ""
    },
    "placement_pools": [
        {
            "key": "default-placement",
            "val": {
                "index_pool": "rgw-metadata-pool:my-store.buckets.index",
                "storage_classes": {
                    "STANDARD": {
                        "data_pool": "rgw-data-pool:my-store.buckets.data"
                    }
                },
                "data_extra_pool": "rgw-data-pool:my-store.buckets.non-ec", #only pool is not erasure coded, otherwise use different pool
                "index_type": 0,
                "inline_data": "true"
            }
        }
    ],
    "realm_id": "65a7bf34-42d3-4344-ac40-035e160d7f9e",
    "notif_pool": "rgw-meta-pool:my-store.notif"
}

The following steps need to implement internally in Rook Operator to add this feature assuming zone and pool are already created:

# radosgw-admin zone get --rgw-zone=my-store > zone.json

# modify the zone.json to add the pools as above

# radosgw-admin zone set --rgw-zone=my-store --infile=zone.json

# radosgw-admin period update --commit

After deleting the object store the data in the rados namespace can deleted by Rook Operator using following commands:

# rados -p rgw-meta-pool ls --all | grep my-store

# rados -p rgw-data-pool ls --all | grep my-store

# rados -p <rgw-meta-pool | rgw-data-pool> rm <each object> -N <namespace> (from above list)

For mulisite use case

If there is a zone section in object-store configuration, then the pool creation will configured by the CephObjectZone CR. The CephObjectStore CR will include below section to specify the zone name.

spec:
  zone:
    name: zone1

Gateway

The gateway settings correspond to the RGW service.

• type: Can be s3. In the future support for swift can be added.
• sslCertificateRef: If specified, this is the name of the Kubernetes secret that contains the SSL certificate to be used for secure connections to the object store. The secret must be in the same namespace as the Rook cluster. If it is an opaque Kubernetes Secret, Rook will look in the secret provided at the cert key name. The value of the cert key must be in the format expected by the RGW service: "The server key, server certificate, and any other CA or intermediate certificates be supplied in one file. Each of these items must be in pem form." If the certificate is not specified, SSL will not be configured.
• caBundleRef: If specified, this is the name of the Kubernetes secret (type opaque) that contains ca-bundle to use. The secret must be in the same namespace as the Rook cluster. Rook will look in the secret provided at the cabundle key name.
• port: The service port where the RGW service will be listening (http)
• securePort: The service port where the RGW service will be listening (https)
• instances: The number of RGW pods that will be started for this object store
• placement: The rgw pods can be given standard Kubernetes placement restrictions with nodeAffinity, tolerations, podAffinity, podAntiAffinity, and topologySpreadConstraints similar to placement defined for daemons configured by the cluster CRD.

The RGW service can be configured to listen on both http and https by specifying both port and securePort.

 gateway:
    sslCertificateRef: my-ssl-cert-secret
    securePort: 443
    instances: 1

Multisite

By default, the object store will be created independently from any other object stores and replication to another object store will not be configured. This done by creating a new Ceph realm, zone group, and zone all with the name of the new object store.

If desired to configure the object store to replicate and sync data amongst object-store or Ceph clusters, the zone section would be required.

This section enables the object store to be part of a specified ceph-object-zone.

Specifying this section also ensures that the pool section in the ceph-object-zone is used for the object-store. If pools are specified for the object-store they are neither created nor deleted.

• name: name of the ceph-object-zone the object store is in. This name must be of a ceph-object-zone resource not just of a zone that has been already created.

  zone:
    name: "name"

LDAP

The Ceph Object Gateway supports integrating with LDAP for authenticating and creating users, please refer here. This means that the rgw backend user is also required to be part of groups in the LDAP server otherwise, authentication will fail. The rgw backend user can be generated from CephObjectStoreUser or ObjectBucketClaim CRDs. For the both resources credentials are saved in Kubernetes Secrets which may not be valid with LDAP Server, user need to follow the steps mentioned here.The following settings need to be configured in the RGW server:

rgw ldap binddn =
rgw ldap secret = /etc/ceph/ldap/bindpass.secret
rgw ldap uri =
rgw ldap searchdn =
rgw ldap dnattr =
rgw ldap searchfilter =
rgw s3 auth use ldap = true

So the CRD for the Ceph Object Store will be modified to include the above changes:

spec:
  security
    ldap:
      config:
        uri: ldaps://ldap-server:636
        binddn: "uid=ceph,cn=users,cn=accounts,dc=example,dc=com"
        searchdn: "cn=users,cn=accounts,dc=example,dc=com"
        dnattr: "uid"
        searchfilter: "memberof=cn=s3,cn=groups,cn=accounts,dc=example,dc=com"
      credential:
        volumeSource:
          secret:
            secretName: object-my-store-ldap-creds
            defaultMode: 0600 #required

The config section includes options used for RGW wrt LDAP server. These options are strongly typed rather than string map approach since very less chance to modify in future.

• uri: It specifies the address of LDAP server to use.* binddn: The bind domain for the service account used by RGW server.
• searchdn: The search domain where can it look for the user details.
• dnattr: The attribute being used in the constructed search filter to match a username, this can either be uid or cn.
• searchfilter: A generic search filter. If dnattr is set, this filter is &()'d together with the automatically constructed filter.

The credential defines where the password for accessing ldap server should be sourced from

• volumeSource: this is a standard Kubernetes VolumeSource for the Kerberos keytab file like what is normally used to configure Volumes for a Pod. For example, a Secret or HostPath. There are two requirements for the source's content:
  1. The config file must be mountable via subPath: password. For example, in a Secret, the data item must be named password, or items must be defined to select the key and give it path password. A HostPath directory must have the password file.
  2. The volume or config file must have mode 0600.

The CA bundle for ldap can be added to the caBundleRef option in Gateway settings:

spec:
  gateway:
    caBundleRef: #ldaps-cabundle

Virtual host-style access for buckets

The Ceph Object Gateway supports accessing buckets using virtual host-style addressing, which allows accessing buckets using the bucket name as a subdomain in the endpoint. This is important because AWS (the primary definer of the S3 interface) has deprecated the alternative, path-style access which is Ceph (and Rook's) default deployment. AWS has extended support for path-style access, but Rook/OBC/COSI users have begun to identify applications which do not support the long-deprecated path-style access.

Virtual host-style (vhost-style) addressing requires 2 things:

1. An endpoint that supports wildcard addressing
2. The above DNS endpoint must be added to RGW's rgw_dns_names list

The user can configure this option manually in Rook like below for a wildcard-enabled endpoint:

# ceph config set client.rgw.mystore.a rgw_dns_name my-ingress.mydomain.com,rook-ceph-rgw-my-store.rook-ceph.svc

In this example, if the object store contains a bucket named sample, then the S3 vhost-style access URL would be sample.my-ingress.mydomain.com. More details about the feature can be found in Ceph Documentation. This is supported from Ceph Reef release(v18.0) onwards.

Wildcard DNS addressing can be configured in myriad ways. Some options:

• Kubernetes ingress loadbalancer
• Openshift DNS operator

Rook will implement the following API to allow for user configuration, with detailed design notes to follow.

• hosting (optional) - configures hosting settings for the CephObjectStore
  • advertiseEndpoint (optional) - allow users to definitively specify which endpoint the user wants applications to use by default for internal S3 connections.
  • dnsNames (optional) - RGW will reject any S3 connections to unknown endpoints. Users should add any additional endpoints RGW should accept here.

The below example illustrates a commonly-anticipated configuration, and one that Rook can recommend to users. The user is using an internal, wildcard-enabled K8s service as the advertised endpoint for OBCs. The user also has externally-available ingress that the RGW needs to accept connections from, which serves S3 applications outside the Kubernetes cluster. Both endpoints are wildcard-enabled, but the internal service is preferable to the external ingress to prevent the ingress router from being a cluster-wide S3 bottleneck for Kubernetes applications.

spec:
  hosting:
    advertiseEndpoint:
      dnsName: my-internal-wildcard-service.mydomain.com
      port: 8443
      useTLS: true
    dnsNames:
      - my-external-ingress.mydomain.com

These proposed configurations are necessary for users to enable vhost-style addressing, but they are not only useful when DNS wildcarding is enabled. These configurations are independent and may provide some value to other user scenarios. Be clear in documentation that wildcarding is enabled by these features but not required.

When the user configures DNS wildcarding on an endpoint other than the CephObjectStore service endpoint, Rook should advertise that endpoint to CephObjectStores, OBCs, and COSI Buckets/Accesses as a priority. However, Rook cannot know which endpoints support wildcarding in order to prioritize advertising them. Therefore, allow users to disambiguate for Rook which endpoint should be advertised via advertiseEndpoint.

S3 clients (and therefore OBCs and COSI) generally assume a single endpoint for an object store. Multiple advertised endpoints will not be supported to avoid user and internal/developer confusion.

By default, Rook will advertise the CephObjectStore service endpoint with a priority on advertising the HTTPS (securePort) endpoint. Because the advertised endpoint is primarily relevant for resources internal to the Kubernetes cluster, this default should be sufficient for most users, and this is the behavior expected by users when dnsNames is not configured, so it should be familiar.

When this feature is enabled, there is also ambiguity about which endpoint Rook should use for Admin Ops API communication. Some users have reported issues with Rook using a dnsNames endpoint (or advertiseEndpoint) when they wish to set up ingress certificates after Rook deployment. The obvious alternative is to have Rook always use the CephObjectStore service, but other users have expressed troubles creating certificates or CAs that allow the service endpoint in the past.

Each rgw_dns_name entry must be a valid RFC-1123 hostname, and Rook Operator will perform input validation using k8s apimachinery IsDNS1123Subdomain().

When rgw_dns_name is changed for an RGW cluster, all RGWs need to be restarted. To enforce this in Rook, we can apply the --rgw-dns-name flag, which will restart RGWs with no user action needed.

This is different from CephObjectZone customEndpoints which is used for configuring the object store to replicate and sync data amongst Multisite.

When endpoints are configured in rgw_dns_names, the RGW will reject any incoming connections intended for endpoints not in the list. Therefore, all endpoints that might be used must be added.

For convenience, and to ensure other CephObjectStore configurations are not rendered unusable when users add additional endpoints to dnsNames, the following should be added to the rgw_dns_names list automatically by Rook:

• the advertiseEndpoint.dnsName
• the default service endpoint for the object store (e.g., rook-ceph-rgw-my-store.rook-ceph.svc)
• CephObjectZone customEndpoints

When Rook builds the rgw_dns_names list internally, Rook should remove any duplicate entries. While Rook add endpoints to the list for safety and convenience, users might add the same endpoints, which Rook should not treat as a configuration bug. Rook should also ensure the list ordering is consistent between reconciles.

In order to attempt to strike the best balance for everyone, and to provide the best clarity for users and Rook internally, Rook will always use the service endpoint for admin ops. Rook documentation must inform users that if TLS is enabled, they must give Rook a certificate that accepts the service endpoint. Alternately, if that is not possible, Rook will add an insecureSkipTlsVerification option to the CephObjectStore to allow users to provision a healthy CephObjectStore. This opens users up to machine-in-the-middle attacks, so users should be advised to only use it for test/proof-of-concept clusters, or to work around bugs temporarily.

Rook can refer users to this Kubernetes doc for a suggested way that they can manage certificates in a Kubernetes cluster that work with Kubernetes services like the CephObjectStore service: https://kubernetes.io/docs/tasks/tls/managing-tls-in-a-cluster/

For external CephCephObjectStores (i.e., when spec.gateway.externalRgwEndpoints are set), vhost-style addressing should be configured on the host cluster, and hosting.dnsNames is irrelevant. The default advertiseEndpoint for external CephObjectStores is the first entry in the spec.gateway.externalRgwEndpoints list, which users should be able to override if desired.