Rook allows creation and customization of storage clusters through the custom resource definitions (CRDs). There are primarily three different modes in which to create your cluster.
- Specify host paths and raw devices
- Dynamically provision storage underneath Rook by specifying the storage class Rook should use to consume storage via PVCs
- Create a Stretch cluster that distributes Ceph mons across three zones, while storage (OSDs) is only configured in two zones
Following is an example for each of these approaches. More examples are included later in this doc.
To get you started, here is a simple example of a CRD to configure a Ceph cluster with all nodes and all devices. The Ceph persistent data is stored directly on a host path (Ceph Mons) and on raw devices (Ceph OSDs).
In addition to your CephCluster object, you need to create the namespace, service accounts, and RBAC rules for the namespace you are going to create the CephCluster in. These resources are defined in the example
In a "PVC-based cluster", the Ceph persistent data is stored on volumes requested from a storage class of your choice. This type of cluster is recommended in a cloud environment where volumes can be dynamically created and also in clusters where a local PV provisioner is available.
For a more advanced scenario, such as adding a dedicated device you can refer to the dedicated metadata device for OSD on PVC section.
For environments that only have two failure domains available where data can be replicated, consider the case where one failure domain is down and the data is still fully available in the remaining failure domain. To support this scenario, Ceph has recently integrated support for "stretch" clusters.
Rook requires three zones. Two zones (A and B) will each run all types of Rook pods, which we call the "data" zones. Two mons run in each of the two data zones, while two replicas of the data are in each zone for a total of four data replicas. The third zone (arbiter) runs a single mon. No other Rook or Ceph daemons need to be run in the arbiter zone.
For this example, we assume the desired failure domain is a zone. Another failure domain can also be specified with a known topology node label which is already being used for OSD failure domains.
For more details, see the Stretch Cluster design doc.
Settings can be specified at the global level to apply to the cluster as a whole, while other settings can be specified at more fine-grained levels. If any setting is unspecified, a suitable default will be used automatically.
name: The name that will be used internally for the Ceph cluster. Most commonly the name is the same as the namespace since multiple clusters are not supported in the same namespace.
namespace: The Kubernetes namespace that will be created for the Rook cluster. The services, pods, and other resources created by the operator will be added to this namespace. The common scenario is to create a single Rook cluster. If multiple clusters are created, they must not have conflicting devices or host paths.
true, the cluster will not be managed by Rook but via an external entity. This mode is intended to connect to an existing cluster. In this case, Rook will only consume the external cluster. However, Rook will be able to deploy various daemons in Kubernetes such as object gateways, mds and nfs if an image is provided and will refuse otherwise. If this setting is enabled all the other options will be ignored except
dataDirHostPath. See external cluster configuration. If
cephVersion.imageis left blank, Rook will refuse the creation of extra CRs like object, file and nfs.
cephVersion: The version information for launching the ceph daemons.
image: The image used for running the ceph daemons. For example,
v17.2.0. For more details read the container images section. For the latest ceph images, see the Ceph DockerHub. To ensure a consistent version of the image is running across all nodes in the cluster, it is recommended to use a very specific image version. Tags also exist that would give the latest version, but they are only recommended for test environments. For example, the tag
v15will be updated each time a new Octopus build is released. Using the
v15or similar tag is not recommended in production because it may lead to inconsistent versions of the image running across different nodes in the cluster.
true, allow an unsupported major version of the Ceph release. Currently
pacificare supported. Future versions such as
quincywould require this to be set to
true. Should be set to
dataDirHostPath: The path on the host (hostPath) where config and data should be stored for each of the services. If the directory does not exist, it will be created. Because this directory persists on the host, it will remain after pods are deleted. Following paths and any of their subpaths must not be used:
- On Minikube environments, use
/data/rook. Minikube boots into a tmpfs but it provides some directories where files can be persisted across reboots. Using one of these directories will ensure that Rook's data and configuration files are persisted and that enough storage space is available.
- WARNING: For test scenarios, if you delete a cluster and start a new cluster on the same hosts, the path used by
dataDirHostPathmust be deleted. Otherwise, stale keys and other config will remain from the previous cluster and the new mons will fail to start. If this value is empty, each pod will get an ephemeral directory to store their config files that is tied to the lifetime of the pod running on that node. More details can be found in the Kubernetes empty dir docs.
skipUpgradeChecks: if set to true Rook won't perform any upgrade checks on Ceph daemons during an upgrade. Use this at YOUR OWN RISK, only if you know what you're doing. To understand Rook's upgrade process of Ceph, read the upgrade doc.
continueUpgradeAfterChecksEvenIfNotHealthy: if set to true Rook will continue the OSD daemon upgrade process even if the PGs are not clean, or continue with the MDS upgrade even the file system is not healthy.
dashboard: Settings for the Ceph dashboard. To view the dashboard in your browser see the dashboard guide.
enabled: Whether to enable the dashboard to view cluster status
urlPrefix: Allows to serve the dashboard under a subpath (useful when you are accessing the dashboard via a reverse proxy)
port: Allows to change the default port where the dashboard is served
ssl: Whether to serve the dashboard via SSL, ignored on Ceph versions older than
monitoring: Settings for monitoring Ceph using Prometheus. To enable monitoring on your cluster see the monitoring guide.
enabled: Whether to enable prometheus based monitoring for this cluster
externalMgrEndpoints: external cluster manager endpoints
externalMgrPrometheusPort: external prometheus manager module port. See external cluster configuration for more details.
rulesNamespace: Namespace to deploy prometheusRule. If empty, namespace of the cluster will be used. Recommended:
- If you have a single Rook Ceph cluster, set the
rulesNamespaceto the same namespace as the cluster or keep it empty.
- If you have multiple Rook Ceph clusters in the same Kubernetes cluster, choose the same namespace to set
rulesNamespacefor all the clusters (ideally, namespace with prometheus deployed). Otherwise, you will get duplicate alerts with duplicate alert definitions.
- If you have a single Rook Ceph cluster, set the
network: For the network settings for the cluster, refer to the network configuration settings
mon: contains mon related options mon settings For more details on the mons and when to choose a number other than
3, see the mon health doc.
mgr: manager top level section
count: set number of ceph managers between
2. The default value is 2. If there are two managers, it is important for all mgr services point to the active mgr and not the passive mgr. Therefore, Rook will automatically update all services (in the cluster namespace) that have a label
app=rook-ceph-mgrwith a selector pointing to the active mgr. This commonly applies to services for the dashboard or the prometheus metrics collector.
modules: is the list of Ceph manager modules to enable
crashCollector: The settings for crash collector daemon(s).
disable: is set to
true, the crash collector will not run on any node where a Ceph daemon runs
daysToRetain: specifies the number of days to keep crash entries in the Ceph cluster. By default the entries are kept indefinitely.
logCollector: The settings for log collector daemon.
enabled: if set to
true, the log collector will run as a side-car next to each Ceph daemon. The Ceph configuration option
log_to_filewill be turned on, meaning Ceph daemons will log on files in addition to still logging to container's stdout. These logs will be rotated. (default: false)
periodicity: how often to rotate daemon's log. (default: 24h). Specified with a time suffix which may be 'h' for hours or 'd' for days. Rotating too often will slightly impact the daemon's performance since the signal briefly interrupts the program.
annotations: annotations configuration settings
labels: labels configuration settings
placement: placement configuration settings
resources: resources configuration settings
priorityClassNames: priority class names configuration settings
storage: Storage selection and configuration that will be used across the cluster. Note that these settings can be overridden for specific nodes.
false, indicating if all nodes in the cluster should be used for storage according to the cluster level storage selection and configuration values. If individual nodes are specified under the
useAllNodesmust be set to
nodes: Names of individual nodes in the cluster that should have their storage included in accordance with either the cluster level configuration specified above or any node specific overrides described in the next section below.
useAllNodesmust be set to
falseto use specific nodes and their config. See node settings below.
config: Config settings applied to all OSDs on the node unless overridden by
devices. See the config settings below.
- storage selection settings
- Storage Class Device Sets
onlyApplyOSDPlacement: Whether the placement specific for OSDs is merged with the
false, the OSD placement will be merged with the
allplacement. If true, the
OSD placement will be appliedand the
allplacement will be ignored. The placement for OSDs is computed from several different places depending on the type of OSD:
- For non-PVCs:
- For PVCs:
placement.alland inside the storageClassDeviceSets from the
- For non-PVCs:
disruptionManagement: The section for configuring management of daemon disruptions
true, the operator will create and manage PodDisruptionBudgets for OSD, Mon, RGW, and MDS daemons. OSD PDBs are managed dynamically via the strategy outlined in the design. The operator will block eviction of OSDs by default and unblock them safely when drains are detected.
osdMaintenanceTimeout: is a duration in minutes that determines how long an entire failureDomain like
region/zone/hostwill be held in
noout(in addition to the default DOWN/OUT interval) when it is draining. This is only relevant when
true. The default value is
true, the operator will create and manage MachineDisruptionBudgets to ensure OSDs are only fenced when the cluster is healthy. Only available on OpenShift.
machineDisruptionBudgetNamespace: the namespace in which to watch the MachineDisruptionBudgets.
truethe operator will remove the OSDs that are down and whose data has been restored to other OSDs. In Ceph terms, the OSDs are
safe-to-destroywhen they are removed.
cleanupPolicy: cleanup policy settings
security: security page for key management configuration
Ceph container images¶
Official releases of Ceph Container images are available from Docker Hub.
These are general purpose Ceph container with all necessary daemons and dependencies installed.
|vRELNUM||Latest release in this series (e.g., v15 = Octopus)|
|vRELNUM.Y||Latest stable release in this stable series (e.g., v15.2)|
|vRELNUM.Y.Z||A specific release (e.g., v15.2.5)|
|vRELNUM.Y.Z-YYYYMMDD||A specific build (e.g., v15.2.11-20200419)|
A specific will contain a specific release of Ceph as well as security fixes from the Operating System.
count: Set the number of mons to be started. The number must be between
9. The recommended value is most commonly
3. For highest availability, an odd number of mons should be specified. For higher durability in case of mon loss, an even number can be specified although availability may be lower. To maintain quorum a majority of mons must be up. For example, if there are three mons, two must be up. If there are four mons, three must be up. If there are two mons, both must be up. If quorum is lost, see the disaster recovery guide to restore quorum from a single mon.
allowMultiplePerNode: Whether to allow the placement of multiple mons on a single node. Default is
falsefor production. Should only be set to
truein test environments.
PersistentVolumeSpecused by Rook to create PVCs for monitor storage. This field is optional, and when not provided, HostPath volume mounts are used. The current set of fields from template that are used are
storageresource request and limit. The default storage size request for new PVCs is
10Gi. Ensure that associated storage class is configured to use
volumeBindingMode: WaitForFirstConsumer. This setting only applies to new monitors that are created when the requested number of monitors increases, or when a monitor fails and is recreated. An example CRD configuration is provided below.
stretchCluster: The stretch cluster settings that define the zones (or other failure domain labels) across which to configure the cluster.
failureDomainLabel: The label that is expected on each node where the cluster is expected to be deployed. The labels must be found in the list of well-known topology labels.
subFailureDomain: With a zone, the data replicas must be spread across OSDs in the subFailureDomain. The default is
zones: The failure domain names where the Mons and OSDs are expected to be deployed. There must be three zones specified in the list. This element is always named
zoneeven if a non-default
failureDomainLabelis specified. The elements have two values:
name: The name of the zone, which is the value of the domain label.
arbiter: Whether the zone is expected to be the arbiter zone which only runs a single mon. Exactly one zone must be labeled
true. The two zones that are not the arbiter zone are expected to have OSDs deployed.
If these settings are changed in the CRD the operator will update the number of mons during a periodic check of the mon health, which by default is every 45 seconds.
To change the defaults that the operator uses to determine the mon health and whether to failover a mon, refer to the health settings. The intervals should be small enough that you have confidence the mons will maintain quorum, while also being long enough to ignore network blips where mons are failed over too often.
You can use the cluster CR to enable or disable any manager module. This can be configured like so:
Some modules will have special configuration to ensure the module is fully functional after being enabled. Specifically:
pg_autoscaler: Rook will configure all new pools with PG autoscaling by setting:
osd_pool_default_pg_autoscale_mode = on
Network Configuration Settings¶
If not specified, the default SDN will be used. Configure the network that will be enabled for the cluster and services.
provider: Specifies the network provider that will be used to connect the network interface. You can choose between
selectors: List the network selector(s) that will be used associated by a key.
ipFamily: Specifies the network stack Ceph daemons should listen on.
dualStack: Specifies that Ceph daemon should listen on both IPv4 and IPv6 network stacks.
connections: Settings for network connections using Ceph's msgr2 protocol
encryption: Settings for encryption on the wire to Ceph daemons
enabled: Whether to encrypt the data in transit across the wire to prevent eavesdropping the data on the network. The default is false. When encryption is enabled, all communication between clients and Ceph daemons, or between Ceph daemons will be encrypted. When encryption is not enabled, clients still establish a strong initial authentication and data integrity is still validated with a crc check. IMPORTANT: Encryption requires the 5.11 kernel for the latest nbd and cephfs drivers. Alternatively for testing only, set "mounter: rbd-nbd" in the rbd storage class, or "mounter: fuse" in the cephfs storage class. The nbd and fuse drivers are not recommended in production since restarting the csi driver pod will disconnect the volumes.
enabled: Whether to compress the data in transit across the wire. The default is false. Requires Ceph Quincy (v17) or newer. Also see the kernel requirements above for encryption.
Changing networking configuration after a Ceph cluster has been deployed is NOT supported and will result in a non-functioning cluster.
To use host networking, set
Rook supports addition of public and cluster network for ceph using Multus
The selector keys are required to be
cluster where each represent:
public: client communications with the cluster (reads/writes)
cluster: internal Ceph replication network
If you want to learn more, please read:
Based on the configuration, the operator will do the following:
If only the
publicselector is specified, all communication will happen on that network
If only the
clusterselector is specified, the internal cluster traffic* will happen on that network. All other traffic to mons, OSDs, and other daemons will be on the default network.
clusterselectors are specified the first one will run all the communication network and the second the internal cluster network*
* Internal cluster traffic includes OSD heartbeats, data replication, and data recovery
Only OSD pods will have both Public and Cluster networks attached. The rest of the Ceph component pods and CSI pods will only have the Public network attached. Rook Ceph Operator will not have any networks attached as it proxies the required commands via a sidecar container in the mgr pod.
In order to work, each selector value must match a
NetworkAttachmentDefinition object name in Multus.
multus network provider, an already working cluster with Multus networking is required. Network attachment definition that later will be attached to the cluster needs to be created before the Cluster CRD. The Network attachment definitions should be using whereabouts cni. If Rook cannot find the provided Network attachment definition it will fail running the Ceph OSD pods. You can add the Multus network attachment selection annotation selecting the created network attachment definition on
A valid NetworkAttachmentDefinition will look like following:
- Ensure that
mastermatches the network interface of the host that you want to use.
- Ipam type
whereaboutsis required because it makes sure that all the pods get a unique IP address from the multus network.
- The NetworkAttachmentDefinition should be referenced along with the namespace in which it is present like
public: <namespace>/<name of NAD>. e.g., the network attachment definition are in
- This format is required in order to use the NetworkAttachmentDefinition across namespaces.
- In Openshift, to use a NetworkAttachmentDefinition (NAD) across namespaces, the NAD must be deployed in the
defaultnamespace. The NAD is then referenced with the namespace:
Known limitations with Multus¶
Daemons leveraging Kubernetes service IPs (Monitors, Managers, Rados Gateways) are not listening on the NAD specified in the
selectors. Instead the daemon listens on the default network, however the NAD is attached to the container, allowing the daemon to communicate with the rest of the cluster. There is work in progress to fix this issue in the multus-service repository. At the time of writing it's unclear when this will be supported.
Provide single-stack IPv4 or IPv6 protocol to assign corresponding addresses to pods and services. This field is optional. Possible inputs are IPv6 and IPv4. Empty value will be treated as IPv4. Kubernetes version should be at least v1.13 to run IPv6. Dual-stack is supported as of ceph Pacific. To turn on dual stack see the network configuration section.
In addition to the cluster level settings specified above, each individual node can also specify configuration to override the cluster level settings and defaults. If a node does not specify any configuration then it will inherit the cluster level settings.
name: The name of the node, which should match its
config: Config settings applied to all OSDs on the node unless overridden by
devices. See the config settings below.
- storage selection settings
useAllNodes is set to
true, Rook attempts to make Ceph cluster management as hands-off as possible while still maintaining reasonable data safety. If a usable node comes online, Rook will begin to use it automatically. To maintain a balance between hands-off usability and data safety, Nodes are removed from Ceph as OSD hosts only (1) if the node is deleted from Kubernetes itself or (2) if the node has its taints or affinities modified in such a way that the node is no longer usable by Rook. Any changes to taints or affinities, intentional or unintentional, may affect the data reliability of the Ceph cluster. In order to help protect against this somewhat, deletion of nodes by taint or affinity modifications must be "confirmed" by deleting the Rook-Ceph operator pod and allowing the operator deployment to restart the pod.
For production clusters, we recommend that
useAllNodes is set to
false to prevent the Ceph cluster from suffering reduced data reliability unintentionally due to a user mistake. When
useAllNodes is set to
false, Rook relies on the user to be explicit about when nodes are added to or removed from the Ceph cluster. Nodes are only added to the Ceph cluster if the node is added to the Ceph cluster resource. Similarly, nodes are only removed if the node is removed from the Ceph cluster resource.
Nodes can be added and removed over time by updating the Cluster CRD, for example with
kubectl -n rook-ceph edit cephcluster rook-ceph. This will bring up your default text editor and allow you to add and remove storage nodes from the cluster. This feature is only available when
useAllNodes has been set to
Storage Selection Settings¶
Below are the settings for host-based cluster. This type of cluster can specify devices for OSDs, both at the cluster and individual node level, for selecting which storage resources will be included in the cluster.
false, indicating whether all devices found on nodes in the cluster should be automatically consumed by OSDs. Not recommended unless you have a very controlled environment where you will not risk formatting of devices with existing data. When
true, all devices/partitions will be used. Is overridden by
deviceFilter: A regular expression for short kernel names of devices (e.g.
sda) that allows selection of devices to be consumed by OSDs. If individual devices have been specified for a node then this filter will be ignored. This field uses golang regular expression syntax. For example:
sdb: Only selects the
sdbdevice if found
^sd.: Selects all devices starting with
^sd[a-d]: Selects devices starting with
^s: Selects all devices that start with
^[^r]: Selects all devices that do not start with
devicePathFilter: A regular expression for device paths (e.g.
/dev/disk/by-path/pci-0:1:2:3-scsi-1) that allows selection of devices to be consumed by OSDs. If individual devices or
deviceFilterhave been specified for a node then this filter will be ignored. This field uses golang regular expression syntax. For example:
^/dev/sd.: Selects all devices starting with
^/dev/disk/by-path/pci-.*: Selects all devices which are connected to PCI bus
devices: A list of individual device names belonging to this node to include in the storage cluster.
name: The name of the device (e.g.,
sda), or full udev path (e.g.
/dev/disk/by-id/ata-ST4000DM004-XXXX- this will not change after reboots).
config: Device-specific config settings. See the config settings below
Host-based cluster only supports raw device and partition. Be sure to see the quickstart doc prerequisites for additional considerations.
Below are the settings for a PVC-based cluster.
storageClassDeviceSets: Explained in Storage Class Device Sets
Storage Class Device Sets¶
The following are the settings for Storage Class Device Sets which can be configured to create OSDs that are backed by block mode PVs.
name: A name for the set.
count: The number of devices in the set.
resources: The CPU and RAM requests/limits for the devices. (Optional)
placement: The placement criteria for the devices. (Optional) Default is no placement criteria.
The syntax is the same as for other placement configuration. It supports
It is recommended to configure the placement such that the OSDs will be as evenly spread across nodes as possible. At a minimum, anti-affinity should be added so at least one OSD will be placed on each available nodes.
However, if there are more OSDs than nodes, this anti-affinity will not be effective. Another placement scheme to consider is to add labels to the nodes in such a way that the OSDs can be grouped on those nodes, create multiple storageClassDeviceSets, and add node affinity to each of the device sets that will place the OSDs in those sets of nodes.
Rook will automatically add required nodeAffinity to the OSD daemons to match the topology labels that are found on the nodes where the OSD prepare jobs ran. To ensure data durability, the OSDs are required to run in the same topology that the Ceph CRUSH map expects. For example, if the nodes are labeled with rack topology labels, the OSDs will be constrained to a certain rack. Without the topology labels, Rook will not constrain the OSDs beyond what is required by the PVs, for example to run in the zone where provisioned. See the OSD Topology section for the related labels.
preparePlacement: The placement criteria for the preparation of the OSD devices. Creating OSDs is a two-step process and the prepare job may require different placement than the OSD daemons. If the
preparePlacementis not specified, the
placementwill instead be applied for consistent placement for the OSD prepare jobs and OSD deployments. The
preparePlacementis only useful for
portableOSDs in the device sets. OSDs that are not portable will be tied to the host where the OSD prepare job initially runs.
- For example, provisioning may require topology spread constraints across zones, but the OSD daemons may require constraints across hosts within the zones.
true, the OSDs will be allowed to move between nodes during failover. This requires a storage class that supports portability (e.g.
aws-ebs, but not the local storage provisioner). If
false, the OSDs will be assigned to a node permanently. Rook will configure Ceph's CRUSH map to support the portability.
tuneDeviceClass: For example, Ceph cannot detect AWS volumes as HDDs from the storage class "gp2", so you can improve Ceph performance by setting this to true.
tuneFastDeviceClass: For example, Ceph cannot detect Azure disks as SSDs from the storage class "managed-premium", so you can improve Ceph performance by setting this to true..
volumeClaimTemplates: A list of PVC templates to use for provisioning the underlying storage devices.
resources.requests.storage: The desired capacity for the underlying storage devices.
storageClassName: The StorageClass to provision PVCs from. Default would be to use the cluster-default StorageClass. This StorageClass should provide a raw block device, multipath device, or logical volume. Other types are not supported. If you want to use logical volume, please see known issue of OSD on LV-backed PVC
volumeMode: The volume mode to be set for the PVC. Which should be Block
accessModes: The access mode for the PVC to be bound by OSD.
schedulerName: Scheduler name for OSD pod placement. (Optional)
encrypted: whether to encrypt all the OSDs in a given storageClassDeviceSet
OSD Configuration Settings¶
The following storage selection settings are specific to Ceph and do not apply to other backends. All variables are key-value pairs represented as strings.
metadataDevice: Name of a device to use for the metadata of OSDs on each node. Performance can be improved by using a low latency device (such as SSD or NVMe) as the metadata device, while other spinning platter (HDD) devices on a node are used to store data. Provisioning will fail if the user specifies a
metadataDevicebut that device is not used as a metadata device by Ceph. Notably,
ceph-volumewill not use a device of the same device class (HDD, SSD, NVMe) as OSD devices for metadata, resulting in this failure.
databaseSizeMB: The size in MB of a bluestore database. Include quotes around the size.
walSizeMB: The size in MB of a bluestore write ahead log (WAL). Include quotes around the size.
deviceClass: The CRUSH device class to use for this selection of storage devices. (By default, if a device's class has not already been set, OSDs will automatically set a device's class to either
nvmebased on the hardware properties exposed by the Linux kernel.) These storage classes can then be used to select the devices backing a storage pool by specifying them as the value of the pool spec's
initialWeight: The initial OSD weight in TiB units. By default, this value is derived from OSD's capacity.
primaryAffinity: The primary-affinity value of an OSD, within range
osdsPerDevice**: The number of OSDs to create on each device. High performance devices such as NVMe can handle running multiple OSDs. If desired, this can be overridden for each node and each device.
encryptedDevice**: Encrypt OSD volumes using dmcrypt ("true" or "false"). By default this option is disabled. See encryption for more information on encryption in Ceph.
crushRoot: The value of the
rootCRUSH map label. The default is
default. Generally, you should not need to change this. However, if any of your topology labels may have the value
default, you need to change
crushRootto avoid conflicts, since CRUSH map values need to be unique.
Annotations and Labels¶
Annotations and Labels can be specified so that the Rook components will have those annotations / labels added to them.
You can set annotations / labels for Rook components for the list of key value pairs:
all: Set annotations / labels for all components except
mgr: Set annotations / labels for MGRs
mon: Set annotations / labels for mons
osd: Set annotations / labels for OSDs
prepareosd: Set annotations / labels for OSD Prepare Jobs
monitoring: Set annotations / labels for service monitor
crashcollector: Set annotations / labels for crash collectors
clusterMetadata: Set annotations only to
rook-ceph-mon-endpointsconfigmap and the
rook-ceph-admin-keyringsecrets. These annotations will not be merged with the
allannotations. The common usage is for backing up these critical resources with
kubed. Note the clusterMetadata annotation will not be merged with the
allannotation. When other keys are set,
allwill be merged together with the specific component.
Placement Configuration Settings¶
Placement configuration for the cluster services. It includes the following keys:
all. Each service will have its placement configuration generated by merging the generic configuration under
all with the most specific one (which will override any attributes).
In stretch clusters, if the
arbiter placement is specified, that placement will only be applied to the arbiter. Neither will the
arbiter placement be merged with the
all placement to allow the arbiter to be fully independent of other daemon placement. The remaining mons will still use the
Placement of OSD pods is controlled using the Storage Class Device Set, not the general
A Placement configuration is specified (according to the kubernetes PodSpec) as:
nodeAffinity: kubernetes NodeAffinity
podAffinity: kubernetes PodAffinity
podAntiAffinity: kubernetes PodAntiAffinity
tolerations: list of kubernetes Toleration
topologySpreadConstraints: kubernetes TopologySpreadConstraints
If you use
osd pods, you must write two rules both for
rook-ceph-osd-prepare like the example configuration. It comes from the design that there are these two pods for an OSD. For more detail, see the osd design doc and the related issue.
The Rook Ceph operator creates a Job called
rook-ceph-detect-version to detect the full Ceph version used by the given
cephVersion.image. The placement from the
mon section is used for the Job except for the
Cluster-wide Resources Configuration Settings¶
Resources should be specified so that the Rook components are handled after Kubernetes Pod Quality of Service classes. This allows to keep Rook components running when for example a node runs out of memory and the Rook components are not killed depending on their Quality of Service class.
You can set resource requests/limits for Rook components through the Resource Requirements/Limits structure in the following keys:
mon: Set resource requests/limits for mons
osd: Set resource requests/limits for OSDs. This key applies for all OSDs regardless of their device classes. In case of need to apply resource requests/limits for OSDs with particular device class use specific osd keys below. If the memory resource is declared Rook will automatically set the OSD configuration
osd_memory_targetto the same value. This aims to ensure that the actual OSD memory consumption is consistent with the OSD pods' resource declaration.
osd-<deviceClass>: Set resource requests/limits for OSDs on a specific device class. Rook will automatically detect
nvmedevice classes. Custom device classes can also be set.
mgr: Set resource requests/limits for MGRs
mgr-sidecar: Set resource requests/limits for the MGR sidecar, which is only created when
mgr.count: 2. The sidecar requires very few resources since it only executes every 15 seconds to query Ceph for the active mgr and update the mgr services if the active mgr changed.
prepareosd: Set resource requests/limits for OSD prepare job
crashcollector: Set resource requests/limits for crash. This pod runs wherever there is a Ceph pod running. It scrapes for Ceph daemon core dumps and sends them to the Ceph manager crash module so that core dumps are centralized and can be easily listed/accessed. You can read more about the Ceph Crash module.
logcollector: Set resource requests/limits for the log collector. When enabled, this container runs as side-car to each Ceph daemons.
cleanup: Set resource requests/limits for cleanup job, responsible for wiping cluster's data after uninstall
In order to provide the best possible experience running Ceph in containers, Rook internally recommends minimum memory limits if resource limits are passed. If a user configures a limit or request value that is too low, Rook will still run the pod(s) and print a warning to the operator log.
mgr-sidecar: 100MB limit, 40MB requests
The resources for MDS daemons are not configured in the Cluster. Refer to the Ceph Filesystem CRD instead.
For more information on resource requests/limits see the official Kubernetes documentation: Kubernetes - Managing Compute Resources for Containers
requests: Requests for cpu or memory.
cpu: Request for CPU (example: one CPU core
1, 50% of one CPU core
memory: Limit for Memory (example: one gigabyte of memory
1Gi, half a gigabyte of memory
limits: Limits for cpu or memory.
cpu: Limit for CPU (example: one CPU core
1, 50% of one CPU core
memory: Limit for Memory (example: one gigabyte of memory
1Gi, half a gigabyte of memory
Priority Class Names¶
Priority class names can be specified so that the Rook components will have those priority class names added to them.
You can set priority class names for Rook components for the list of key value pairs:
all: Set priority class names for MGRs, Mons, OSDs, and crashcollectors.
mgr: Set priority class names for MGRs. Examples default to system-cluster-critical.
mon: Set priority class names for Mons. Examples default to system-node-critical.
osd: Set priority class names for OSDs. Examples default to system-node-critical.
crashcollector: Set priority class names for crashcollectors.
The specific component keys will act as overrides to
Rook-Ceph will monitor the state of the CephCluster on various components by default. The following CRD settings are available:
healthCheck: main ceph cluster health monitoring section
Currently three health checks are implemented:
mon: health check on the ceph monitors, basically check whether monitors are members of the quorum. If after a certain timeout a given monitor has not joined the quorum back it will be failed over and replace by a new monitor.
osd: health check on the ceph osds
status: ceph health status check, periodically check the Ceph health state and reflects it in the CephCluster CR status field.
The liveness probe and startup probe of each daemon can also be controlled via
startupProbe respectively. The settings are valid for
osd. Here is a complete example for both
The probe's timing values and thresholds (but not the probe itself) can also be overridden. For more info, refer to the Kubernetes documentation.
For example, you could change the
mgr probe by applying:
Changing the liveness probe is an advanced operation and should rarely be necessary. If you want to change these settings then modify the desired settings.
The operator is regularly configuring and checking the health of the cluster. The results of the configuration and health checks can be seen in the
status section of the CephCluster CR.
Ceph is constantly monitoring the health of the data plane and reporting back if there are any warnings or errors. If everything is healthy from Ceph's perspective, you will see
If Ceph reports any warnings or errors, the details will be printed to the status. If further troubleshooting is needed to resolve these issues, the toolbox will likely be needed where you can run
ceph commands to find more details.
capacity of the cluster is reported, including bytes available, total, and used. The available space will be less that you may expect due to overhead in the OSDs.
conditions represent the status of the Rook operator.
- If the cluster is fully configured and the operator is stable, the
Readycondition is raised with
ClusterCreatedreason and no other conditions. The cluster will remain in the
Readycondition after the first successful configuration since it is expected the storage is consumable from this point on. If there are issues preventing the storage layer from working, they are expected to show as Ceph health errors.
- If the cluster is externally connected successfully, the
Readycondition will have the reason
- If the operator is currently being configured or the operator is checking for update, there will be a
- If there was a failure, the condition(s) status will be
messagewill give a summary of the error. See the operator log for more details.
There are several other properties for the overall status including:
state: A summary of the overall current state of the cluster, which is somewhat duplicated from the conditions for backward compatibility.
storage.deviceClasses: The names of the types of storage devices that Ceph discovered in the cluster. These types will be
hddunless they have been overridden with the
version: The version of the Ceph image currently deployed.
Here are several samples for configuring Ceph clusters. Each of the samples must also include the namespace and corresponding access granted for management by the Ceph operator. See the common cluster resources below.
Storage configuration: All devices¶
Storage Configuration: Specific devices¶
Individual nodes and their config can be specified so that only the named nodes below will be used as storage resources. Each node's 'name' field should match their 'kubernetes.io/hostname' label.
To control where various services will be scheduled by kubernetes, use the placement configuration sections below. The example under 'all' would have all services scheduled on kubernetes nodes labeled with 'role=storage-node' and tolerate taints with a key of 'storage-node'.
To control how many resources the Rook components can request/use, you can set requests and limits in Kubernetes for them. You can override these requests/limits for OSDs per node when using
useAllNodes: false in the
node item in the
Before setting resource requests/limits, please take a look at the Ceph documentation for recommendations for each component: Ceph - Hardware Recommendations.
The topology of the cluster is important in production environments where you want your data spread across failure domains. The topology can be controlled by adding labels to the nodes. When the labels are found on a node at first OSD deployment, Rook will add them to the desired level in the CRUSH map.
The complete list of labels in hierarchy order from highest to lowest is:
For example, if the following labels were added to a node:
These labels would result in the following hierarchy for OSDs on that node (this command can be run in the Rook toolbox):
Ceph requires unique names at every level in the hierarchy (CRUSH map). For example, you cannot have two racks with the same name that are in different zones. Racks in different zones must be named uniquely.
Note that the
host is added automatically to the hierarchy by Rook. The host cannot be specified with a topology label. All topology labels are optional.
When setting the node labels prior to
CephCluster creation, these settings take immediate effect. However, applying this to an already deployed
CephCluster requires removing each node from the cluster first and then re-adding it with new configuration to take effect. Do this node by node to keep your data safe! Check the result with
ceph osd tree from the Rook Toolbox. The OSD tree should display the hierarchy for the nodes that already have been re-added.
To utilize the
failureDomain based on the node labels, specify the corresponding option in the CephBlockPool
This configuration will split the replication of volumes across unique racks in the data center setup.
Using PVC storage for monitors¶
In the CRD specification below three monitors are created each using a 10Gi PVC created by Rook using the
local-storage storage class.
In the CRD specification below, 3 OSDs (having specific placement and resource values) and 3 mons with each using a 10Gi PVC, are created by Rook using the
local-storage storage class.
Dedicated metadata and wal device for OSD on PVC¶
In the simplest case, Ceph OSD BlueStore consumes a single (primary) storage device. BlueStore is the engine used by the OSD to store data.
The storage device is normally used as a whole, occupying the full device that is managed directly by BlueStore. It is also possible to deploy BlueStore across additional devices such as a DB device. This device can be used for storing BlueStore’s internal metadata. BlueStore (or rather, the embedded RocksDB) will put as much metadata as it can on the DB device to improve performance. If the DB device fills up, metadata will spill back onto the primary device (where it would have been otherwise). Again, it is only helpful to provision a DB device if it is faster than the primary device.
You can have multiple
volumeClaimTemplates where each might either represent a device or a metadata device. So just taking the
storage section this will give something like:
Note that Rook only supports three naming convention for a given template:
- "data": represents the main OSD block device, where your data is being stored.
- "metadata": represents the metadata (including block.db and block.wal) device used to store the Ceph Bluestore database for an OSD.
- "wal": represents the block.wal device used to store the Ceph Bluestore database for an OSD. If this device is set, "metadata" device will refer specifically to block.db device. It is recommended to use a faster storage class for the metadata or wal device, with a slower device for the data. Otherwise, having a separate metadata device will not improve the performance.
The bluestore partition has the following reference combinations supported by the ceph-volume utility:
- A single "data" device.
- A "data" device and a "metadata" device.
- A "data" device and a "wal" device. A WAL device can be used for BlueStore’s internal journal or write-ahead log (block.wal), it is only useful to use a WAL device if the device is faster than the primary device (data device). There is no separate "metadata" device in this case, the data of main OSD block and block.db located in "data" device.
- A "data" device, a "metadata" device and a "wal" device.
To determine the size of the metadata block follow the official Ceph sizing guide.
With the present configuration, each OSD will have its main block allocated a 10GB device as well a 5GB device to act as a bluestore database.
The minimum supported Ceph version for the External Cluster is Luminous 12.2.x.
The features available from the external cluster will vary depending on the version of Ceph. The following table shows the minimum version of Ceph for some of the features:
|Dynamic provisioning RBD||12.2.X|
|Configure extra CRDs (object, file, nfs)1||13.2.3|
|Dynamic provisioning CephFS||14.2.3|
External Cluster configuration¶
Source cluster: The cluster providing the data, usually configured by cephadm
Consumer cluster: The K8s cluster that will be consuming the external source cluster
Commands on the source Ceph cluster¶
In order to configure an external Ceph cluster with Rook, we need to extract some information in order to connect to that cluster.
- Run the python script create-external-cluster-resources.py for creating all users and keys.
--namespace: Namespace where CephCluster will run, for example
--format bash: The format of the output
--rbd-data-pool-name: The name of the RBD data pool
--rbd-metadata-ec-pool-name: (optional) Provides the name of erasure coded RBD metadata pool
--rados-namespace: (optional) Divides a pool into separate logical namespaces
--cephfs-filesystem-name: (optional) The name of the filesystem
--cephfs-metadata-pool-name: (optional) Provides the name of the cephfs metadata pool
--cephfs-data-pool-name: (optional) Provides the name of the CephFS data pool
--rgw-endpoint: (optional) The RADOS Gateway endpoint in the format
--rgw-pool-prefix: (optional) The prefix of the RGW pools. If not specified, the default prefix is
--rgw-tls-cert-path: (optional) RADOS Gateway endpoint TLS certificate file path
--rgw-skip-tls: (optional) Ignore TLS certification validation when a self-signed certificate is provided (NOT RECOMMENDED)
--monitoring-endpoint: (optional) Ceph Manager prometheus exporter endpoints (comma separated list of
entries of active and standby mgrs)
--monitoring-endpoint-port: (optional) Ceph Manager prometheus exporter port
--ceph-conf: (optional) Provide a Ceph conf file
--cluster-name: (optional) Ceph cluster name
--output: (optional) Output will be stored into the provided file
--dry-run: (optional) Prints the executed commands without running them
--run-as-user: (optional) Provides a user name to check the cluster's health status, must be prefixed by
--restricted-auth-permission: (optional) Restrict cephCSIKeyrings auth permissions to specific pools, and cluster. Mandatory flags that need to be set are
--cephfs-filesystem-nameflag can also be passed in case of CephFS user restriction, so it can restrict users to particular CephFS filesystem.
!!! note Restricting the csi-users per pool, and per cluster will require creating new csi-users and new secrets for that csi-users, so it can only be used for new deployments.
1 2 3 4
--upgrade: (optional) Upgrades the 'Ceph CSI keyrings (For example: client.csi-cephfs-provisioner) with new permissions needed for the new cluster version and older permission will still be applied.
This will upgrade all the default csi-users(non-restricted)
Restricted users created using --restricted-auth-permission flag (For example: client.csi-rbd-node-rookStorage-replicapool) need to pass mandatory flags: '--rbd-data-pool-name(if it is a rbd user), --cluster-name and --run-as-user' flags while upgrading, in case of cephfs users if you have passed --cephfs-filesystem-name flag while creating csi-users then while upgrading it will be mandatory too
!!! note An existing non-restricted user cannot be downgraded to a restricted user by upgrading. Admin needs to create a new restricted user for this by re-running the script.The upgrade flag should only be used to append new permissions to users. It shouldn't be used for changing a csi user already applied permissions. For example, you shouldn't change the pool(s) a user has access to.
Copy the bash output.
Commands on the K8s consumer cluster¶
Paste the above output from
create-external-cluster-resources.pyinto your current shell to allow importing the source data.
Run the import script.
Verify the consumer cluster is connected to the source ceph cluster:
StorageClass will also be created, verify its creation.
cephfsStorageClass would be respective name for RBD and CephFS StorageClass.
For CephFS StorageClass you also need to export
CEPHFS_POOL_NAMEor you can pass these parameters with a CLI flag (--cephfs-data-pool-name,--cephfs-filesystem-name) while running the python script. For creating ECRBDStorageClass you need to pass --rbd-metadata-ec-pool-name CLI flag or export
Then you can now create a persistent volume based on these StorageClass.
CephCluster example (management)¶
The following CephCluster CR represents a cluster that will perform management tasks on the external cluster. It will not only act as a consumer but will also allow the deployment of other CRDs such as CephFilesystem or CephObjectStore. You would need to inject the admin keyring for that.
The corresponding YAML example:
Deleting a CephCluster¶
During deletion of a CephCluster resource, Rook protects against accidental or premature destruction of user data by blocking deletion if there are any other Rook-Ceph Custom Resources that reference the CephCluster being deleted. Rook will warn about which other resources are blocking deletion in three ways until all blocking resources are deleted:
- An event will be registered on the CephCluster resource
- A status condition will be added to the CephCluster resource
- An error will be added to the Rook-Ceph Operator log
Rook has the ability to cleanup resources and data that were deployed when a CephCluster is removed. The policy settings indicate which data should be forcibly deleted and in what way the data should be wiped. The
cleanupPolicy has several fields:
confirmation: Only an empty string and
yes-really-destroy-dataare valid values for this field. If this setting is empty, the cleanupPolicy settings will be ignored and Rook will not cleanup any resources during cluster removal. To reinstall the cluster, the admin would then be required to follow the cleanup guide to delete the data on hosts. If this setting is
yes-really-destroy-data, the operator will automatically delete the data on hosts. Because this cleanup policy is destructive, after the confirmation is set to
yes-really-destroy-dataRook will stop configuring the cluster as if the cluster is about to be destroyed.
sanitizeDisks: sanitizeDisks represents advanced settings that can be used to delete data on drives.
method: indicates if the entire disk should be sanitized or simply ceph's metadata. Possible choices are 'quick' (default) or 'complete'
dataSource: indicate where to get random bytes from to write on the disk. Possible choices are 'zero' (default) or 'random'. Using random sources will consume entropy from the system and will take much more time then the zero source
iteration: overwrite N times instead of the default (1). Takes an integer value
allowUninstallWithVolumes: If set to true, then the cephCluster deletion doesn't wait for the PVCs to be deleted. Default is false.
To automate activation of the cleanup, you can use the following command. WARNING: DATA WILL BE PERMANENTLY DELETED:
Nothing will happen until the deletion of the CR is requested, so this can still be reverted. However, all new configuration by the operator will be blocked with this cleanup policy enabled.
Rook waits for the deletion of PVs provisioned using the cephCluster before proceeding to delete the cephCluster. To force deletion of the cephCluster without waiting for the PVs to be deleted, you can set the allowUninstallWithVolumes to true under spec.CleanupPolicy.
Configure an object store, shared filesystem, or NFS resources in the local cluster to connect to the external Ceph cluster ↩