ceph-mgr orchestrator modules


This is developer documentation, describing Ceph internals that are only relevant to people writing ceph-mgr orchestrator modules.

In this context, orchestrator refers to some external service that provides the ability to discover devices and create Ceph services. This includes external projects such as ceph-ansible, DeepSea, and Rook.

An orchestrator module is a ceph-mgr module (ceph-mgr module developer’s guide) which implements common management operations using a particular orchestrator.

Orchestrator modules subclass the Orchestrator class: this class is an interface, it only provides method definitions to be implemented by subclasses. The purpose of defining this common interface for different orchestrators is to enable common UI code, such as the dashboard, to work with various different backends.

digraph G { subgraph cluster_1 { volumes [label="mgr/volumes"] rook [label="mgr/rook"] dashboard [label="mgr/dashboard"] orchestrator_cli [label="mgr/orchestrator_cli"] orchestrator [label="Orchestrator Interface"] ansible [label="mgr/ansible"] ssh [label="mgr/ssh"] deepsea [label="mgr/deepsea"] label = "ceph-mgr"; } volumes -> orchestrator dashboard -> orchestrator orchestrator_cli -> orchestrator orchestrator -> rook -> rook_io orchestrator -> ansible -> ceph_ansible orchestrator -> deepsea -> suse_deepsea orchestrator -> ssh rook_io [label="Rook"] ceph_ansible [label="ceph-ansible"] suse_deepsea [label="DeepSea"] rankdir="TB"; }

Behind all the abstraction, the purpose of orchestrator modules is simple: enable Ceph to do things like discover available hardware, create and destroy OSDs, and run MDS and RGW services.

A tutorial is not included here: for full and concrete examples, see the existing implemented orchestrator modules in the Ceph source tree.


Stateful service

a daemon that uses local storage, such as OSD or mon.

Stateless service

a daemon that doesn’t use any local storage, such as an MDS, RGW, nfs-ganesha, iSCSI gateway.


arbitrary string tags that may be applied by administrators to nodes. Typically administrators use labels to indicate which nodes should run which kinds of service. Labels are advisory (from human input) and do not guarantee that nodes have particular physical capabilities.

Drive group

collection of block devices with common/shared OSD formatting (typically one or more SSDs acting as journals/dbs for a group of HDDs).


choice of which node is used to run a service.

Key Concepts

The underlying orchestrator remains the source of truth for information about whether a service is running, what is running where, which nodes are available, etc. Orchestrator modules should avoid taking any internal copies of this information, and read it directly from the orchestrator backend as much as possible.

Bootstrapping nodes and adding them to the underlying orchestration system is outside the scope of Ceph’s orchestrator interface. Ceph can only work on nodes when the orchestrator is already aware of them.

Calls to orchestrator modules are all asynchronous, and return completion objects (see below) rather than returning values immediately.

Where possible, placement of stateless services should be left up to the orchestrator.

Completions and batching

All methods that read or modify the state of the system can potentially be long running. To handle that, all such methods return a Completion object. Orchestrator modules must implement the process method: this takes a list of completions, and is responsible for checking if they’re finished, and advancing the underlying operations as needed.

Each orchestrator module implements its own underlying mechanisms for completions. This might involve running the underlying operations in threads, or batching the operations up before later executing in one go in the background. If implementing such a batching pattern, the module would do no work on any operation until it appeared in a list of completions passed into process.

Some operations need to show a progress. Those operations need to add a ProgressReference to the completion. At some point, the progress reference becomes effective, meaning that the operation has really happened (e.g. a service has actually been started).


Given a list of Completion instances, process any which are incomplete.

Callers should inspect the detail of each completion to identify partial completion/progress information, and present that information to the user.

This method should not block, as this would make it slow to query a status, while other long running operations are in progress.

class orchestrator.Completion(_first_promise=None, value=<object object>, on_complete=None, name=None)

Combines multiple promises into one overall operation.

Completions are composable by being able to call one completion from another completion. I.e. making them re-usable using Promises E.g.:

>>> return Orchestrator().get_hosts().then(self._create_osd)

where get_hosts returns a Completion of list of hosts and _create_osd takes a list of hosts.

The concept behind this is to store the computation steps explicit and then explicitly evaluate the chain:

>>> p = Completion(on_complete=lambda x: x*2).then(on_complete=lambda x: str(x))
... p.finalize(2)
... assert p.result = "4"

or graphically:

+---------------+      +-----------------+
|               | then |                 |
| lambda x: x*x | +--> | lambda x: str(x)|
|               |      |                 |
+---------------+      +-----------------+

Sets the whole completion to be faild with this exception and end the evaluation.

property has_result

Has the operation already a result?

For Write operations, it can already have a result, if the orchestrator’s configuration is persistently written. Typically this would indicate that an update had been written to a manifest, but that the update had not necessarily been pushed out to the cluster.


property is_errored

Has the completion failed. Default implementation looks for self.exception. Can be overwritten.

property is_finished

Could the external operation be deemed as complete, or should we wait? We must wait for a read operation only if it is not complete.

property needs_result

Could the external operation be deemed as complete, or should we wait? We must wait for a read operation only if it is not complete.

property progress_reference

ProgressReference. Marks this completion as a write completeion.

property result

The result of the operation that we were waited for. Only valid after calling Orchestrator.process() on this completion.


Force a string.

class orchestrator.ProgressReference(message, mgr, completion=None)
completion = None

The completion can already have a result, before the write operation is effective. progress == 1 means, the services are created / removed.

property progress

if a orchestrator module can provide a more detailed progress information, it needs to also call progress.update().


In general, stateless services do not require any specific placement rules, as they can run anywhere that sufficient system resources are available. However, some orchestrators may not include the functionality to choose a location in this way, so we can optionally specify a location when creating a stateless service.

OSD services generally require a specific placement choice, as this will determine which storage devices are used.

Error Handling

The main goal of error handling within orchestrator modules is to provide debug information to assist users when dealing with deployment errors.

class orchestrator.OrchestratorError

General orchestrator specific error.

Used for deployment, configuration or user errors.

It’s not intended for programming errors or orchestrator internal errors.

class orchestrator.NoOrchestrator(msg='No orchestrator configured (try `ceph orchestrator set backend`)')

No orchestrator in configured.

class orchestrator.OrchestratorValidationError

Raised when an orchestrator doesn’t support a specific feature.

In detail, orchestrators need to explicitly deal with different kinds of errors:

  1. No orchestrator configured

    See NoOrchestrator.

  2. An orchestrator doesn’t implement a specific method.

    For example, an Orchestrator doesn’t support add_host.

    In this case, a NotImplementedError is raised.

  3. Missing features within implemented methods.

    E.g. optional parameters to a command that are not supported by the backend (e.g. the hosts field in Orchestrator.update_mons() command with the rook backend).

    See OrchestratorValidationError.

  4. Input validation errors

    The orchestrator_cli module and other calling modules are supposed to provide meaningful error messages.

    See OrchestratorValidationError.

  5. Errors when actually executing commands

    The resulting Completion should contain an error string that assists in understanding the problem. In addition, Completion.is_errored() is set to True

  6. Invalid configuration in the orchestrator modules

    This can be tackled similar to 5.

All other errors are unexpected orchestrator issues and thus should raise an exception that are then logged into the mgr log file. If there is a completion object at that point, Completion.result() may contain an error message.

Excluded functionality

  • Ceph’s orchestrator interface is not a general purpose framework for managing linux servers – it is deliberately constrained to manage the Ceph cluster’s services only.

  • Multipathed storage is not handled (multipathing is unnecessary for Ceph clusters). Each drive is assumed to be visible only on a single node.

Host management


Add a host to the orchestrator inventory.


host – hostname


Remove a host from the orchestrator inventory.


host – hostname


Report the hosts in the cluster.

The default implementation is extra slow.


list of InventoryNodes

Inventory and status

Orchestrator.get_inventory(node_filter=None, refresh=False)

Returns something that was created by ceph-volume inventory.


list of InventoryNode

class orchestrator.InventoryFilter(labels=None, nodes=None)

When fetching inventory, use this filter to avoid unnecessarily scanning the whole estate.

Typical use: filter by node when presenting UI workflow for configuring

a particular server. filter by label when not all of estate is Ceph servers, and we want to only learn about the Ceph servers. filter by label when we are interested particularly in e.g. OSD servers.

class ceph.deployment.inventory.Devices(devices)

A container for Device instances with reporting

class ceph.deployment.inventory.Device(path, sys_api=None, available=None, rejected_reasons=None, lvs=None, device_id=None)
Orchestrator.describe_service(service_type=None, service_id=None, node_name=None, refresh=False)

Describe a service (of any kind) that is already configured in the orchestrator. For example, when viewing an OSD in the dashboard we might like to also display information about the orchestrator’s view of the service (like the kubernetes pod ID).

When viewing a CephFS filesystem in the dashboard, we would use this to display the pods being currently run for MDS daemons.


list of ServiceDescription objects.

class orchestrator.ServiceDescription(nodename=None, container_id=None, container_image_id=None, container_image_name=None, service=None, service_instance=None, service_type=None, version=None, rados_config_location=None, service_url=None, status=None, status_desc=None)

For responding to queries about the status of a particular service, stateful or stateless.

This is not about health or performance monitoring of services: it’s about letting the orchestrator tell Ceph whether and where a service is scheduled in the cluster. When an orchestrator tells Ceph “it’s running on node123”, that’s not a promise that the process is literally up this second, it’s a description of where the orchestrator has decided the service should run.

Service Actions

Orchestrator.service_action(action, service_type, service_name=None, service_id=None)

Perform an action (start/stop/reload) on a service.

Either service_name or service_id must be specified:

  • If using service_name, perform the action on that entire logical service (i.e. all daemons providing that named service).

  • If using service_id, perform the action on a single specific daemon instance.

  • action – one of “start”, “stop”, “restart”, “redeploy”, “reconfig”

  • service_type – e.g. “mds”, “rgw”, …

  • service_name – name of logical service (“cephfs”, “us-east”, …)

  • service_id – service daemon instance (usually a short hostname)

Return type


OSD management


Create one or more OSDs within a single Drive Group.

The principal argument here is the drive_group member of OsdSpec: other fields are advisory/extensible for any finer-grained OSD feature enablement (choice of backing store, compression/encryption, etc).

  • drive_group – DriveGroupSpec

  • all_hosts – TODO, this is required because the orchestrator methods are not composable Probably this parameter can be easily removed because each orchestrator can use the “get_inventory” method and the “drive_group.host_pattern” attribute to obtain the list of hosts where to apply the operation

  • osd_ids – list of OSD IDs

  • destroy – marks the OSD as being destroyed. See OSD Replacement

Note that this can only remove OSDs that were successfully created (i.e. got an OSD ID).

class ceph.deployment.drive_group.DeviceSelection(paths=None, model=None, size=None, rotational=None, limit=None, vendor=None, all=False)

Used within ceph.deployment.drive_group.DriveGroupSpec to specify the devices used by the Drive Group.

Any attributes (even none) can be included in the device specification structure.

all = None

Matches all devices. Can only be used for data devices

limit = None

Limit the number of devices added to this Drive Group. Devices are used from top to bottom in the output of ceph-volume inventory

model = None

A wildcard string. e.g: “SDD*” or “SanDisk SD8SN8U5”

paths = None

List of absolute paths to the devices.

rotational = None

is the drive rotating or not

size = None

Size specification of format LOW:HIGH. Can also take the the form :HIGH, LOW: or an exact value (as ceph-volume inventory reports)

vendor = None

Match on the VENDOR property of the drive

class ceph.deployment.drive_group.DriveGroupSpec(host_pattern, data_devices=None, db_devices=None, wal_devices=None, journal_devices=None, data_directories=None, osds_per_device=None, objectstore='bluestore', encrypted=False, db_slots=None, wal_slots=None, osd_id_claims=None, block_db_size=None, block_wal_size=None, journal_size=None)

Describe a drive group in the same form that ceph-volume understands.

block_db_size = None

Set (or override) the “bluestore_block_db_size” value, in bytes

block_wal_size = None

Set (or override) the “bluestore_block_wal_size” value, in bytes

data_devices = None

A ceph.deployment.drive_group.DeviceSelection

data_directories = None

A list of strings, containing paths which should back OSDs

db_devices = None

A ceph.deployment.drive_group.DeviceSelection

db_slots = None

How many OSDs per DB device

encrypted = None

true or false

host_pattern = None

An fnmatch pattern to select hosts. Can also be a single host.

journal_devices = None

A ceph.deployment.drive_group.DeviceSelection

journal_size = None

set journal_size is bytes

objectstore = None

filestore or bluestore

osd_id_claims = None

Optional: mapping of OSD id to DeviceSelection, used when the created OSDs are meant to replace previous OSDs on the same node. See OSD Replacement

osds_per_device = None

Number of osd daemons per “DATA” device. To fully utilize nvme devices multiple osds are required.

wal_devices = None

A ceph.deployment.drive_group.DeviceSelection

wal_slots = None

How many OSDs per WAL device

Instructs the orchestrator to enable or disable either the ident or the fault LED.

class orchestrator.DeviceLightLoc

Describes a specific device on a specific host. Used for enabling or disabling LEDs on devices.

hostname as in orchestrator.Orchestrator.get_hosts()

device_id: e.g. ABC1234DEF567-1R1234_ABC8DE0Q.

See ceph osd metadata | jq '.[].device_ids'

OSD Replacement

See Replacing an OSD for the underlying process.

Replacing OSDs is fundamentally a two-staged process, as users need to physically replace drives. The orchestrator therefor exposes this two-staged process.

Phase one is a call to Orchestrator.remove_osds() with destroy=True in order to mark the OSD as destroyed.

Phase two is a call to Orchestrator.create_osds() with a Drive Group with

DriveGroupSpec.osd_id_claims set to the destroyed OSD ids.

Stateless Services

class orchestrator.StatelessServiceSpec(name, placement=None)

Details of stateless service creation.

Request to orchestrator for a group of stateless services such as MDS, RGW or iscsi gateway


Create a new MDS cluster


Remove an MDS cluster


Update / redeploy existing MDS cluster Like for example changing the number of service instances.


Create a new MDS zone


Remove a RGW zone


Update / redeploy existing RGW zone Like for example changing the number of service instances.

class orchestrator.NFSServiceSpec(name, pool=None, namespace=None, placement=None)

Create a new MDS cluster


Remove a NFS cluster


Update / redeploy existing NFS cluster Like for example changing the number of service instances.



Report on what versions are available to upgrade to


List of strings

Orchestrator.upgrade_start(image, version)

If an upgrade is currently underway, report on where we are in the process, or if some error has occurred.


UpgradeStatusSpec instance

class orchestrator.UpgradeStatusSpec



Report whether we can talk to the orchestrator. This is the place to give the user a meaningful message if the orchestrator isn’t running or can’t be contacted.

This method may be called frequently (e.g. every page load to conditionally display a warning banner), so make sure it’s not too expensive. It’s okay to give a slightly stale status (e.g. based on a periodic background ping of the orchestrator) if that’s necessary to make this method fast.


True doesn’t mean that the desired functionality is actually available in the orchestrator. I.e. this won’t work as expected:

>>> if OrchestratorClientMixin().available()[0]:  # wrong.
...     OrchestratorClientMixin().get_hosts()

two-tuple of boolean, string


Describes which methods this orchestrator implements


True doesn’t mean that the desired functionality is actually possible in the orchestrator. I.e. this won’t work as expected:

>>> api = OrchestratorClientMixin()
... if api.get_feature_set()['get_hosts']['available']:  # wrong.
...     api.get_hosts()

It’s better to ask for forgiveness instead:

>>> try:
...     OrchestratorClientMixin().get_hosts()
... except (OrchestratorError, NotImplementedError):
...     ...

Dict of API method names to {'available': True or False}

Client Modules

class orchestrator.OrchestratorClientMixin

A module that inherents from OrchestratorClientMixin can directly call all Orchestrator methods without manually calling remote.

Every interface method from Orchestrator is converted into a stub method that internally calls OrchestratorClientMixin._oremote()

>>> class MyModule(OrchestratorClientMixin):
...    def func(self):
...        completion = self.add_host('somehost')  # calls `_oremote()`
...        self._orchestrator_wait([completion])
...        self.log.debug(completion.result)

Useable in the Dashbord that uses a global mgr