Back to Ceph

OSD Service

doc/cephadm/services/osd.rst

21.0.037.5 KB
Original Source

OSD Service


.. _libstoragemgmt: https://github.com/libstorage/libstoragemgmt

List Devices

ceph-volume scans each host in the cluster periodically in order to determine the devices that are present and responsive. It is also determined whether each is eligible to be used for new OSDs in a block, DB, or WAL role.

To print a list of devices discovered by cephadm, run this command:

.. prompt:: bash #

ceph orch device ls [--hostname=...] [--wide] [--refresh]

Example:

.. code-block:: console

Hostname Path Type Serial Size Health Ident Fault Available srv-01 /dev/sdb hdd 15P0A0YFFRD6 300G Unknown N/A N/A No srv-01 /dev/sdc hdd 15R0A08WFRD6 300G Unknown N/A N/A No srv-01 /dev/sdd hdd 15R0A07DFRD6 300G Unknown N/A N/A No srv-01 /dev/sde hdd 15P0A0QDFRD6 300G Unknown N/A N/A No srv-02 /dev/sdb hdd 15R0A033FRD6 300G Unknown N/A N/A No srv-02 /dev/sdc hdd 15R0A05XFRD6 300G Unknown N/A N/A No srv-02 /dev/sde hdd 15R0A0ANFRD6 300G Unknown N/A N/A No srv-02 /dev/sdf hdd 15R0A06EFRD6 300G Unknown N/A N/A No srv-03 /dev/sdb hdd 15R0A0OGFRD6 300G Unknown N/A N/A No srv-03 /dev/sdc hdd 15R0A0P7FRD6 300G Unknown N/A N/A No srv-03 /dev/sdd hdd 15R0A0O7FRD6 300G Unknown N/A N/A No

In the above examples you can see fields named Health, Ident, and Fault. This information is provided by integration with libstoragemgmt. By default, this integration is disabled because libstoragemgmt may not be 100% compatible with your hardware. To direct Ceph to include these fields, enable cephadm's "enhanced device scan" option as follows:

.. prompt:: bash #

ceph config set mgr mgr/cephadm/device_enhanced_scan true

Note that the columns reported by ceph orch device ls may vary from release to release.

The --wide option shows device details, including any reasons that the device might not be eligible for use as an OSD. Example (Reef):

.. code-block:: console

HOST PATH TYPE DEVICE ID SIZE AVAILABLE REFRESHED REJECT REASONS davidsthubbins /dev/sdc hdd SEAGATE_ST20000NM002D_ZVTBJNGC17010W339UW25 18.1T No 22m ago Has a FileSystem, Insufficient space (<10 extents) on vgs, LVM detected nigeltufnel /dev/sdd hdd SEAGATE_ST20000NM002D_ZVTBJNGC17010C3442787 18.1T No 22m ago Has a FileSystem, Insufficient space (<10 extents) on vgs, LVM detected

.. warning:: Although the libstoragemgmt library issues standard SCSI (SES) inquiry calls, there is no guarantee that your hardware and firmware properly implement these standards. This can lead to erratic behavior and even bus resets on some older hardware. It is therefore recommended that, before enabling this feature, you first test your hardware's compatibility with libstoragemgmt to avoid unplanned interruptions to services.

There are a number of ways to test compatibility, but the simplest is
to use the cephadm shell to call ``libstoragemgmt`` directly: ``cephadm shell
lsmcli ldl``. If your hardware is supported you should see something like
this: 

.. code-block:: console

  Path     | SCSI VPD 0x83    | Link Type | Serial Number      | Health Status
  ----------------------------------------------------------------------------
  /dev/sda | 50000396082ba631 | SAS       | 15P0A0R0FRD6       | Good
  /dev/sdb | 50000396082bbbf9 | SAS       | 15P0A0YFFRD6       | Good

After enabling libstoragemgmt support, the output will look something like this:

.. prompt:: bash #

ceph orch device ls

.. code-block:: console

Hostname Path Type Serial Size Health Ident Fault Available srv-01 /dev/sdb hdd 15P0A0YFFRD6 300G Good Off Off No srv-01 /dev/sdc hdd 15R0A08WFRD6 300G Good Off Off No :

In this example, libstoragemgmt has confirmed the health of the drives and the ability to interact with the identification and fault LEDs on the drive enclosures. For further information about interacting with these LEDs, refer to :ref:devices.

.. note:: The current release of libstoragemgmt (1.8.8) supports SCSI, SAS, and SATA based local drives only. There is no official support for NVMe devices (PCIe), SAN LUNs, or exotic/complex metadevices.

Retrieve Exact Size of Block Devices

Run a command of the following form to discover the exact size of a block device. The value returned here is used by the orchestrator when filtering based on size:

.. prompt:: bash #

cephadm shell ceph-volume inventory </dev/sda> --format json | jq .sys_api.human_readable_size

The exact size in GB is the size reported in TB, multiplied by 1024.

Example

The following provides a specific example of this command based upon the general form of the command above:

.. prompt:: bash #

cephadm shell ceph-volume inventory /dev/sdc --format json | jq .sys_api.human_readable_size

.. code-block:: console

"3.64 TB"

This indicates that the exact device size is 3.64 TB, or 3727.36 GB.

This procedure was developed by Frédéric Nass. See this thread on the [ceph-users] mailing list <https://lists.ceph.io/hyperkitty/list/[email protected]/message/5BAAYFCQAZZDRSNCUPCVBNEPGJDARRZA/>_ for discussion of this matter.

.. _cephadm-deploy-osds:

Deploy OSDs

Listing Storage Devices

In order to deploy an OSD, there must be an available storage device or devices on which the OSD will be deployed.

Run this command to display an inventory of storage devices on all cluster hosts:

.. prompt:: bash #

ceph orch device ls

A storage device is considered available if all of the following conditions are met:

  • The device must have no partitions.
  • The device must not have any LVM state.
  • The device must not be mounted.
  • The device must not contain a file system.
  • The device must not contain a Ceph BlueStore OSD.
  • The device must be >= 5 GB.

Ceph will not provision an OSD on a device that is not available.

Creating New OSDs

There are multiple ways to create new OSDs:

  • Consume any available and unused storage device:

    .. prompt:: bash #

    ceph orch apply osd --all-available-devices

  • Create an OSD from a specific device on a specific host:

    .. prompt:: bash #

    ceph orch daemon add osd <host>:<device-path>

    For example:

    .. prompt:: bash #

    ceph orch daemon add osd host1:/dev/sdb

  • Advanced OSD creation from specific devices on a specific host:

    .. prompt:: bash #

    ceph orch daemon add osd host1:data_devices=/dev/sda,/dev/sdb,db_devices=/dev/sdc,osds_per_device=2

  • Create an OSD on a specific LVM logical volume on a specific host:

    .. prompt:: bash #

    ceph orch daemon add osd <host>:<lvm-path>

    For example:

    .. prompt:: bash #

    ceph orch daemon add osd host1:/dev/vg_osd/lvm_osd1701

  • You can use :ref:drivegroups to categorize devices based on their properties. This is useful to clarify which devices are available to consume. Properties include device type (SSD or HDD), device model names, size, and the hosts on which the devices exist:

    .. prompt:: bash #

    ceph orch apply -i spec.yml

.. warning:: When deploying new OSDs with cephadm, ensure that the ceph-osd package is not installed on the target host. If it is installed, conflicts may arise in the management and control of the OSD that may lead to errors or unexpected behavior.

  • New OSDs created using ceph orch daemon add osd are added under osd.default as managed OSDs with a valid spec.

    To attach an existing OSD to a different managed service, ceph orch osd set-spec-affinity command can be used:

    .. prompt:: bash #

    ceph orch osd set-spec-affinity <service_name> <osd_id(s)>

    For example:

    .. prompt:: bash #

    ceph orch osd set-spec-affinity osd.default_drive_group 0 1

Dry Run

The --dry-run flag causes the orchestrator to present a preview of what will happen without actually creating the OSDs.

For example:

.. prompt:: bash #

ceph orch apply osd --all-available-devices --dry-run

.. code-block:: console

NAME HOST DATA DB WAL all-available-devices node1 /dev/vdb - - all-available-devices node2 /dev/vdc - - all-available-devices node3 /dev/vdd - -

.. _cephadm-osd-declarative:

Declarative State

The effect of ceph orch apply is persistent. This means that drives that are added to the system after the ceph orch apply command completes will be automatically detected and added to the cluster as specified. It also means that drives that become available (e.g. by zapping) after the ceph orch apply command completes will be automatically found and added to the cluster.

We will examine the effects of the following command:

.. prompt:: bash #

ceph orch apply osd --all-available-devices

After running the above command:

  • When you add new drives to the cluster, they will automatically be used to create new OSDs.
  • When you remove an OSD and clean the LVM physical volume, a new OSD will be created automatically.

If you want to avoid this behavior (disable automatic creation of OSD on available devices), use the unmanaged parameter:

.. prompt:: bash #

ceph orch apply osd --all-available-devices --unmanaged=true

.. note::

Keep these three facts in mind:

- The default behavior of ``ceph orch apply`` causes ``cephadm`` to constantly reconcile. This means that ``cephadm`` creates OSDs as soon as new drives are detected.

- Setting ``unmanaged: True`` disables the creation of OSDs. If ``unmanaged: True`` is set, nothing will happen even if you apply a new OSD service.

- ``ceph orch daemon add`` creates OSDs, but does not add an OSD service.

For more on cephadm, see also :ref:cephadm-spec-unmanaged.

.. _cephadm-osd-removal:

Remove an OSD

Removing an OSD from a cluster involves two steps:

#. Evacuating all placement groups (PGs) from the OSD #. Removing the PG-free OSD from the cluster

The following command performs these two steps:

.. prompt:: bash #

ceph orch osd rm <osd_id(s)> [--replace] [--force] [--zap]

Example:

.. prompt:: bash #

ceph orch osd rm 0 ceph orch osd rm 1138 --zap

Expected output:

.. code-block:: console

Scheduled OSD(s) for removal

OSDs that are not safe to destroy will be rejected. Adding the --zap flag directs the orchestrator to remove all LVM and partition information from the OSD's drives, leaving it a blank slate for redeployment or other reuse.

.. note:: After removing OSDs, if the OSDs' drives become available, cephadm may automatically try to deploy more OSDs on these drives if they match an existing drivegroup spec. If you deployed the OSDs you are removing with a spec and don't want any new OSDs deployed on the drives after removal, it's best to modify the drivegroup spec before removal. Either set unmanaged: true to stop it from picking up new drives, or modify it in some way that it no longer matches the drives used for the OSDs you wish to remove. Then re-apply the spec. For more info on drivegroup specs see :ref:drivegroups. For more info on the declarative nature of cephadm in reference to deploying OSDs, see :ref:cephadm-osd-declarative.

Monitoring OSD State During OSD Removal

You can query the state of OSD operations during the process of removing OSDs by running the following command:

.. prompt:: bash #

ceph orch osd rm status

Expected output:

.. code-block:: console

OSD_ID HOST STATE PG_COUNT REPLACE FORCE STARTED_AT 2 cephadm-dev done, waiting for purge 0 True False 2020-07-17 13:01:43.147684 3 cephadm-dev draining 17 False True 2020-07-17 13:01:45.162158 4 cephadm-dev started 42 False True 2020-07-17 13:01:45.162158

When no PGs are left on the OSD, it will be decommissioned and removed from the cluster.

.. note:: After removing an OSD, if you wipe the LVM physical volume in the device used by the removed OSD, a new OSD will be created. For more information on this, read about the unmanaged parameter in :ref:cephadm-osd-declarative.

Stopping OSD Removal

It is possible to stop queued OSD removals by using the following command:

.. prompt:: bash #

ceph orch osd rm stop <osd_id(s)>

Example:

.. prompt:: bash #

ceph orch osd rm stop 4

Expected output:

.. code-block:: console

Stopped OSD(s) removal

This resets the state of the OSD and takes it off the removal queue.

.. _cephadm-replacing-an-osd:

Replacing an OSD

.. prompt:: bash #

ceph orch osd rm <osd_id(s)> --replace [--force]

Example:

.. prompt:: bash #

ceph orch osd rm 4 --replace

Expected output:

.. code-block:: console

Scheduled OSD(s) for replacement

This follows the same procedure as the procedure in the :ref:cephadm-osd-removal section, with one exception: the OSD is not permanently removed from the CRUSH hierarchy, but is instead assigned the destroyed flag.

.. note:: The new OSD that will replace the removed OSD must be created on the same host as the OSD that was removed.

Preserving the OSD ID

The destroyed flag is used to determine which OSD IDs will be reused in the next OSD deployment.

If you use :ref:OSDSpecs <drivegroups> for OSD deployment, your newly added drives will be assigned the OSD IDs of their replaced counterparts. This assumes that the new drives still match the OSDSpecs.

Use the --dry-run flag to ensure that the ceph orch apply osd command will do what you intend. The --dry-run flag shows what the outcome of the command will be without executing any changes. When you are satisfied that the command will do what you want, run the command without the --dry-run flag.

.. tip::

The name of your OSDSpec can be retrieved with the command ceph orch ls.

Alternatively, you can use an OSDSpec file:

.. prompt:: bash #

ceph orch apply -i <osd_spec_file> --dry-run

Expected output:

.. code-block:: console

NAME HOST DATA DB WAL <name_of_osd_spec> node1 /dev/vdb - -

When this output reflects your intent, omit the --dry-run flag to execute the deployment.

Erasing Devices (Zapping Devices)

Erase (zap) a device so that it can be reused. zap calls ceph-volume zap on the remote host.

.. prompt:: bash #

ceph orch device zap <hostname> <path>

Example command:

.. prompt:: bash #

ceph orch device zap my_hostname /dev/sdx

.. note:: If the unmanaged flag is not set, cephadm automatically deploys drives that match the OSDSpec. For example, if you specify the all-available-devices option when creating OSDs, when you zap a device the cephadm orchestrator automatically creates a new OSD on the device. To disable this behavior, see :ref:cephadm-osd-declarative.

.. _osd_autotune:

Automatically Tuning OSD Memory

OSD daemons will adjust their memory consumption based on the :confval:osd_memory_target config option. If Ceph is deployed on dedicated nodes that are not sharing memory with other services, cephadm will automatically adjust the per-OSD memory consumption target based on the total amount of RAM and the number of deployed OSDs. This allows the full use of available memory, and adapts when OSDs or RAM are added or removed.

.. warning:: Cephadm sets :confval:osd_memory_target_autotune to true by default which is usually not appropriate for converged architectures, where a given node is used for both Ceph and compute purposes.

Cephadm will use a fraction :confval:mgr/cephadm/autotune_memory_target_ratio of available memory, subtracting memory consumed by non-autotuned daemons (non-OSDs and OSDs for which :confval:osd_memory_target_autotune is false), and then divide the balance by the number of OSDs.

The final targets are reflected in the config database with options like the below:

.. code-block:: console

WHO MASK LEVEL OPTION VALUE osd host:foo basic osd_memory_target 126092301926 osd host:bar basic osd_memory_target 6442450944

Both the limits and the current memory consumed by each daemon are visible from the ceph orch ps output in the MEM LIMIT column:

.. code-block:: console

NAME HOST PORTS STATUS REFRESHED AGE MEM USED MEM LIMIT VERSION IMAGE ID CONTAINER ID
osd.1 dael running (3h) 10s ago 3h 72857k 117.4G 17.0.0-3781-gafaed750 7015fda3cd67 9e183363d39c
osd.2 dael running (81m) 10s ago 81m 63989k 117.4G 17.0.0-3781-gafaed750 7015fda3cd67 1f0cc479b051
osd.3 dael running (62m) 10s ago 62m 64071k 117.4G 17.0.0-3781-gafaed750 7015fda3cd67 ac5537492f27

To exclude an OSD from memory autotuning, disable the autotune option for that OSD and also set a specific memory target. For example,

.. prompt:: bash #

ceph config set osd.123 osd_memory_target_autotune false ceph config set osd.123 osd_memory_target 16G

.. _drivegroups:

Advanced OSD Service Specifications

:ref:orchestrator-cli-service-spec\s of type osd provide a way to use the properties of drives to describe a Ceph cluster's layout. Service specifications are an abstraction used to tell Ceph which drives to transform into OSDs and which configurations to apply to those OSDs. :ref:orchestrator-cli-service-spec\s make it possible to target drives for transformation into OSDs even when the Ceph cluster operator does not know the specific device names and paths associated with those disks.

:ref:orchestrator-cli-service-spec\s make it possible to define a YAML or JSON file that can be used to reduce the amount of manual work involved in creating OSDs.

.. note:: We recommend that advanced OSD specs include the service_id field. OSDs created using ceph orch daemon add or ceph orch apply osd --all-available-devices are placed in the plain osd service. Failing to include a service_id in your OSD spec causes the Ceph cluster to mix the OSDs from your spec with those OSDs, which can potentially result in the overwriting of service specs created by cephadm to track them. Newer versions of cephadm block OSD specs that do not include the service_id.

For example, instead of running the following command:

.. prompt:: bash [monitor.1]#

ceph orch daemon add osd <host>:<path-to-device>

for each device and each host, we can create a .yaml or .json file that allows us to describe the layout. Here is the most basic example:

Create a file called (for example) osd_spec.yml:

.. code-block:: yaml

service_type: osd
service_id: default_drive_group  # custom name of the osd spec
placement:
  host_pattern: '*'              # which hosts to target
spec:
  data_devices:                  # the type of devices you are applying specs to
    all: true                    # a filter, check below for a full list

This means :

#. Turn any available device (ceph-volume decides which are available) into an OSD on all hosts that match the glob pattern '*'. The glob pattern matches registered hosts from ceph orch host ls. See :ref:cephadm-services-placement-by-pattern-matching for more on using host_pattern matching to use devices for OSDs.

#. Pass osd_spec.yml to osd create by using the following command:

.. prompt:: bash [monitor.1]#

 ceph orch apply -i /path/to/osd_spec.yml

This specification is applied to all the matching hosts to deploy OSDs.

Strategies more complex than the one specified by the all filter are possible. See :ref:osd_filters for details.

A --dry-run flag can be passed to the apply osd command to display a synopsis of the proposed layout.

Example

.. prompt:: bash [monitor.1]#

ceph orch apply -i /path/to/osd_spec.yml --dry-run

.. _osd_filters:

Filters

.. note:: Filters are applied using an AND operation by default. This means that a drive must match all filter criteria to be selected. This behavior can be adjusted by setting filter_logic: OR in the OSD specification.

Filters are used to select sets of drives for OSD data or WAL+DB offload based on various attributes. These attributes are gathered by ceph-volume's drive inventory. Retrieve these attributes with this command:

.. prompt:: bash #

ceph-volume inventory </path/to/drive>

Vendor or Model ^^^^^^^^^^^^^^^

Specific drives can be targeted by vendor brand, manufacturer) or model (SKU):

.. code-block:: yaml

model: drive_model_name

or

.. code-block:: yaml

vendor: drive_vendor_name

Size ^^^^

Specific drive capacities can be targeted with size:

.. code-block:: yaml

size: size_spec

Size specs


Size specifications can be of the following forms:

  • LOW:HIGH
  • :HIGH
  • LOW:
  • EXACT

We explore examples below.

To match only drives of an exact capacity:

.. code-block:: yaml

size: '10T'

Note that drive capacity is often not an exact multiple of units, so it is often best practice to match drives within a range of sizes as shown below. This handles future drives of the same class that may be of a different model and thus slightly different in size. Or say you have 10 TB drives today but may add 16 TB drives next year:

.. code-block:: yaml

size: '10T:40T'

To match only drives that are less than or equal to 1701 GB in size:

.. code-block:: yaml

size: ':1701G'

To include drives equal to or greater than 666 GB in size:

.. code-block:: yaml

size: '666G:'

The supported units of size are Megabyte (M), Gigabyte (G) and Terabyte (T). The B (byte) suffix for units is also acceptable: MB, GB, TB.

Rotational ^^^^^^^^^^

This gates based on the 'rotational' attribute of each drive, as indicated by the kernel. This attribute is usually as expected for bare HDDs and SSDs installed in each node. Exotic or layered device presentations may however be reported differently than you might expect or desire:

  • Network-accessed SAN LUNs attached to the node
  • Composite devices presented by dCache, Bcache, OpenCAS, etc.

In such cases you may align the kernel's reporting with your expectations by adding a udev rule to override the default behavior. The below rule was used for this purpose to override the rotational attribute on OSD nodes with no local physical drives and only attached SAN LUNs. It is not intended for deployment in all scenarios; you will have to determine what is right for your systems. If by emplacing such a rule you summon eldritch horrors from beyond spacetime, that's on you.

.. code-block:: none

ACTION=="add|change", KERNEL=="sd[a-z]*", ATTR{queue/rotational}="0"
ACTION=="add|change", KERNEL=="dm*", ATTR{queue/rotational}="0"
ACTION=="add|change", KERNEL=="nbd*", ATTR{queue/rotational}="0"

Spec file syntax:

.. code-block:: yaml

rotational: 0 | 1

1 to match all drives that the kernel indicates are rotational

0 to match all drives that are non-rotational (SATA, SATA, NVMe SSDs, SAN LUNs, etc)

All ^^^

This matches all drives that are available, i.e. they are free of partitions, GPT labels, etc.

.. note:: This may only be specified for data_devices.

.. code-block:: yaml

all: true

Limiter ^^^^^^^

If filters are specified but you wish to limit the number of drives that they match, use the limit attribute. This is useful when one uses some drives for non-Ceph purposes, or when multiple OSD strategies are intended.

.. code-block:: yaml

limit: 2

For example, when using vendor to match all drives branded VendorA but you wish to use at most two of them per host as OSDs, specify a limit:

.. code-block:: yaml

data_devices:
  vendor: VendorA
  limit: 2

.. note:: limit is usually appropriate in only certain specific scenarios.

Additional Options

There are multiple optional settings that specify the way OSDs are deployed. Add these options to an OSD spec for them to take effect.

This example deploys encrypted OSDs on all unused drives. Note that if Linux MD mirroring is used for the boot, /var/log, or other volumes this spec may grab replacement or added drives before you can employ them for non-OSD purposes. The unmanaged attribute may be set to pause automatic deployment until you are ready.

.. code-block:: yaml

service_type: osd
service_id: example_osd_spec
placement:
  host_pattern: '*'
spec:
  data_devices:
    all: true
  encrypted: true

Ceph Squid onwards support TPM2 token enrollment for LUKS2 devices. Add the tpm2 attribute to the OSD spec:

.. code-block:: yaml

service_type: osd
service_id: example_osd_spec_with_tpm2
placement:
  host_pattern: '*'
spec:
  data_devices:
    all: true
  encrypted: true
  tpm2: true

A full list of supported attributes:

.. py:currentmodule:: ceph.deployment.drive_group

.. autoclass:: DriveGroupSpec :members: :exclude-members: from_json

Examples

The Simple Case

When all cluster nodes have identical drives and we wish to use them all as OSDs with offloaded WAL+DB:

.. code-block:: none

10 HDDs
Vendor: VendorA
Model: HDD-123-foo
Size: 4TB

2 SAS/SATA SSDs
Vendor: VendorB
Model: MC-55-44-ZX
Size: 512GB

This is a common arrangement and can be described easily:

.. code-block:: yaml

service_type: osd
service_id: osd_spec_default
placement:
  host_pattern: '*'
spec:
  data_devices:
    model: HDD-123-foo       # Note, HDD-123 would also be valid
  db_devices:
    model: MC-55-44-XZ       # Same here, MC-55-44 is valid

However, we can improve the OSD specification by filtering based on properties of the drives instead of specific models, as models may change over time as drives are replaced or added:

.. code-block:: yaml

service_type: osd
service_id: osd_spec_default
placement:
  host_pattern: '*'
spec:
  data_devices:
    rotational: 1            # The kernel flags as HDD
  db_devices:
    rotational: 0            # The kernel flags as SSD (SAS/SATA/NVMe)

Here designate all HDDs to be data devices (OSDs) and all SSDs to be used for WAL+DB offload.

If you know that drives larger than 2 TB should always be used as data devices, and drives smaller than 2 TB should always be used as WAL/DB devices, you can filter by size:

.. code-block:: yaml

service_type: osd
service_id: osd_spec_default
placement:
  host_pattern: '*'
spec:
  data_devices:
    size: '2TB:'           # Drives larger than 2 TB
  db_devices:
    size: ':2TB'           # Drives smaller than 2TB

.. note:: All of the above OSD specs are equally valid. Which you use depends on taste and on how much you expect your node layout to change.

Multiple OSD Specs for a Single Host

Here we specify two distinct strategies for deploying OSDs across multiple types of media, usually for use by separate pools:

.. code-block:: none

10 HDDs
Vendor: VendorA
Model: HDD-123-foo
Size: 4TB

12 SAS/SATA SSDs
Vendor: VendorB
Model: MC-55-44-ZX
Size: 512GB

2 NVME SSDs
Vendor: VendorC
Model: NVME-QQQQ-987
Size: 256GB
  • 10 HDD OSDs use 2 SATA/SAS SSDs for WAL+DB offload
  • 10 SATA/SAS SSD OSDs share 2 NVMe SSDs for WAL+DB offload

This can be specificed with two service specs in the same file:

.. code-block:: yaml

service_type: osd
service_id: osd_spec_hdd
placement:
  host_pattern: '*'
spec:
  data_devices:             # Select all drives the kernel identifies as HDDs
    rotational: 1           #  for OSD data
  db_devices:
    model: MC-55-44-XZ      # Select only this model for WAL+DB offload
    limit: 2                # Select at most two for this purpose
  db_slots: 5               # Chop the DB device into this many slices and
                            #  use one for each of this many HDD OSDs
---
service_type: osd
service_id: osd_spec_ssd    # Unique so it doesn't overwrite the above
placement:
  host_pattern: '*'
spec:                       # This scenario is uncommon
  data_devices:
    model: MC-55-44-XZ      # Select drives of this model for OSD data
  db_devices:               # Select drives of this brand for WAL+DB. Since the
    vendor: VendorC         #   data devices are SAS/SATA SSDs this would make sense for NVMe SSDs
  db_slots: 2               # Back two slower SAS/SATA SSD data devices with each NVMe slice

This would create the desired layout by using all HDDs as data devices with two SATA/SAS SSDs assigned as dedicated DB/WAL devices, each backing five HDD OSDs. The remaining ten SAS/SATA SSDs will be used as OSD data devices, with VendorC NVMEs SSDs assigned as dedicated DB/WAL devices, each serving two SAS/SATA OSDs. We call these _hybrid OSDs.

Multiple Hosts with the same Disk Layout

When a cluster comprises hosts with different drive layouts, or a complex constellation of multiple media types, it is recommended to apply multiple OSD specs, each matching only one set of hosts. Typically you will have a single spec for each type of host.

The service_id must be unique: if a new OSD spec with an already applied service_id is applied, the existing OSD spec will be superseded. Cephadm will then create new OSD daemons on unused drives based on the new spec definition. Existing OSD daemons will not be affected. See :ref:cephadm-osd-declarative.

Example:

Nodes 1-5:

.. code-block:: none

20 HDDs
Vendor: VendorA
Model: SSD-123-foo
Size: 4TB
2 SSDs
Vendor: VendorB
Model: MC-55-44-ZX
Size: 512GB

Nodes 6-10:

.. code-block:: none

5 NVMEs
Vendor: VendorA
Model: SSD-123-foo
Size: 4TB
20 SSDs
Vendor: VendorB
Model: MC-55-44-ZX
Size: 512GB

You can specify a placement to target only certain nodes.

.. code-block:: yaml

service_type: osd
service_id: disk_layout_a
placement:
  label: disk_layout_a
spec:
  data_devices:
    rotational: 1           # All drives identified as HDDs
  db_devices:
    rotational: 0           # All drives identified as SSDs
---
service_type: osd
service_id: disk_layout_b
placement:
  label: disk_layout_b
spec:
  data_devices:
    model: MC-55-44-XZ      # Only this model
  db_devices:
    model: SSD-123-foo      # Only this model

This applies different OSD specs to different hosts that match hosts tagged with ceph orch labels via the placement filter. For more information, see :ref:orchestrator-cli-placement-spec.

.. note::

Assuming each host has a unique disk layout, each OSD 
spec must have a unique ``service_id``.

Dedicated WAL + DB

All previous cases colocated the WALs with the DBs. It is however possible to deploy the WAL on a separate device if desired.

.. code-block:: none

20 HDDs
Vendor: VendorA
Model: SSD-123-foo
Size: 4TB

2 SAS/SATA SSDs
Vendor: VendorB
Model: MC-55-44-ZX
Size: 512GB

2 NVME SSDs
Vendor: VendorC
Model: NVME-QQQQ-987
Size: 256GB

The OSD spec for this case would look like the following, using the model filter:

.. code-block:: yaml

service_type: osd
service_id: osd_spec_default
placement:
  host_pattern: '*'
spec:
  data_devices:
    model: MC-55-44-XZ
  db_devices:
    model: SSD-123-foo
  wal_devices:
    model: NVME-QQQQ-987

It is also possible to specify device paths as below, when every matched host is expected to present devices identically.

.. code-block:: yaml

service_type: osd
service_id: osd_using_paths
placement:
  hosts:
    - node01
    - node02
spec:
  data_devices:
    paths:
    - /dev/sdb
  db_devices:
    paths:
    - /dev/sdc
  wal_devices:
    paths:
    - /dev/sdd

In most cases it is preferable to accomplish this with other filters including size or vendor so that OSD services adapt when Linux or an HBA may enumerate devices differently across boots, or when drives are added or replaced.

It is possible to specify a crush_device_class parameter to be applied to OSDs created on devices matched by the paths filter:

.. code-block:: yaml

service_type: osd
service_id: osd_using_paths
placement:
  hosts:
    - node01
    - node02
crush_device_class: ssd
spec:
  data_devices:
    paths:
    - /dev/sdb
    - /dev/sdc
  db_devices:
    paths:
    - /dev/sdd
  wal_devices:
    paths:
    - /dev/sde

The crush_device_class attribute may be specified at OSD granularity via the paths keyword with the following syntax:

.. code-block:: yaml

service_type: osd
service_id: osd_using_paths
placement:
  hosts:
    - node01
    - node02
crush_device_class: ssd
spec:
  data_devices:
    paths:
    - path: /dev/sdb
      crush_device_class: ssd
    - path: /dev/sdc
      crush_device_class: nvme
  db_devices:
    paths:
    - /dev/sdd
  wal_devices:
    paths:
    - /dev/sde

.. _cephadm-osd-activate:

Activate Existing OSDs

If a host's operating system has been reinstalled, existing OSDs must be activated again. cephadm provides a wrapper for :ref:ceph-volume-lvm-activate that activates all existing OSDs on a host.

The following procedure explains how to use cephadm to activate OSDs on a host that has had its operating system reinstalled.

This example applies to two hosts: ceph01 and ceph04.

  • ceph01 is a host equipped with an admin keyring.
  • ceph04 is the host with the recently reinstalled operating system.

#. Install cephadm and podman on the host. The command for installing these utilities will depend upon the operating system of the host.

#. Retrieve the public key.

.. prompt:: bash ceph01#

  cd /tmp ; ceph cephadm get-pub-key > ceph.pub

#. Copy the key (from ceph01) to the freshly reinstalled host (ceph04):

.. prompt:: bash ceph01#

  ssh-copy-id -f -i ceph.pub root@<hostname>

#. Retrieve the private key in order to test the connection:

.. prompt:: bash ceph01#

  cd /tmp ; ceph config-key get mgr/cephadm/ssh_identity_key > ceph-private.key

#. From ceph01, modify the permissions of ceph-private.key:

.. prompt:: bash ceph01#

  chmod 400 ceph-private.key

#. Log in to ceph04 from ceph01 to test the connection and configuration:

.. prompt:: bash ceph01#

  ssh -i /tmp/ceph-private.key ceph04

#. While logged into ceph01, remove ceph.pub and ceph-private.key:

.. prompt:: bash ceph01#

  cd /tmp ; rm ceph.pub ceph-private.key

#. If you run your own container registry, instruct the orchestrator to log into it on each host:

.. prompt:: bash #

  ceph cephadm registry-login my-registry.domain <user> <password>

When the orchestrator performs the registry login, it will attempt to deploy any missing daemons to the host. This includes crash, node-exporter, and any other daemons that the host ran before its operating system was reinstalled.

To be clea: cephadm attempts to deploy missing daemons to all hosts managed by cephadm, when cephadm determines that the hosts are online. In this context, "online" means "present in the output of the ceph orch host ls command and with a status that is not offline or maintenance. If it is necessary to log in to the registry in order to pull the images for the missing daemons, then deployment of the missing daemons will fail until the process of logging in to the registry has been completed.

.. note:: This step is not necessary if you do not run your own container registry. If your host is still in the "host list", which can be retrieved by running the command ceph orch host ls, you do not need to run this command.

#. Activate the OSDs on the host that has recently had its operating system reinstalled:

.. prompt:: bash #

  ceph cephadm osd activate ceph04

This command causes cephadm to scan all existing disks for OSDs. This command will make cephadm deploy any missing daemons to the host specified.

This procedure was developed by Eugen Block in Feburary of 2025, and a blog post pertinent to its development can be seen here: Eugen Block's "Cephadm: Activate existing OSDs" blog post <https://heiterbiswolkig.blogs.nde.ag/2025/02/06/cephadm-activate-existing-osds/>_.

.. note:: It is usually not safe to run ceph orch restart osd.myosdservice on a running cluster, as attention is not paid to CRUSH failure domains, and parallel OSD restarts may lead to temporary data unavailability or in rare cases even data loss.

Further Reading

  • :ref:ceph-volume
  • :ref:rados-index