Migrate Ceph Deployment Architecture

The all-in-one Ceph charm has been deprecated in favor of splitting out the distinct functions of the Ceph monitor cluster and the Ceph OSDs. The Ceph charm itself is receiving limited maintenance and all new features are being added to the ceph-osd or ceph-mon charms. Unfortunately, there currently exists no way of migrating existing deployments using the legacy architecture to the preferred architecture.

Problem Description

Deploying the all-in-one ceph charm was difficult to deploy properly and often resulted in too many monitors deployed. This (and other) problems eventually lead to splitting up the all-in-one ceph deployment into its different compositional parts - ceph-mon and ceph-osd. New deployments are encouraged to use the new architecture, but users which have deployed the previous architecture have no way to migrate to the new architecture.

In order to migrate to the new architecture, users will need to:

  1. Deploy the ceph-mon charm into the environment without bootstrapping a new monitor cluster.
  2. Relate the ceph application to the new ceph-mon application. The relation between these two applications is defined in the new ceph-bootstrap interface, documented herein.
  3. Update the deployments using the ceph-client relation to point to the new ceph-mon application. This should include:
    • ceph-radosgw
    • nova-compute
    • cinder-ceph
    • cinder
    • cinder-backup
    • glance
  4. Deploy the ceph-osd charm alongside the current ceph charms. The ceph-osd charm will not reformat any of the currently running OSD devices.
  5. Remove the all-in-one ceph application.

Note: the configuration values will not be validated in any part of this migration scenario. If the user deploys the ceph-osd application or the ceph-mon application with different config values than that of the all-in-one ceph charms, then there may be unexpected changes to the environment.

Proposed Change

Most of the basic plumbing already exists to support the basic parts of this migration. There are some parts that are actually missing an need attention to detail.

Config Changes to charm-ceph-mon

A new config option will be added to the ceph-mon charm indicating that the charm should not do an initial bootstrap of the monitor cluster. When this value is set to true, the charm will not:

  1. bootstrap a new monitor cluster
  2. automatically generate an fsid

This new configuration option will be defined as follows:

no-bootstrap:
  type: boolean
  default: False
  description: |
    Causes the charm to not do any of the initial bootstrapping of the
    Ceph monitor cluster. This is only intended to be used when migrating
    from the ceph all-in-one charm to a ceph-mon / ceph-osd deployment.
    Refer to the Charm Deployment guide at https://docs.openstack.org/charm-deployment-guide/latest/
    for more information.

New ceph-bootstrap Interface

Notably, there exists no way of adding a relation between the ceph-mon charm and the all-in-one ceph charm. To address this, a new interface will be added to the ceph and ceph-mon charms called ceph-bootstrap, enabling the necessary information for joining the cluster to be shared. This is essentially the same information that’s shared on the peer mon interface for the ceph charm itself.

Since the charms are not reactive, no new interface repository is required. The information exchanged will contain the following:

ceph-bootstrap:
  - name: fsid
    type: string
    desc: |
      The fsid of the already bootstrapped monitor cluster

  - name: ceph-public-address
    type: string
    desc: |
      The public address that should be used by the charm for each of the
      units in the relation.

  - name: mon-key
    type: string
    desc: |
      The key used to authorize a monitor node for joining a ceph mon
      cluster.

In order to join this relation, the ceph-mon charm will require that the no-bootstrap config option be set to True and the monitor cluster is not already bootstrapped locally. If either of these is not valid, the ceph-mon charm will fail to join the relation.

Changes to charm-ceph

In addition to implementing the ceph-bootstrap interface, the all-in-one ceph charm needs to properly clean up when removing itself. It should not remove any OSDs or Ceph packages as this would interrupt the operational ceph cluster. However, the charm needs to remove its ceph.conf file as a registered alternative during its stop hook.

Alternatives

One alternative would be to not offer a migration to the new architecture and leave the deployments stuck. This has rather unfortunate side effect with respect to user happiness and has implications regarding the overall lifetime of the ceph charm.

Implementation

Assignee(s)

Primary assignee:
james-page
Additional assignee(s):
billy-olsen

Gerrit Topic

Use Gerrit topic charm-ceph-migration for all patches related to this spec.

git-review -t charm-ceph-migration

Work Items

charm-ceph-mon

  • Add new config flag
  • Implement new bootstrap interface
  • Remove additional monmap entries when removing the bootstrap relation

charm-ceph-osd

  • Ensure that the charm supports multiple monitor relations

charm-ceph

  • Add support for new bootstrap interface
  • Remove registered alternative for ceph.conf in the stop hook
  • Update charm README

charm-ceph-radosgw

  • Ensure that the charm supports multiple monitor relations

charm-helpers

  • Ensure that CephContext supports multiple monitor relations (consumed by charm-glance, charm-inder-ceph, charm-nova, etc)

openstack-charm-testing

  • Update all next.yaml and stable.yaml bundles to use split architecture
  • Add legacy bundle which deploys an all-in-one ceph bundle

docs

  • Update charm deployment guide to clearly describe this process
  • Update release notes with implementation note and reference to the deployment guide.

Repositories

None

Documentation

This will need to be carefully documented in the following places:

  1. Charm Deployment Guide
  2. The Ceph Charm’s README needs to refer to the migration process documentation and officially marked as deprecated.

Security

No new security implications or this change.

Testing

  • Create an appropriate functional test that can be run at release gating (mojo, bundle tester, etc)
  • Test the impact of running clients on the migration. This is important because the monitor information is exposed via the libvirt domain XML files which are created and may have an impact on running clients. Specifically, need to verify the impact to: - nova-compute w/ rbd backed instances - nova-compute w/ rbd cinder attached volumes - glance images w/ rbd backing

Dependencies

None