Introduce new volume attach APIs

https://blueprints.launchpad.net/cinder/+spec/add-new-attach-apis

Problem description

The current volume attach process for Cinder Volumes is somewhat inconsistent and rather unclear in terms of exactly how it works. Some of this has been a result of growth and feature adds, things like optional parameters have made the process a bit more complex than it needs to be.

The interpretation of things like initialize_connection has sort of diverged a bit, and due to some requirements for different backends things have become a bit hacky. We have the potential for race conditions during the attach/detach process and most of all we seem to have a situation where adding things like multi-attach support are extremely complex and difficult.

Rather than continuing to add on to the monolith, it might be a good time to step back and simplify the flow altogether.

Use Cases

Attaching Cinder volumes to Nova Instances, Bare Metal or whatever.

Proposed change

Rather than using reserve, initialize and attach with numerous parameters to chain together an attach, and try to modify things to make multi-attach work this specs proposes that we instead introduce a new simplified and more robust set of attach API’s.

API Operations:

  • attachment_create

    This call is used to create a volume attachment.

    If the connector is not specified this call will only complete the volume reservation step and create an empty volume attachment. This can be used by Nova to call it from the API and start the volume attaching process by checking that the volume is available and creates the empty attachment. To finalize the operation Nova will call ‘attachment_update’.

    In case the connector is specified it will reserve the volume and create the connection between the volume and the host and update the attachment record. This can be used for instance for bare metal use cases, when the connector is available immediately.

    In both cases the call returns the attachment_id to the caller.

    Args:

    volume_id instance_uuid connector (optional)

    Returns:

    attachment_id

  • attachment_update

    This call is used to update an existing volume attachment. For this call the connector is mandatory.

    In case of an empty attachment it also finalizes attaching the volume and updates the attachment record accordingly.

    As the attachment exists in every scenario this call is invoked the attachment_id is mandatory.

    Args:

    attachment_id connector

    Returns:

    connection_info

  • attachment_get

    This call will fetch the specified attachment object (by ID). In addition this call will go out to the Cinder backend driver and query it for any shared connection info. The purpose being that we return the attachment object that was requested as well as a list of any shared attachment_id’s.

    Args:

    attachment_id

    Returns:

    Attachment info

    {id: <uuid-of-the-attachment-object>,

    project_id: <project-id-associated-with-attachment>, volume_id: <uuid-of-volume-associated-with-this-attachment-record>, instance_uuid: <uuid-of-instance-attachment-is-used-for>, attached_host: <hostname-of-compute-node-attached-to>, mountpoint: <mountpoint-inside-instance-that-is-expected-to-be-used>, attach_time: <time-we-created-the-attachment>, detach_time: <time-we-detached>, attach_status: <status-of-the-attachment>, attach_mode: <ro|rw>, shared_connections_by_attachment: [<attachment_id>, …]}

  • attachment_remove

    Removes/Detaches an existing attachment, marks a volume as available if no more attachments to it left.

    Args:

    attachment_id

    Returns:

    List of attachment_id’s for any shared connections

Alternatives

Keep trying to hack our way around the existing attach/detach flows we have. Deal with the split brain issue we have between Nova and Cinder.

It is possible that we could rewrite the existing methods and enforce stricter interface usage and we’d probably be ok, but the code has become pretty convoluted and ugly to deal with and it might be easier to wrap the existing calls up into a managing wrapper and then over time iterate over the internal pieces.

Data model impact

We need to store the connector when making attachments. As the connector can vary based on the back end and can also easily exceed 255 chars in length adding it to the ‘volume_attachment’ table is not a viable option. As opposed to that we’ll add an ‘attachment_specs’ table that will slurp in the k/v entries in the connector.

REST API impact

This will introduce at least the 4 new calls mentioned in the proposal section, these will be part of a new API object named attachment. We’ll treat the attachment as a first class object in Cinder.

The new API will be released under a new microversion.

Security impact

Notifications impact

Other end user impact

The old API’s will still work until we decide to deprecate them.

Performance Impact

Other deployer impact

Developer impact

For now this is avoiding changes to the drivers in Cinder and focusing on just the API for the attach/detach call up until the Manager layer.

Implementation

Condense initialize_connection and attach down into a single attachment_create call.

The intent here is that Cinder shouldn’t know or care so much about what a consumer is doing with a volume or attachment. We should only care that they desire to attach or detach and that’s about it. In order to do this properly we will need an attachment_id that the consumer should store and reference going forward when they are done with the attachment. This means it’s up to the consumer to decide when they are truly done and they want to destroy the attachment as opposed to now where we try and guess for them based on clues.

Assignee(s)

Primary assignee:
john-griffith

Work Items

  1. Implement API calls and wrappers in Cinder
  2. Implement cinderclient calls to expose them
  3. Implement changes in Nova

Dependencies

Testing

Once the API design is finalized the same Tempest test coverage will be added as we have for the current attach/detach API. In parallel to this activity the Nova implementation to pick up the new API shall be started as well which will add an additional layer of testing.

Documentation Impact

Will need to update any and all documents about how attach/detach of volumes in Cinder works.

References