cinder-specs/specs/kilo/cinder-objects.rst

Cinder Objects

https://blueprints.launchpad.net/cinder/+spec/cinder-objects

The goal of this spec is to introduce objects into cinder. An object is used to bundle data with methods that can operate on them. By implementing this spec, we would be able to pass rich objects over RPC and decouple the database schema implementation from SQLAlchemy, enabling support for NoSQL databases. Database schema independence is the biggest gain with this proposal.

Problem description

There are a few problems that exists today in cinder:

• The database implementation is directly tied to SQLAlchemy. This limits support for NoSQL databases. It is a major reason behind this proposal.
• Cinder sends ids over RPC (e.g. volume id, snapshot id, consistency group id, or backup id). These ids are used to query the SQLAlchemy objects, resulting in a database hit.
• There are multiple places where database calls are made in cinder, and this could result in race conditions.

Use Cases

Proposed change

There is an approved proposal to put nova.objects into oslo as oslo.versionedobjects (https://review.openstack.org/#/c/127532/), which would abstract versionable internal objects for use with RPC. The plan is to adopt oslo.versionedobjects and abstract volumes, snapshots, backups, consistency groups, and quotas.

Since it will take some time before versionedobjects goes into the oslo library, the plan is to sync up with dansmith to get an early version of it for cinder and transition to oslo.versionedobjects when it is ready.

The proposed changes are:

• Objects will be database schema independent. The type of database connection can be determined from the value of "connection" in cinder.conf, which could be mysql, mongodb, db2, etc. Depending on the type of connection, the appropriate objects (e.g. cinder.objects.volume or cinder.objects.nosql.volume) would be imported when objects.register_all() is called.
• Objects will be sent over RPC instead of ids, reducing the additional database hits to query the SQLAlchemy objects.
• Objects will be versioned, allowing older objects to be detected and taking appropriate actions when the objects are incompatible, i.e. fail gracefully or emulate older behavior.
• To handle cases where the underlying volume, snapshot, backup, consistency group, or quota could have been changed when the object is in flight, additional arguments (e.g. expected_volume_state and expected_task_state) would be included in the object save method, which would check against the in-database copy of the volume, snapshot, backup, consistency group, or quota before updates are made and fail gracefully or appropriately sync up the data.

The work will be done incrementally in stages, as follows.

• Obtain an early version of oslo.versionedobjects to be used in cinder.
• Abstract volume snapshot into versionedobject.
• Abstract volume backup into versionedobject.
• Iteratively convert cinder internals to use volume snapshot object.

The goal is to have oslo.versionedobjects code base in by the end of kilo-2, and the abstraction of a few cinder internals, e.g. volume snapshots and backups, by that time also.

Alternatives

Cinder could continue to be tightly tied to SQLAlchemy, limiting support to only SQL databases. Also, the existing implementation of sending ids over RPC can be left as-is.

There are proposals for state machine/state enforcement and RPC versioning, which would help with rolling upgrades and improve the stability of cinder. However, you would still take an additional database hit when the volume, snapshot, backup, or consistency group ids are used to query the SQLAlchemy objects.

Data model impact

None. Cinder objects provides an abstraction layer in cinder, allowing for database schema independence. These objects are a replacement for SQLAlchemy objects that are being used to represent service, volume, volume type, snapshot, quota, backup, consistency group and consistency group snapshot throughout cinder internals.

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

Unknown. There might be a minor hit in performance since SQLAlchemy objects are abstracted into cinder objects.

Other deployer impact

None

Developer impact

Today, direct database calls are made to create, read, update, and delete volumes, volume types, snapshots, quotas, backups, consistency groups and consistency group snapshots, and others. With the switch to an object model, these calls will be made against the appropriate object methods. It will take some time to convert cinder internals over to the object model, so the existing convention of direct database calls should be accepted until all object models are in place.

Implementation

Assignee(s)

Primary assignee:

thang-pham

Other contributors:

Any help is welcome

Work Items

• Obtain an early version of oslo.versionedobjects to be used in cinder.
• Abstract volume snapshot into versionedobject.
• Abstract volume backup into versionedobject.
• Iteratively convert cinder internals to use volume snapshot object.

Dependencies

None

Testing

By introducing objects, we are not changing the end user facing API or functionality of cinder. We are passing objects over RPC instead of ids. Given this, no new tempest tests will be created. However, relevant unit tests will be created to test the new object code base.

Documentation Impact

None

References

• cinder-specs: https://review.openstack.org/#/c/130044/
• oslo-specs: https://review.openstack.org/#/c/127532/
• Proof of concept: https://review.openstack.org/#/c/131873/