Detach Service from Compute_Node

https://blueprints.launchpad.net/nova/+spec/detach-service-from-computenode

Remove the nested dependency in between Service and ComputeNode

Problem description

There is no good reason to keep a dependency in between a service, which is the representation of the message bus and a compute_node, which is a collection of resources for the solely use of the scheduler. The fact that they are related to each other means that the resource tracker ends up needing to “find” its compute node record by first looking up the service record for the ‘compute’ topic and the host for the resource tracker, and then grabs the first compute_node record that is related to the service record that matches that query. There is no reason to do this in the resource tracker other than the fact that right now the compute_nodes table has a service_id field and a relation to the services table.

It also carries a dependency on the compute_nodes table as there is a foreign key on a separate table for something totally unrelated to the Scheduler, which prevents the Scheduler to be split unless we continue to carry that relationship.

Use Cases

This is a refactoring effort helping out to split the scheduler by reducing the dependencies it has to manage.

Project Priority

This blueprint is part of the ‘scheduler’ refactoring effort, defined as a 3rd priority for the Kilo release.

Proposed change

Instead of having a relationship using a foreign key, the proposal will consist of adding a new field called ‘host’ for compute_nodes and a unique constraint on (host, hypervisor_hostname). Also, service_id field will be marked as deprecated and not updated in the compute_nodes table and ComputeNode object field service_id will be left unset. SQLA relationship on service will be deleted and Service object will keep a compute_node field but will actually not use this relationship.

Implementation proposal can be found in the patch series [1].

Alternatives

Only change DB API to remove the relationship without changing callers but it would create some confusion and obfuscate the need of modifying accessors.

Data model impact

Most of the change is about changing the model, but let’s rephrase it. compute_nodes.service relationship will be deleted, compute_nodes.service_id will be marked as deprecated and not updated by Kilo compute nodes and compute_nodes.host will be added as a String (identical to Service.host field).

As it was agreed during Summit, no data migrations will happen for updating either when creating the host column (for populating its values) or when downgrading by repopulating service_id.

Instead, data migration (here service_id to host) will be managed at the Object level (here ComputeNode) each time a save operation will happen by querying Service object to get the host value and set service_id to NULL.

There is no sense to keep a specific ID while the tuple (host, node) is identified as the source of truth for idenfifying compute nodes.

ComputeNode object will still continue to have a service field but it will no longer use the relationship to get that info. In parallel, Service object will continue to have a nested ComputeNode object for backwards compatibility but won’t also use the relationship to get that object.

REST API impact

In order to preserve API stability, we will still provide service information when querying compute nodes but this extra information will be on a case-by-case basis thanks to an extra flag passed to the DB API asking to join service and compute_nodes tables on service.host == compute_nodes.host. We expect no performance penalty as it is already done this way in db.compute_node_get_all() with an INTEGER matching instead of a VARCHAR(255).

No changes in the API model.

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

None

Other deployer impact

An external tool could be provided for migrating offline existing compute nodes which don’t yet have the host field set.

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

sbauza

Other contributors:

None

Work Items

Code was already posted as a patch series [1] :

  • Add host field to compute_nodes table

  • Add extra methods for querying this new field

  • Use these extra methods instead of querying Service for getting the node(s)

  • Make service info optional when querying compute nodes

  • Remove query by service_id on compute_nodes

  • Do not provide by default service info when querying compute nodes

  • Deprecate service_id field from compute_nodes and delete service relationship

Dependencies

None

Testing

Current Tempest and unittests already cover this.

Documentation Impact

None