Allow provider re-parenting in placement

https://storyboard.openstack.org/#!/story/2008764

This spec proposes to allow re-parenting and un-parenting (or orphaning) RPs via PUT /resource_providers/{uuid} API in Placement.

Problem description

Today placement API only allows change the parent of an RP from None to a valid RP UUID. However there are use case when moving an RP between parents make sense.

Use Cases

• An existing PGPU RP needs to be moved under the NUMA RP when NUMA is modeled.

• We have a neutron bug that introduced an unwanted change causing that SRIOV PF RPs was created under the root RP instead of under the neutron agent RP. We can fix the broken logic in neutron but we cannot fix the already wrongly parented RP in the DB via the placement API.

Proposed change

Re-parenting is rejected today and the code has the following comment :

TODO(jaypipes): For now, “re-parenting” and “un-parenting” are not possible. If the provider already had a parent, we don’t allow changing that parent due to various issues, including:

• if the new parent is a descendant of this resource provider, we introduce the possibility of a loop in the graph, which would be very bad

• potentially orphaning heretofore-descendants

So, for now, let’s just prevent re-parenting…

The first reason is moot as the loop check is already needed and implemented for the case when the parent is updated from None to an RP.

The second reason does not make sense to me. By moving an RP under another RP all the descendants should be moved as well. Similarly how the None -> UUID case works today. So I don’t see how can we orphan any RP by re-parenting.

I see the following possible cases of move:

• RP moved upwards, downwards, side-wards in the same RP tree

• RP moved to a different tree

• RP moved to top level, becoming a new root RP

From placement perspective every case results in one or more valid RP trees.

Based on the data model if there was allocations against the moved RP those allocations will still refer to the RP after the move. This means that if a consumer has allocation against a single RP tree before the move might have allocation against multiple trees after the RP move. Such consumer is already supported today.

An RP move might invalidate the original intention of the consumer. If the consumer used an allocation candidate query to select and allocate resources then by such query the consumer defined a set of rules (e.g. in_tree, same_subtree) the allocation needs to fulfill. The rules might not be valid after an RP is moved. However placement never promised to keep such invariant as that would require the storage of the rules and correlating allocation candidate queries and allocations. Moreover such issue can already be created with the POST /reshape API as well. Therefore keeping any such invariant is the responsibility of the client. So I propose to start supporting all form of RP re-parenting in a new placement API microversion.

Alternatives

See the API alternatives below.

Data model impact

None

REST API impact

In a new microversion allow changing the parent_uuid of a resource provider to None or to any valid RP uuid that does not cause a loop in any of the trees via the PUT /resource_providers/{uuid} API.

Protecting against unwanted changes

As noted above re-parenting can significantly change the RP model in the Placement database. So such action needs to be done carefully. While the Placement API is already admin only by default, the request is raised on the Xena PTG for extra safety measures against unintentional parent changes. During the spec discussion every the reviewer expressed the view that such safety measure is not really needed. So this spec only propose to use the new microversion and extensive documentation to signal the new behavior.

Still there is the list of alternatives discussed during the review:

• Do nothing: While it is considered not safe enough during the PTG, during the spec review we ended up choosing this as the main solution.

• A new query parameter: A new query parameter is proposed for the PUT /resource_providers/{uuid} API called allow_reparenting the default value of the query parameter is False and the re-parenting cases defined in this spec is only accepted by Placement if the request contains the new query parameter with the True. It is considered hacky to add a query parameter for a PUT request.

• A new field in the request body: This new field would have the same meaning as the proposed query parameter but it would be put into the request body. It is considered non-RESTful as such field is not persisted or returned as the result of the PUT request as it does not belong to the representation of the ResourceProvider entity the PUT request updates.

• A new Header: Instead of a new query paramtere use a new HTTP header x-openstack-placement-allow-provider-reparenting:True. As the name shows this needs a lot more context encoded in it to be specific for the API it modifies while the query parameter already totally API specific.

• Use a PATCH request for updating the parent: While this would make the parent change more explicit it would also cause great confusion for the client for multiple reasons:

  1. Other fields of the same resource provider entity can be updated via the PUT request, but not the parent_uuid field.

  2. Changing the parent_uuid field from None to a valid RP uuid is supported by the PUT request but to change it from one RP uuid to another would require a totally different PATCH request.

• Use a sub resource: Signal the explicit re-parenting either in a form of PUT /resource-providers/{uuid}/force or PUT /resource-providers/{uuid}/parent_uuid/{parent}. While the second option seems to be acceptable to multiple reviewers, I think it will be confusing similarly to PATCH. It would create another way to update a field of an entity while other fields still updated directly on the parent resource.

Security impact

None

Notifications impact

N/A

Other end user impact

None

Performance Impact

The loop detection and the possible update of all the RPs in the changed subtree with a new root_provider_id needs extra processing. However the re-parenting operation is considered very infrequent. So the overall Placement performance is not affected.

Other deployer impact

None

Developer impact

None

Upgrade impact

None

Implementation

Assignee(s)

Primary assignee:

balazs-gibizer

Feature Liaison

Feature liaison:

None

Work Items

• Add a new microversion to the Placement API. Implement an extended loop detection and update root_provider_id of the subtree if needed.

• Mark the new microversion osc-placement as supported.

Dependencies

None

Testing

• Unit testing

• Gabbit API testing

Documentation Impact

• API doc needs to be updated. Warn the user that this is a potentially dangerous operation.

References

None

History

Revisions

Release Name	Description
Xena	Introduced