Queens Project Priorities¶
List of efforts the Nova development team is prioritizing for reviews in the Queens release (in no particular order).
In Pike, the control plane was made multi-cell aware. However, there are some limitations. Priorities in Queens are related to removing some of those limitations.
Efficient multi-cell instance listing: Improve the performance of listing instances across multiple cells and merge sort the results. This is achieved by concurrently querying the cells and then sorting the results as they are processed.
Alternate hosts: Support rescheduling across compute hosts within a cell during the initial create or migration of an instance by having the scheduler provide a primary selected host and a list of alternate hosts. The alternate hosts will be used within the cell to reschedule in case the primary selected host fails to build / migrate the instance. This avoids the need for the cell conductor service to need to communicate back up to the scheduler service.
Migration allocations: Remove technical debt introduced in the Pike release with how resource allocations are tracked in the Placement service during a move operation (cold and live migrate). Rather than “double” allocations for an instance on both the source and destination compute node resource providers, the instance allocation will be created on the destination node and the current source node allocation will be temporarily tracked against the migration record until the operation completes, at which point the migration allocation against the source node is deleted.
Nested resource providers: Add the ability to model a tree of resource providers within the Placement service such that more complicated resource relationships can be tracked and used during scheduling, such as a compute node resource provider with child physical functions which in turn have child virtual functions for modeling SR-IOV capabilities.
This is a two-part effort.
Use the Cinder volume attachments API: Introduced with the block storage 3.27 API microversion, Cinder can accurately track multiple attachments for a single volume, including storing the storage backend
connection_infoand compute host
connector, which historically is stored in the Nova
block_device_mappingsdatabase table. This is an internal plumbing change to Nova and will be transparent to end users of the compute API. This will improve the separation of duties between the compute and block storage services, reduce technical debt in the compute service long-term, and build a foundation on which volume multi-attach support can be added.
Support multi-attachable volumes: Once Nova can support new-style volume attachments we can work on adding the changes to the API and at least the libvirt driver to attach a volume to more than one instance. There will be some changes needed to the block storage API for modeling volume storage backends that use shared targets, and Cinder will introduce policy rules so operators can configure if and how you can use multi-attach volumes, but basic support should be available, including boot from a multi-attach volume.