Support traits in the Ironic driver¶
To allow more granular scheduling of Ironic resources, Ironic is exposing traits for each ironic node, which in turn must be reported up to placement via the ironic driver.
Sometimes a flavor needs to select only a subset of the ironic nodes with a given resource class, rather than just any node of a given resource class.
Eventually, it is expected that ironic may re-configure a node based on the requested trait, and a node reporting a trait may mean that such re-configuration is available for that node, be that BIOS configuration or RAID configuration changes. For the moment there is no agreed way forward for this approach, so this part of the problem will be considered out of scope.
Consider flavor Gold targeting resource class CUSTOM_GOLD. Some users want a kind of Gold++ where it also targets CUSTOM_GOLD but in addition requires specific set of CPU flags that are not available on all nodes with the resource class of CUSTOM_GOLD.
Another use case is being able to dedicate specific Ironic nodes for use only by a specific set of projects. The remainder of the hosts are for general use. If a user has a dedicated pool of resources, they have the ability to pick if they create an instance in their dedicated pool or in the general pool. Other users are only able to build in the general pool. One way to bisect the nodes like this is assigning traits such as CUSTOM_IRONIC_NODE_PROJECT_B and CUSTOM_IRONIC_NODE_GENERAL_USE to the appropriate ironic nodes. Then there is a public flavor to target the general pool of hosts, and a private project specific flavor that targets their dedicated pool.
All this work depends on the ability for a flavor to have a required set of traits and for Nova to be able to request from placement resource providers that have the requested set of traits. This is all implemented in the two blueprints:
There are two main parts to this spec:
sending requested traits back to ironic
getting ironic node traits into placement
When Nova boot is called, the ironic driver already sets capabilities related extra specs from the requested flavor on the ironic node via the Ironic API. This is done in the virt drivers _add_instance_info_to_node method. It is set on the path /instance_info/capabilities. In a similar way we will also set the traits related flavor extra specs on a path /instance_info/traits. Note, this requires no API changes on the Ironic side, the path is a PATCH requests JSONPath change identifier.
Currently the nova virt driver has a get_inventory call to list the inventory of a given compute_node. This change will add a get_traits(nodename) call to the virt driver interface to fetch the traits for a given nodename. In a similar way to get_inventory, this will use the cached node details. For drivers that don’t override the new get_traits call we will raise a NotImplementedError.
These traits for each ironic node need to be checked against the current traits for the associated Resource Provider, and updated if needed. This is likely going to be done in the scheduler report client, in a similar way to the existing set_inventory_for_provider method that currently creates the resource provider and updates its inventory for the appropriate node. Internally the pattern in _ensure_resource_provider method that ensures the resource provider is in the correct state will be used to ensure the traits are updated correctly. The existing resource provider APIs will be used to update the traits.
We are considering the Ironic API as the single source of truth for the Traits for a given Node. So should someone set any traits directly on the Placement API, they will be overridden on the next virt driver sync, with will reset the traits to what is in the Ironic API.
It is hoped the virt driver will move away from get_inventory() and towards update_provider_tree(). While that would change details on the implementation the key data flow is the same. This spec is rather urgent because the move to Resource Classes makes placement more rigid, and stops the ability to build on a large ironic host with a smaller flavor. Sadly some people rely on that feature.
We could allow admins to set traits directly via the placement API, but this is a bit strange when Nova is creating the resource provider. Its possible ironic could create the resource provider on Nova’s behalf. An additional complication is that we really want the traits to be populated by ironic inspector, in a similar way capabilities are set by inspector rules today.
We could allow admins to extend the list of traits with a configuration variable on the compute host, and that could be behaviour for all drivers that don’t implement get_traits, but for the moment this has been ignored because it is not relevant to the Ironic driver.
Data model impact¶
REST API impact¶
None, uses exiting placement APIs.
Other end user impact¶
There will be increased API calls to placement when the inventory is updated.
Other deployer impact¶
A deployer will now be able to set traits in Ironic via the dependent ironic spec: https://review.openstack.org/#/c/504531/
This spec is about the Nova virt driver sending these traits to placement.
- Primary assignee:
John Garbutt (johnthetubaguy)
add get_traits to the ironic driver
push traits to the placement API
on boot set requested traits on the ironic node
New APIs in Ironic to set traits on a node: https://review.openstack.org/#/c/504531/
There is a loose dependency on Nova adding support for request traits on bp add-trait-support-in-allocation-candidates and bp request-traits-in-nova Without those two blueprints this feature can’t be tested end to end.
Need functional tests that prove we can select the correct ironic node on traits alone, by correctly configuring a flavor.
Related details are mostly covered in the Ironic docs around using resource classes and Nova flavors. This should be expanded to detail how traits can also be used.