KEP-5027: DRA: admin-controlled device attributes

Implementation History
ALPHA Provisional
Created 2024-01-10
Latest v1.33
Related Enhancements
Ownership
Owning SIG
SIG Scheduling
Primary Authors
@pohly

KEP-5027 : DRA: admin-controlled device attributes

Release Signoff Checklist

Items marked with (R) are required prior to targeting to a milestone / release.

  • (R) Enhancement issue in release milestone, which links to KEP dir in kubernetes/enhancements (not the initial KEP PR)
  • (R) KEP approvers have approved the KEP status as implementable
  • (R) Design details are appropriately documented
  • (R) Test plan is in place, giving consideration to SIG Architecture and SIG Testing input (including test refactors)
    • e2e Tests for all Beta API Operations (endpoints)
    • (R) Ensure GA e2e tests meet requirements for Conformance Tests
    • (R) Minimum Two Week Window for GA e2e tests to prove flake free
  • (R) Graduation criteria is in place
  • (R) Production readiness review completed
  • (R) Production readiness review approved
  • “Implementation History” section is up-to-date for milestone
  • User-facing documentation has been created in kubernetes/website , for publication to kubernetes.io
  • Supporting documentation—e.g., additional design documents, links to mailing list discussions/SIG meetings, relevant PRs/issues, release notes

Summary

With Dynamic Resource Allocation (DRA), DRA drivers publish information about the devices that they manage in ResourceSlices. This information is used by the scheduler when selecting devices for user requests in ResourceClaims.

This KEP adds a Kubernetes API that privileged users, typically cluster administrators or control plane controllers, can use to override or extend that information. This can be permanent as part of the installation of a DRA driver to adapt the driver to the cluster or temporary as part of cluster maintenance. An extension of the API adds taints .

Motivation

Goals

  • Enable admin-controlled device taints.

  • Enable updating how devices are seen in the cluster without having to use driver-specific APIs which influence what a driver puts into ResourceSlices.

Non-Goals

  • At least for alpha: extend kubectl to provide a unified view of devices together with all patches that apply to them.

Proposal

The intent to patch device attributes must be recorded persistently so that it is preserved even when a ResourceSlice gets removed or updated. To achieve this, a new cluster-scoped ResourceSlicePatch type gets added. A single ResourceSlicePatch object specifies device attributes that apply to all devices matching a CEL expression (i.e. the same way as users select devices in a ResourceClaim) and/or some additional criteria (device class, driver/pool/device name).

The scheduler must merge these additional attributes with the ones provided by the DRA drivers on-the-fly while it gathers information about available devices.

Notes/Constraints/Caveats

Users who look at ResourceSlices to figure out which devices are available also need to consider ResourceSlicePatches to get the full picture. Copying from the ResourceSlicePatch spec into the ResourceSlice status could help here, but would not be instantaneous and potentially cause write amplification (one ResourceSlicePatch affecting many different devices) and therefore is not part of this proposal.

Perhaps kubectl describe resourceslices can be extended to include the additional information. For now this is out of scope.

Risks and Mitigations

Creating a ResourceSlicePatch is racing with on-going scheduling attempts. This is unavoidable. Removing a device from a ResourceSlice has the same problem: updates need to reach the scheduler before it can consider them. Evaluating a patch on the client-side instead of [having a controller update slices]((#storing-result-of-patching-in-resourceslice) mitigates this risk by shortening the time window where updates must be sent to the scheduler.

From a security perspective, permission to patch device attributes is expected to be limited to privileged users who already have the ability to add or remove DRA drivers, so there won’t be a substantial difference.

Performance in the scheduler could be an issue. This will be mitigated by caching the patched devices and (re-)applying patches only when they or the device definitions change, which should be rare.

Patching directly in the informer event handlers may be fast enough. If it turns out to slow down those handlers too much, then a workqueue with workers may be needed to decouple updating the cache from the events which trigger updating and to avoid slowing down the informers.

The scheduler’s “slice changed” cluster events must be driven by that cache, not the original informers, otherwise a ResourceSlice or ResourceSlicePatch change could trigger a pod scheduling attempt before the slice cache is up-to-date again.

Design Details

API

The ResourceSlicePatch is a cluster-scoped type in the resource.k8s.io API group, initially in v1alpha3 (the alpha version in Kubernetes 1.32). Because it may be useful to clean up after disabling the feature and because the device taint feature also uses this type, it gets served unconditionally as long as the v1alpha3 version is enabled. Fields related specifically to this KEP are feature-gated.

type ResourceSlicePatch struct {
    metav1.TypeMeta
    // Standard object metadata
    // +optional
    metav1.ObjectMeta

    // Changing the spec automatically increments the metadata.generation number.
    Spec ResourceSlicePatchSpec
}

type ResourceSlicePatchSpec struct {
    // Devices defines how to patch device attributes and taints.
    Devices DevicePatch
}

// DevicePatch selects one or more devices by class, driver, pool, device names
// and/or CEL selectors. All of these criteria must be satisfied by a device, otherwise
// it is ignored by the patch. A DevicePatch with no selection criteria is
// valid and matches all devices.
type DevicePatch struct {
    // Filter defines which device(s) the patch is applied to.
    //
    // +optional
    Filter *DevicePatchFilter

    // If a ResourceSlice and a DevicePatch define the same attribute or
    // capacity, the value of the DevicePatch is used. If multiple
    // different DevicePatches match the same device, then the one with
    // the highest priority wins. If priorities are equal, the older
    // patch wins, where "older" is determined based on the creation time.
    // This ensures that adding a new patch does not
    // accidentally change the effect of some existing patch unless
    // that is clearly intended according to the priority. Updates
    // do not change the creation time, so it could still happen that
    // a more recent change is preferred because it happens to be in
    // an older DevicePatch. Overall it is better to set the
    // priority to different values to avoid such ambiguities.
    //
    // +optional
    Priority *int

    // Attributes defines the set of attributes to patch for matching devices.
    // The name of each attribute must be unique in that set and
    // include the domain prefix.
    //
    // In contrast to attributes in a ResourceSlice, entries here are allowed to
    // be marked as empty by setting their null field. Such entries remove the
    // corresponding attribute in a ResourceSlice, if there is one, instead of
    // overriding it. Because entries get removed and are not allowed in
    // slices, CEL expressions do not need to deal with null values.
    //
    // The maximum number of attributes and capacities in the DevicePatch combined is 32.
    // This is an alpha field and requires enabling the DRAAdminControlledDeviceAttributes
    // feature gate.
    //
    // +optional
    // +featureGate:DRAAdminControlledDeviceAttributes
    Attributes map[FullyQualifiedName]NullableDeviceAttribute

    // ^^^
    // The size limit is the same as for attributes and capacities in a ResourceSlice.
    // We could make it larger here because we are less constrained by overall object
    // size, but it seems unnecessary.

    // Capacity defines the set of capacities to patch for matching devices.
    // The name of each capacity must be unique in that set and
    // include the domain prefix.
    //
    // Removing a capacity is not supported. It can be reduced to 0 instead.
    //
    // The maximum number of attributes and capacities in the DevicePatch combined is 32.
    // This is an alpha field and requires enabling the DRAAdminControlledDeviceAttributes
    // feature gate.
    //
    // +optional
    // +featureGate:DRAAdminControlledDeviceAttributes
    Capacity map[FullyQualifiedName]DeviceCapacity

    // ^^^^
    // The assumption here is that all device types will have attributes and capacities,
    // similar to the current BasicDevice type. Therefore the overrides are not made
    // specific to certain device types.
}

// DevicePatchFilter defines which device(s) a DevicePatch applies to.
// All criteria defined here must be satisfied for a device to be
// patched.
type DevicePatchFilter struct {
    // If DeviceClassName is set, the selectors defined there must be
    // satisfied by a device to be patched. This field corresponds
    // to class.metadata.name.
    //
    // +optional
    DeviceClassName *string

    // If driver is set, only devices from that driver are patched.
    // This fields corresponds to slice.spec.driver.
    //
    // +optional
    Driver *string

    // If pool is set, only devices in that pool are patched.
    // This fields corresponds to slice.spec.pool.name.
    //
    // Setting also the driver name may be useful to avoid
    // ambiguity when different drivers use the same pool name,
    // but this is not required because selecting pools from
    // different drivers may also be useful, for example when
    // drivers with node-local devices use the node name as
    // their pool name.
    //
    // +optional
    Pool *string

    // If device is set, only devices with that name are patched.
    // This field corresponds to slice.spec.devices[].name.
    //
    // Setting also driver and pool may be required to avoid ambiguity,
    // but is not required.
    //
    // +optional
    Device *string

    // Selectors define criteria which must be satisfied by a
    // device to be patched. All selectors must be satisfied.
    //
    // +optional
    // +listType=atomic
    Selectors []DeviceSelector

    // ^^^
    //
    // Selectors is a list for the same reason why a request has a list: at some point
    // we might have entries which use some yet to be defined mechanism which isn't
    // CEL.
}

To distinguish intentionally empty attributes from attributes which have some future, unknown content and thus only seem to be empty to an older client, a special “null value” gets introduced:

// NullableDeviceAttribute must have exactly one field set.
// It has the exact same fields as a DeviceAttribute plus `null` as
// an additional alternative.
type NullableDeviceAttribute struct {
    DeviceAttribute

    // NullValue, if set, marks an intentionally empty attribute.
    //
    // +optional
    // +oneOf=ValueType
    NullValue *NullValue `json:"null,omitempty" ...`
}

// ^^^
// `NullableDeviceAttribute` as an extension ensures that the OpenAPI
// for ResourceSlice remains unchanged. Using the same type with
// a `NullValue` that can be set only in one type is less clear.

type NullValue struct {}

Merging ResourceSlicePatches and ResourceSlices

Helper code which keeps an up-to-date list of devices with all patches added to them will be provided as part of k8s.io/dynamic-resource-allocation. It will be based on informers such that evaluating the filter only is necessary when ResourceSlices or ResourceSlicePatches change.

A CEL expression that fails to evaluate to a boolean for a device (runtime error like looking up an attribute that isn’t defined, wrong result type, etc.) is considered faulty. The patch then does not apply to the device where it failed and an event will be generated for the ResourceSlicePatch with the faulty CEL expression.

Test Plan

[X] I/we understand the owners of the involved components may require updates to existing tests to make this code solid enough prior to committing the changes necessary to implement this enhancement.

Prerequisite testing updates

None.

Unit tests

v1.32.0:

  • k8s.io/dynamic-resource-allocation/structured: 91.3%
  • k8s.io/kubernetes/pkg/apis/resource/validation: 98.6%
Integration tests

Additional scenarios will be added to test/integration/scheduler_perf, not just for correctness but also to evaluate a potential performance impact.

e2e tests

One E2E test scenario is to change attributes and then run pods which select devices based on those modified attributes such that unmodified devices don’t match.

  • :

Graduation Criteria

Alpha

  • Feature implemented behind a feature flag
  • Initial e2e tests completed and enabled

Beta

  • Gather feedback from developers and surveys
  • Additional tests are in Testgrid and linked in KEP

GA

Upgrade / Downgrade Strategy

Patching devices gets disabled when downgrading to a release without support for it or when disabling the feature. The effect is that pods get scheduled as if the ResourceSlicePatches didn’t exist. Because they are completely stand-alone, there is no effect on ResourceSlices or ResourceClaims.

Version Skew Strategy

During version skew where the apiserver supports the feature and the scheduler doesn’t, users can create ResourceSlicePatches without encountering errors or warnings, but they won’t have any effect.

Production Readiness Review Questionnaire

Feature Enablement and Rollback

It is possible to disable the feature through the feature gate while leaving the API group enabled. This enables cleanup through the API.

Re-enabling is supported because ResourceSlicePatches remain in etcd even if they are inaccessible.

How can this feature be enabled / disabled in a live cluster?
  • Feature gate (also fill in values in kep.yaml)
    • Feature gate name: DRAAdminControlledDeviceAttributes
    • Components depending on the feature gate:
      • kube-apiserver
      • kube-scheduler
  • Other
    • Describe the mechanism: resource.k8s.io/v1alpha3 API group
    • Will enabling / disabling the feature require downtime of the control plane? Yes, in the apiserver.
    • Will enabling / disabling the feature require downtime or reprovisioning of a node? No.
Does enabling the feature change any default behavior?

No.

Can the feature be disabled once it has been enabled (i.e. can we roll back the enablement)?

Yes. The behavior of scheduling changes when it was in use. Running applications are not affected.

What happens if we reenable the feature if it was previously rolled back?

Admin-controlled attributes take effect again for scheduling. Running applications are not affected because allocations are never updated once they are made.

Note that this is different for reenabling device taints (KEP 5055): that can cause pod to get evicted.

Are there any tests for feature enablement/disablement?

This will be covered through unit tests for the apiserver and scheduler.

Rollout, Upgrade and Rollback Planning

How can a rollout or rollback fail? Can it impact already running workloads?
What specific metrics should inform a rollback?
Were upgrade and rollback tested? Was the upgrade->downgrade->upgrade path tested?
Is the rollout accompanied by any deprecations and/or removals of features, APIs, fields of API types, flags, etc.?

Monitoring Requirements

How can an operator determine if the feature is in use by workloads?
How can someone using this feature know that it is working for their instance?
  • Events
    • Event Reason:
  • API .status
    • Condition name:
    • Other field:
  • Other (treat as last resort)
    • Details:
What are the reasonable SLOs (Service Level Objectives) for the enhancement?
What are the SLIs (Service Level Indicators) an operator can use to determine the health of the service?
  • Metrics
    • Metric name:
    • [Optional] Aggregation method:
    • Components exposing the metric:
  • Other (treat as last resort)
    • Details:
Are there any missing metrics that would be useful to have to improve observability of this feature?

Dependencies

Does this feature depend on any specific services running in the cluster?

Scalability

Applying patches to devices scales with number of ResourceSlicePatches * number of devices when CEL selectors need to be evaluated. Without them, filtering scales with number of ResourceSlicePatches * number of ResourceSlices but then may still need to compare device names and of course modify selected devices.

Will enabling / using this feature result in any new API calls?

A fixed, small number of clients (primarily the scheduler) need to start watching ResourceSlicePatches.

Will enabling / using this feature result in introducing new API types?

ResourceSlicePatches must be created explicitly by admins or controller operated by admins. Kubernetes itself does not create them.

The number of ResourceSlicePatches is expected to be orders of magnitude smaller than the number of ResourceSlices.

Will enabling / using this feature result in any new calls to the cloud provider?

No.

Will enabling / using this feature result in increasing size or count of the existing API objects?

No.

Will enabling / using this feature result in increasing time taken by any operations covered by existing SLIs/SLOs?

Pod scheduling should be as fast as would be without this feature, because in both cases it starts with listing all devices. That information is local and comes either from an informer cache or a cache of patched devices.

Will enabling / using this feature result in non-negligible increase of resource usage (CPU, RAM, disk, IO, …) in any components?

Filtering and patching are local operations, with no impact on the cluster. To prevent doing the same work repeatedly, it will be implemented so that it gets done once and then only processes changes. This increases CPU and RAM consumption. But even if all devices should get patched (which is unlikely), memory will be shared between objects in the informer cache and in the patch cache, so it will not be doubled.

Can enabling / using this feature result in resource exhaustion of some node resources (PIDs, sockets, inodes, etc.)?

No, because the feature is not used on nodes.

Troubleshooting

How does this feature react if the API server and/or etcd is unavailable?
What are other known failure modes?
What steps should be taken if SLOs are not being met to determine the problem?

Implementation History

  • 1.33: first KEP revision, implementation postponed until there is a more specific need for it

Drawbacks

Distributing information across different objects of different types makes it harder for users to get a complete view.

Alternatives

Admin-intent in ResourceSlice

Instead of ResourceSlicePatch as a separate type, new fields in the ResourceSlice status could be modified by an admin. That has the problem that the ResourceSlice object might get deleted while doing cluster maintenance like a driver update, in which case the admin intent would get lost.

Storing result of patching in ResourceSlice

A controller could read ResourceSlicePatches and apply them to ResourceSlices. Then consumers like the scheduler and users would only need to look at ResourceSlices. This has several drawbacks.

We would need to duplicate the attributes in the slice status. If we didn’t and directly modified the spec, this patch controller and the CSI driver as the owner of the slice spec would fight against each other. Also, after removing a patch the original state must be available somewhere, otherwise the controller cannot restore it.

Duplicating the attributes might make a slice too large. The limits were chosen so that we have some space left for a status, but not enough for a status that is potentially as large as the spec.

Creating a single ResourceSlicePatch could force the controller to update a potentially large number of ResourceSlices. When using rate limiting, updating them all will take longer than client-side patching. When not using rate limiting, this could overwhelm the apiserver.