KEP-4049: Storage Capacity Scoring of Nodes for Dynamic Provisioning

Implementation History
BETA Implementable
Created 2023-05-30
Latest v1.37
Milestones
Alpha v1.33
Beta v1.37
Stable TBD
Related Enhancements
Ownership
Owning SIG
SIG Storage
Participating SIGs
SIG Scheduling
Primary Authors
@cupnes

KEP-4049: Storage Capacity Scoring of Nodes for Dynamic Provisioning

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

This KEP proposes adding a way to score nodes for dynamic provisioning of PVs. This scoring method is based on storage capacity in the VolumeBinding plugin. By considering the amount of free space that nodes have, it is possible to dynamically schedule pods on the node that has the most or least free space.

Motivation

Storage capacity needs to be considered when:

  • we want to resize after a node-local PV is scheduled. In this case we need to select a node with as much free space as possible.
  • we want to select a node with less free node space to reduce the number of nodes as much as possible.

Goals

  • To modify the scoring logic to count on dynamic provisioning in addition to the current, considering only static provisioning.

Non-Goals

  • To change how to score nodes for static provisioning.

Proposal

  • Node scores based on available space can be taken into account when performing dynamic provisioning.

Cluster admin can configure the scoring logic using a new field in VolumeBindingArgs of kubescheduler.config.k8s.io. The scoring logic is global for the whole cluster and we propose two values:

  • Prefer a node with the least allocatable.
  • Prefer a node with the maximum allocatable.

Considering the common scenario of local storage, we want to leave room for volume expansion after node allocation. The default setting is to prefer a node with the maximum allocatable.

User Stories (Optional)

Story 1 (Optional)

We want to leave room for volume expansion after node allocation. In this case, we want to allocate the node that has the maximum amount of free space.

Story 2 (Optional)

We want to reduce the number of nodes as much as possible to reduce costs when using a cloud environment. In this case, we want to allocate the node that has the smallest amount of sufficiently free space left.

Notes/Constraints/Caveats (Optional)

Risks and Mitigations

RiskImpactMitigation
Misconfiguration of storage capacity scoring parametersMediumProvide documentation
Potential performance overhead due to additional scoring calculationsLowOptimize scoring algorithms
Loss of optimized scheduling after downgrading to a version without this featureMediumExplain the impact of downgrading in documentation

Design Details

We modify the existing VolumeBinding plugin to achieve scoring of nodes for dynamic provisioning.

Modify stateData to be able to store StorageCapacity

We modify the struct called PodVolumes contained in stateData to score nodes for dynamic provisioning.

The struct of stateData is as follows:

type stateData struct {
	...
	// podVolumesByNode holds the pod's volume information found in the Filter
	// phase for each node
	// it's initialized in the PreFilter phase
	podVolumesByNode map[string]*PodVolumes
	...
}

By making the following changes to PodVolumes, CSIStorageCapacity can be stored.

+ type DynamicProvision struct {
+ 	PVC      *v1.PersistentVolumeClaim
+ 	Capacity *storagev1.CSIStorageCapacity
+ }

type PodVolumes struct {
	StaticBindings []*BindingInfo
-   DynamicProvisions []*v1.PersistentVolumeClaim
+ 	DynamicProvisions []*DynamicProvision
}

Get the capacity of nodes for dynamic provisioning

Add CSIStorageCapacity to the return value of the volumeBinder.hasEnoughCapacity method. This returns the DynamicProvision.Capacity field in the case of dynamic provisioning.

- func (b *volumeBinder) hasEnoughCapacity(provisioner string, claim *v1.PersistentVolumeClaim, storageClass *storagev1.StorageClass, node *v1.Node) (bool, error) {
+ func (b *volumeBinder) hasEnoughCapacity(provisioner string, claim *v1.PersistentVolumeClaim, storageClass *storagev1.StorageClass, node *v1.Node) (bool, *storagev1.CSIStorageCapacity, error) {
	quantity, ok := claim.Spec.Resources.Requests[v1.ResourceStorage]
	if !ok {
		// No capacity to check for.
- 		return true, nil
+ 		return true, nil, nil
	}

	// Only enabled for CSI drivers which opt into it.
	driver, err := b.csiDriverLister.Get(provisioner)
	if err != nil {
		if apierrors.IsNotFound(err) {
			// Either the provisioner is not a CSI driver or the driver does not
			// opt into storage capacity scheduling. Either way, skip
			// capacity checking.
- 			return true, nil
+ 			return true, nil, nil
		}
- 		return false, err
+ 		return false, nil, err
	}
	if driver.Spec.StorageCapacity == nil || !*driver.Spec.StorageCapacity {
- 		return true, nil
+ 		return true, nil, nil
	}

	// Look for a matching CSIStorageCapacity object(s).
	// TODO (for beta): benchmark this and potentially introduce some kind of lookup structure (https://github.com/kubernetes/enhancements/issues/1698#issuecomment-654356718).
	capacities, err := b.csiStorageCapacityLister.List(labels.Everything())
	if err != nil {
- 		return false, err
+ 		return false, nil, err
	}

  sizeInBytes := quantity.Value()
	for _, capacity := range capacities {
		if capacity.StorageClassName == storageClass.Name &&
			capacitySufficient(capacity, sizeInBytes) &&
			b.nodeHasAccess(node, capacity) {
			// Enough capacity found.
- 			return true, nil
+ 			return true, capacity, nil
		}
	}

	// TODO (?): this doesn't give any information about which pools where considered and why
	// they had to be rejected. Log that above? But that might be a lot of log output...
	klog.V(4).InfoS("Node has no accessible CSIStorageCapacity with enough capacity for PVC",
		"node", klog.KObj(node), "PVC", klog.KObj(claim), "size", sizeInBytes, "storageClass", klog.KObj(storageClass))
- 	return false, nil
+ 	return false, nil, nil
}

Scoring of nodes for dynamic provisioning

The Score method in the current VolumeBinding plug-in scores nodes considering only static provisioning. The scoring applies to every entry in podVolumes.StaticBindings.

In this KEP, add the scoring of nodes for dynamic provisioning in the Score method of the VolumeBinding plugin. The scoring applies to every entry in podVolumes.DynamicProvisions where Capacity is not equal to nil.

Scoring for dynamic provisioning is executed if there are no StaticBindings. In other words, if there is only static provisioning or both static and dynamic provisioning, the scoring will be done as usual for static provisioning. Then, if there is only dynamic provisioning, the following will be set to classResources and passed to the scorer function:

  • Requested: provision.PVC.Spec.Resources.Requests[v1.ResourceName(v1.ResourceStorage)]
  • Capacity: CSIStorageCapacity

By doing this, we can calculate scores to nodes for dynamic provisioning in a way that is based on the Shape setting of VolumeBindingArgs, and which takes into account the amount of free space the nodes have.

// Score invoked at the score extension point.
func (pl *VolumeBinding) Score(ctx context.Context, cs *framework.CycleState, pod *v1.Pod, nodeName string) (int64, *framework.Status) {
	if pl.scorer == nil {
		return 0, nil
	}
	state, err := getStateData(cs)
	if err != nil {
		return 0, framework.AsStatus(err)
	}
	podVolumes, ok := state.podVolumesByNode[nodeName]
        if !ok {
		return 0, nil
	}
-       // group by storage class
+
        classResources := make(classResourceMap)
-       for _, staticBinding := range podVolumes.StaticBindings {
-               class := staticBinding.StorageClassName()
-               storageResource := staticBinding.StorageResource()
-               if _, ok := classResources[class]; !ok {
-                       classResources[class] = &StorageResource{
-                               Requested: 0,
-                               Capacity:  0,
+       if len(podVolumes.StaticBindings) != 0 {
+               // group static biding volumes by storage class
+               for _, staticBinding := range podVolumes.StaticBindings {
+                       class := staticBinding.StorageClassName()
+                       storageResource := staticBinding.StorageResource()
+                       if _, ok := classResources[class]; !ok {
+                               classResources[class] = &StorageResource{
+                                       Requested: 0,
+                                       Capacity:  0,
+                               }
+                       }
+                       classResources[class].Requested += storageResource.Requested
+                       classResources[class].Capacity += storageResource.Capacity
+               }
+       } else {
+               // group dynamic biding volumes by storage class
+               for _, provision := range podVolumes.DynamicProvisions {
+                       if provision.Capacity == nil {
+                               continue
+                       }
+                       class := *provision.PVC.Spec.StorageClassName
+                       if _, ok := classResources[class]; !ok {
+                               classResources[class] = &StorageResource{
+                                       Requested: 0,
+                                       Capacity:  0,
+                               }
                        }
+                       requestedQty := provision.PVC.Spec.Resources.Requests[v1.ResourceName(v1.ResourceStorage)]
+                       classResources[class].Requested += requestedQty.Value()
+                       classResources[class].Capacity += provision.Capacity.Capacity.Value()
                }
-               classResources[class].Requested += storageResource.Requested
-               classResources[class].Capacity += storageResource.Capacity
        }
+
        return pl.scorer(classResources), nil
}

Users can select the scoring logic from the following options in VolumeBindingArgs. The scoring logic is the same among all Pod + PVC(s).

  • (a) Prefer a node with the least allocatable.
  • (b) Prefer a node with the maximum allocatable.

Considering the common scenario of local storage, we want to leave room for volume expansion after node allocation. The default setting is to prefer a node with the maximum allocatable.

Conditions for scoring static or dynamic provisioning

About the Score function, the score will be calculated with the existing way (only static provisioning is taken into account) if at least one PVC was statically provisioned. Otherwise, the score will be calculated from dynamic provisioning.

Implementation idea:

func (pl *VolumeBinding) Score(ctx context.Context, cs *framework.CycleState, pod *v1.Pod, nodeName string) (int64, *framework.Status) {
	...

+ 	if len(static) != 0 {
+ 		return static_score, nil;	// Same value as the current method
+ 	} else {
+ 		return dynamic_score, nil;	// Propose in this KEP
+ 	}
- 	return pl.scorer(classResources), nil
}

Feature Gate Consolidation

The StorageCapacityScoring feature gate will now control the functionality previously managed by the VolumeCapacityPriority feature gate, which will be deprecated. This consolidation focuses on enabling node scoring based on storage capacity, limited to the behaviors necessary for StorageCapacityScoring. Specifically, the utilization shape points have been supported because they are required for StorageCapacityScoring. However, the weight of storage class has not been implemented (ref1 , ref2 ), and there are no plans to require it for StorageCapacityScoring, so it will not be implemented. For more details on the original proposal, see KEP-1845 .

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

Nothing in particular.

Unit tests

The following unit tests have been written. Most are already merged into the master branch; a minor fix is under review in kubernetes/kubernetes#138497 .

Integration tests

The scoring function is tested in test/integration/volumescheduling/storage_capacity_scoring_test.go.

e2e tests

The following e2e tests are under review in kubernetes/kubernetes#138497 :

  • When only static provisioning is available, or a mixture of static provisioning and dynamic provisioning is available:
    • Does it pass traditional tests?
  • When only dynamic provisioning is available (single CSI driver case):
    • Is the Pod placed on the node with the largest available space by default?
    • When VolumeBindingArgs is set to “Prefer a node with the maximum allocatable”, is the Pod placed on the node with the largest available space?
    • Does the Pod placement fail if no node meets the requested size?
    • Even when the Pod is recreated, is the placement in the node performed as expected above?

Graduation Criteria

Alpha

  • Add StorageCapacityScoring feature gate
  • E2e tests completed

Beta

  • One release with positive feedback from users

GA

  • No users complaining about the new behavior

Upgrade / Downgrade Strategy

  1. Upgrading the cluster to support storage capacity scoring for dynamic provisioning:

    • After the upgrade, the scheduler will be able to score nodes based on their storage capacity for dynamic provisioning. This will involve additional checks and calculations to ensure that nodes with sufficient capacity are prioritized.
    • Existing configurations and API usage will remain compatible, but administrators may need to review and adjust their storage class configurations to fully leverage the new scoring mechanism.

  2. Downgrading the cluster to a version without storage capacity scoring for dynamic provisioning:

    • If the cluster is downgraded, the scheduler will revert to the previous behavior where storage capacity scoring for dynamic provisioning is not considered.
    • Any Pods created after the upgrade will still exist, but their scheduling will no longer take storage capacity into account, potentially leading to less optimal placement.
    • No additional changes to invocations or configurations are required, but administrators should be aware that the enhanced scheduling capabilities will be lost.

Version Skew Strategy

This enhancement is confined to the kube-scheduler component. A kube-scheduler at n-1 (without the StorageCapacityScoring feature gate) reverts to the current scoring behavior — scoring for VolumeBinding based on static provisioning only. This is acceptable and identical to the previous behavior. No changes are made to kubelet, kube-proxy, kube-controller-manager, or any node-level components (CSI, CRI, or CNI).

Production Readiness Review Questionnaire

Feature Enablement and Rollback

How can this feature be enabled / disabled in a live cluster?
  • Feature gate (also fill in values in kep.yaml)
    • Feature gate name: StorageCapacityScoring
    • Components depending on the feature gate: kube-scheduler
Does enabling the feature change any default behavior?

The scheduling behavior is changed if this function is enabled.

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

Yes, this feature can be disabled after it has been enabled by setting the feature gate to false again. In doing so, the scoring for VolumeBinding will revert to the current method. This change won’t affect the behavior of existing Pods.

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

Re-enabling the feature from a rolled-back state will result in scheduling that considers dynamic provisioning. There will be no impact on existing running Pods.

Are there any tests for feature enablement/disablement?

Yes. The unit tests in TestVolumeBinding cover both cases:

  • Feature gate enabled (fts: feature.Features{EnableStorageCapacityScoring: true}): the scoring tests listed in the Unit tests section above (e.g., TestVolumeBinding/storage_capacity_score ) verify that dynamic provisioning scoring based on CSIStorageCapacity works correctly.
  • Feature gate disabled (no fts field, defaulting to false): test cases such as TestVolumeBinding/unbound_claims_no_matches verify that when the feature gate is disabled, PreScore returns Skip for dynamic-provisioning-only scenarios, meaning behavior reverts to static provisioning scoring only.

Rollout, Upgrade and Rollback Planning

How can a rollout or rollback fail? Can it impact already running workloads?

Turning the feature gate flag on/off only changes scheduling scoring. So there is no possibility of impacting workloads that are already running.

What specific metrics should inform a rollback?

A spike on metric schedule_attempts_total{result="error|unschedulable"} when this feature gate is enabled.

Were upgrade and rollback tested? Was the upgrade->downgrade->upgrade path tested?

This was tested manually before the transition to beta using a cluster with the TopoLVM CSI driver. The test covered enabling StorageCapacityScoring (with the corresponding CSI driver configuration), disabling it (reverting the CSI driver configuration), and re-enabling it. In all three steps, Pod scheduling completed successfully with no unexpected behavior.

Is the rollout accompanied by any deprecations and/or removals of features, APIs, fields of API types, flags, etc.?

Yes, the VolumeCapacityPriority feature gate is deprecated in favor of the new StorageCapacityScoring feature gate.

Monitoring Requirements

How can an operator determine if the feature is in use by workloads?

If enabled, this feature applies to all workloads which uses delay binding PVCs. Also non-zero value of metric plugin_execution_duration_seconds{plugin="VolumeBinding",extension_point="Score"} is a sign indicating this feature is in use. Unfortunately, there is no way to distinguish whether only static provisioning is being considered (the current behavior) or both static and dynamic provisioning are being considered (the new behavior).

How can someone using this feature know that it is working for their instance?

By default, pods that use only dynamically provisioned PVCs will be scheduled to nodes with the most free space. This behavior can be configured to prefer nodes with the least free space instead.

  • Other (treat as last resort)
    • Details: Check which node the pod was scheduled to (kubectl get pod <pod> -o wide) and verify it is the node with the most available storage capacity by running kubectl get csistoragecapacities -A. By default, the VolumeBinding plugin prefers the node with the most available capacity for dynamic provisioning. This can be configured to prefer the node with the least available capacity instead via the Shape setting in VolumeBindingArgs.
What are the reasonable SLOs (Service Level Objectives) for the enhancement?

Metric plugin_execution_duration_seconds{plugin="VolumeBinding",extension_point="Score"} <= 100ms on 90-percentile.

(This follows the same SLO established in KEP-1845, as the additional overhead introduced by this KEP is limited to arithmetic calculations for scoring.)

What are the SLIs (Service Level Indicators) an operator can use to determine the health of the service?
  • Metrics
    • Metric name: plugin_execution_duration_seconds{plugin="VolumeBinding",extension_point="Score"}
    • [Optional] Aggregation method:
    • Components exposing the metric: kube-scheduler
Are there any missing metrics that would be useful to have to improve observability of this feature?

Nothing in particular.

Dependencies

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

Yes. This feature depends on a CSI driver that opts into storage capacity tracking by setting StorageCapacity: true in its CSIDriver spec. If no such driver is deployed, the feature has no effect on scheduling.

  • CSI driver (with StorageCapacity: true in its CSIDriver spec)
    • Usage description: The scheduler reads CSIStorageCapacity objects published by the CSI driver’s external-provisioner sidecar to determine available storage capacity on each node, which is used to score nodes for dynamic provisioning.
      • Impact of its outage on the feature: CSIStorageCapacity objects are no longer updated. Scoring continues based on stale capacity data and may place pods on suboptimal nodes.
      • Impact of its degraded performance or high-error rates on the feature: Stale or infrequently updated CSIStorageCapacity data causes scoring to reflect outdated capacity values. Pods may be placed on suboptimal nodes, reducing the effectiveness of the feature. Scoring returns to accurate behavior once the driver resumes updating the capacity objects.

Scalability

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

No.

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

No.

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?

Yes, it may affect the time taken by scheduling, but the impact is negligible.

The additional overhead introduced by this KEP is limited to arithmetic calculations in the Score phase: iterating over DynamicProvisions per node and summing the requested and available capacity values. This is equivalent in complexity to the existing static provisioning scoring. Importantly, the CSIStorageCapacity objects are already fetched and cached during the Filter phase, so no additional API calls or I/O occur during scoring. The SLO for plugin_execution_duration_seconds{plugin="VolumeBinding",extension_point="Score"} is therefore the same as established by KEP-1845 (<= 100ms on 90-percentile).

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

No.

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

No, this feature will not exhaust node resources such as PIDs, sockets, or inodes.

Troubleshooting

How does this feature react if the API server and/or etcd is unavailable?

The behavior in such cases does not change. This proposal only modifies one of the plugins in the kube-scheduler.

What are other known failure modes?

  • [Pods are placed on suboptimal nodes due to stale CSIStorageCapacity data]

    • Detection: Pods end up on nodes that do not match the intended scoring strategy configured via the Shape setting of VolumeBindingArgs — for example, placed on a node with less remaining capacity than expected when preferring the maximum allocatable. PVC provisioning may also fail after scheduling if the actual capacity on the chosen node is insufficient. Run kubectl get csistoragecapacities -A and inspect the creationTimestamp to check whether the data is fresh.
    • Mitigations: CSIStorageCapacity objects are created and updated by the CSI driver’s external-provisioner sidecar, not by the scheduler. If the objects do not reflect current node storage state, the root cause lies in the external-provisioner. Delete the stale CSIStorageCapacity objects so that the external-provisioner recreates them, or restart the external-provisioner Pod to trigger a forced resync. As an immediate workaround, disable the StorageCapacityScoring feature gate; scoring for VolumeBinding will revert to the previous method and will not be affected by stale capacity information.
    • Diagnostics: Check kube-scheduler logs at verbosity level 5 or higher for messages related to CSIStorageCapacity lookups. Verify that CSIStorageCapacity objects (kubectl get csistoragecapacities -A) reflect current node storage state. If objects are stale, also check the external-provisioner logs for sync errors or update failures.
    • Testing: Unit and integration tests verify that scoring correctly reflects the capacity values reported in CSIStorageCapacity objects. Note that detecting staleness is outside the scheduler’s responsibility; stale capacity data is not covered by tests.

  • [Fallback to scoring only with static provisioning when both static and dynamic provisioning exist]

    • Detection: When both static and dynamic provisioning are involved, the scoring is done only with static provisioning. This is by design but may be unexpected for users who expect storage capacity for dynamic provisioning to always be scored.
    • Mitigations: This is expected behavior documented in the KEP. Users who need scoring for dynamic provisioning should use only dynamically provisioned PVCs for a given pod.
    • Diagnostics: Inspect the pod’s PVC list. If any PVC has already been bound to an existing PV (kubectl get pvc), that pod will use scoring only with static provisioning.
    • Testing: Unit tests verify that static provisioning takes precedence when both exist.
What steps should be taken if SLOs are not being met to determine the problem?

Check the kube-scheduler logs.

Implementation History

  • v1.33: Alpha release
  • v1.37: Beta release

Drawbacks

  • The implementation of storage capacity scoring for dynamic provisioning may introduce complexity in the scheduling process. This could potentially lead to increased scheduling latency as the scheduler performs additional checks and calculations.

Alternatives

Weighting Static Provisioning Scores and Dynamic Provisioning Scores

The scoring function will return the sum of the static score and the dynamic score, each multiplied by their respective weights. The weights are determined by the ratio of static and dynamic capacities.

Implementation idea for the Score function:

func (pl *VolumeBinding) Score(...) (int64, *framework.Status) {
  ...
  return (static_weight) * static_score + (1-static_weight) * dynamic_score;
}

Ultimately, the current design was chosen. The reasons are as follows:

  • Conflict issue: In this approach, there is a possibility that the static provisioning and dynamic provisioning scores could cancel each other out, leading to inaccurate scoring.
  • Feasibility of implementation: The current design was deemed more feasible and clearer in terms of implementation.

Infrastructure Needed (Optional)