Skip to content

Latest commit

 

History

History
190 lines (146 loc) · 10.4 KB

File metadata and controls

190 lines (146 loc) · 10.4 KB

AEP-4016: Support for in place updates in VPA

Summary

VPA applies its recommendations with a mutating webhook, during pod creation. It can also evict pods expecting that it will apply the recommendation when the pod is recreated. This is a disruptive process so VPA has some mechanism to avoid too frequent disruptive updates.

This proposal allows VPA to apply its recommendations more frequently, with less disruption by using the in-place update feature which is an alpha feature available in Kubernetes 1.27. This proposal enables only core uses of in place updates in VPA with intention of gathering more feedback. Any more advanced uses of in place updates in VPA (like applying different recommendations during pod initialization) will be introduced as separate enhancement proposals.

Goals

  • Allow VPA to actuate without disruption,
  • Allow VPA to actuate more frequently, when it can actuate without disruption,
  • Allow VPA to actuate in situations where actuation by eviction doesn't work.

Non-Goals

  • Improved handling of injected sidecars
    • Separate AEP will improve VPAs handling of injected sidecars.

Proposal

Add new supported values of UpdateMode:

  • InPlaceOnly and
  • InPlaceOrRecreate.

Context

In-place update of pod resources KEP is available in alpha in Kubernetes 1.27. The feature allows changing container resources while the container is running. It also adds ResizePolicy field to Container. This field indicates for an individual resource if a container needs to be restarted by kubelet when the resource is changed. For example it may be the case that a Container automatically adapts to a change in CPU, but needs to be restarted for a change in Memory to take effect.

Design Details

In the update modes existing before (Initial and Recreate) only admission controller was changing pod spec (updater was responsible for evicting pods so that admission controller could change them during admission).

In the newly added InPlaceOnly and InPlaceOrRecreate modes VPA Updater will execute in-place updates (which change pod spec without involving admission controller).

Applying Updates During Pod Admission

For VPAs in InPlaceOnly and InPlaceOrRecreate modes VPA Admission Controller will apply updates to starting pods, like it does for VPAs in Initial, Auto, and Recreate modes.

Applying Disruption-free Updates

When an update only changes resources for which a container indicates that it doesn't require a restart, then VPA can attempt to actuate the change without disrupting the pod. VPA Updater will:

  • attempt to actuate such updates in place,
  • attempt to actuate them if difference between recommendation and request is at least 10%
    • even if pod has been running for less than 12h,
  • if necessary execute only partial updates (if some changes would require a restart).

Applying Disruptive Updates

In both InPlaceOnly and InPlaceOrRecreate modes VPA updater will attempt to apply updates that require container restart in place. It will update them under conditions that would trigger update with Recreate mode. That is it will apply disruptive updates if:

  • Any container has a request below the corresponding LowerBound or
  • Any container has a request above the corresponding UpperBound or
  • Difference between sum of pods requests and sum of recommendation Targets is more than 10% and the pod has been running undisrupted for at least 12h.
    • Successful disruptive update counts as disruption here (and prevents further disruptive updates to the pod for 12h).

In InPlaceOrRecreate mode (but not in InPlaceOnly mode) VPA updater will evict pod to actuate a recommendation if it attempted to apply the recommendation in place and failed.

VPA updater will consider that the update failed if:

  • The pod has .status.resize: Infeasible or
  • The pod has .status.resize: Deferred and more than 1 minute elapsed since the update or
  • The pod has .status.resize: InProgress and more than 1 hour elapsed since the update:
    • There seems to be a bug where containers that say they need to be restarted get stuck in update, hopefully it gets fixed and we don't have to worry about this by beta.
  • Patch attempt will return an error.
    • If the attempt fails because it would change pods QoS:
      • InPlaceOrRecreate will treat it as any other failure and consider evicting the pod.
      • InPlaceOnly will consider applying request slightly lower than the limit.

Those failure modes shouldn't disrupt pod operations, only update. If there are problems that can disrupt pod operation we should consider not implementing the InPlaceOnly mode.

Comparison of UpdateModes

VPA updater considers the following conditions when deciding if it should apply an update:

  • CanEvict:
    • Pod is Pending or
    • There are enough running pods in the controller.
  • Quick OOM:
    • Any container in the pod had a quick OOM (by default less than 10 minutes after the container started) and
    • There is difference between sum of recommendations and sum of current requests over containers (see [defaultPriorityProcessor.GetUpdatePriority]).
  • Long-lived pod - started enough time ago (by default 12h)
  • Significant change - difference between sum of requests over containers and sum of recommendations are different enough (by default 10%).
  • Outside recommended range:
    • At least one container has at least one resource request lower than the lower bound of the corresponding recommendation or
    • At least one container has at least one resource request higher than the upper bound of the corresponding recommendation.
  • Disruption-free update - doesn't change any resources for which the relevant container specifies RestartPolicy: RestartContainer.

Auto / Recreate evicts pod if:

  • CanEvict returns true for the pod, and it meets at least one of the following conditions:
    • Quick OOM,
    • Outside recommended range,
    • Long-lived pod with significant change.

InPlaceOnly and InPlaceOrRecreate will attempt to apply a disruption-free update in place if it meets at least one of the following conditions:

  • Quick OOM,
  • Outside recommended range,
  • Significant change.

InPlaceOnly and InPlaceOrRecreate when considering a disruption-free update in place ignore some conditions that influence eviction decission in the Recreate mode:

  • CanEvict won't be checked and
  • Pods with significant change can be updated even if they are not long-lived.

InPlaceOnly and InPlaceOrRecreate will attempt to apply updates that are not disruption-free in place under the same conditions that apply to updates in the Recreate mode.

InPlaceOrRecreate will attempt to apply updates that by eviction when:

  • VPA already attempted to apply the update in-place and failed and
  • it meets conditions for applying in the Recreate mode.

Test Plan

The following test scenarios will be added to e2e tests. Both InPlaceOnly and InPlaceOrRecreate modes will be tested and they should behave the same:

  • Admission controller applies recommendation to pod controlled by VPA.
  • Disruption-free in-place update applied to all containers of a pod (request in recommendation bounds).
  • Partial disruption-free update applied to some containers of a pod, some disruptive changes skipped (request in recommendation bounds).
  • Disruptive in-place update applied to all containers of a pod (request out ouf recommendation bounds).

There will be also scenarios testing differences between InPlaceOnly and InPlaceOrRecreate modesL

  • Disruptive in-place update will fail. In InPlaceOnly pod should not be evicted, in the InPlaceOrRecreate pod should be evicted and the recommendation applied.
  • VPA attempts an update that would change Pods QoS (RequestsOnly scaling, request initially < limit, recommendation equal to limit). In InPlaceOnly pod should not be evicted, request slightly lower than the recommendation will be applied. In the InPlaceOrRecreate pod should be evicted and the recommendation applied.

Implementation History

  • 2023-05-10: initial version