Skip to content

Latest commit

 

History

History
159 lines (129 loc) · 5.71 KB

replica-sets.md

File metadata and controls

159 lines (129 loc) · 5.71 KB

VirtualMachineInstanceReplicaSet

Overview

In order to allow scaling up or down similar VMIs, add a way to specify VirtualMachineInstance templates and the amount of replicas required, to let the runtime create these VirtualMachineInstances.

Use-cases

Ephemeral VMIs

Scaling ephemeral VMIs which only have read-only mounts, or work with a backing store, to keep temporary data, which can be deleted after the VMI gets undefined.

Implementation

A new object VirtualMachineInstanceReplicaSet, backed with a controller will be implemented:

apiVersion: kubevirt.io/v1alpha2
kind: VirtualMachineInstanceReplicaSet
metadata:
  name: myreplicaset
spec:
  replicas: 3
  selector:
    matchLabels:
      mylabel: mylabel
  template:
    metadata:
      name: test
      labels:
        mylabel: mylabel
    spec:
      domain:
        devices:
      [...]

spec.template is equal to a VirtualMachineInstanceSpec. spec.replicas specifies how many instances should be created out of spec.template. spec.selector contains selectors, which need to match spec.template.metadata.labels.

The status looks like this:

status:
  conditions: null
  replicas: 3
  readyReplicas : 2

In case of a scaling error, a ReplicaFailure condition is added to the status.conditions. Further it shows the number of VirtualMachineInstances which are in a non-final state and which match spec.selector in the status.replicas field. status.readyReplicas indicates how many of these replicas meet the ready condition.

Note that at the moment when writing this proposal, there exist no readiness checks for VirtualMachineInstances in Kubevirt. Therefore a VirtualMachineInstance is considered to be ready, when reported by virt-handler as running or migrating.

In case of a delete failure:

status:
  conditions:
  - type: "ReplicaFailure"
    status: True
    reason: "FailureDelete"
    message: "no permission to delete VMIs"
    lastTransmissionTime: "..."
  replicas: 4
  readyReplicas: 3

In case of a create failure:

status:
  conditions:
  - type: "ReplicaFailure"
    status: True
    reason: "FailureCreate"
    message: "no permission to create VMIs"
    lastTransmissionTime: "..."
  replicas: 2
  readyReplicas: 3

Guarantees

The VirtualMachineInstanceReplicaSet does not guarantee that there will never be more than the wanted replicas active in the cluster. Based on readiness checks, unknown VirtualMachineInstance states and graceful deletes, it might decide to already create new replicas in advance, to make sure that the amount of ready replicas stays close to the expected replica count.

Chosen Implementation

The implementation of the VirtualMachineInstanceReplicaSet is a reimplementation of the ReplicaSet for Pods in Kubernetes. It does not wrap around a Kubernetes ReplicaSet.

There two hard reasons why one was chose over the other:

  1. Allow possible VirtualMachineInstance related optimizations by implementing a custom deletion order. When scaling down, it can make sense to take migrating VirtualMachineInstances into account. A possible deletion order would be: not ready VMIs, migrating VMIs, ready VMIs. This would be impossible to represent by wrapping around the ReplicaSet.
  2. Make Live Migrations an orthogonal feature for every VirtualMachineInstance controller. If a controller does not have to care about the VirtualMachineInstance/Pod relationship, it does not have to care in any special way about migrations. By wrapping around existing controllers, we will always have to hide migration target Pods from the controller, until the VirtualMachineInstance was migrated, and then have to make it somehow visible again. This solution leads to tricky synchronization problems which may not even be completely solvable.

There are several soft reasons why one was chosen over the other:

  1. Provide a clear VirtualMachineInstance abstraction via our VirtualMachineInstance object. This allows domain specific modelling of controllers.
  2. Hide the whole relationship between VirtualMachineInstances and Pods behind on controller, so that others don't have to care about it.
  3. The actual business-logic of scaling the VirtualMachineInstances is very easy to understand and simple to implement. We get almost all the necessary infrastructure code from k8s libraries. The infrastructure code has a similar complexity (entity listeners, creating/deleting resources, ... ) for both implementations. For different types of controllers this argument may not apply (e.g. DaemonSet equivalent for VirtualMachineInstances).
  4. Because of the simplicity of the business-logic, the equal complexity regarding to the controller infrastructure and the fact that there exists a ReplicaSet reference implementation from which we can learn/copy, it seems beneficial to really express our domain and avoid design errors, by implementing another complex flow around the existing ReplicaSet.

Milestones

  • Basic functionality
  • Support label changes
  • Define a well known scale-down order
  • Support graceful delete [1]
  • Support controller references [2]
  • Support adopting orphaned Pods [2]

The basic functionality includes scaling up, down and reporting errors if scaling does not work. In this stage it is the full responsibility of the user to attach labels to the VMIs in a way, so that selectors of multiple VirtualMachineInstanceReplicaSets don't overlap.

References

  1. https://kubernetes.io/docs/tasks/run-application/force-delete-stateful-set-pod/#delete-pods
  2. https://github.com/kubernetes/community/blob/58b1c30d95719749068497ba35dfe4c64b21aa72/contributors/design-proposals/api-machinery/controller-ref.md