YuniKorn Concepts

Internal

• YuniKorn

YuniKorn Core

YuniKorn Core Concepts

Kuberentes Implementation

• A namespace can have a "queue" if annotated with "yunikorn.apache.org/queue". A namespace can have a "parent queue" is annotated with "yunikorn.apache.org/parentqueue".
• An allocation can be in one of two states ("Pending" and "In-Progress"). A pending allocation is one which has been decided upon by YuniKorn but has not yet been communicated to the default scheduler via PreFilter()/Filter(). Once PreFilter()/Filter() pass, the allocation transitions to "In-Progress" to signify that the default scheduler is responsible for fulfilling the allocation. Once PostBind() is called in the plugin to signify completion of the allocation, it is removed.

Allocation

Allocation Ask

Each allocation has a key. The key is used by both the scheduler and the resource manager to track allocations. The key does not have to be the resource manager's internal allocation ID, such as the pod name.

If the allocation specifies an application ID, the application must be registered in advance, otherwise we get "failed to find application ..."

Each allocation ask requests resources.

Each allocation ask has a priority.

What is "MaxAllocations"?

An application goes into a state transition after the first allocation ask, to "Running".

Identity

An application is submitted under a certain identity, that consists of a user and one or more groups.

TO PARSE:

• https://yunikorn.apache.org/docs/user_guide/usergroup_resolution/
• https://yunikorn.apache.org/docs/design/scheduler_configuration/#user-definition

Group

The identity an application is submitted under may be associated with one or more groups.

Resource Manager (RM)

YuniKorn communicates with various implementation of resource management systems (Kubernetes, YARN) via a standard interface defined in the yunikorn-scheduler-interface package.

Node

Can a node be declared to be part of a partition with the "si/node-partition" label? It seems that the node Attributes partially come from the node labels.

In the Kubernetes implementation, the node is first added, then updated.

As part of handling the RMNodeUpdateEvent, RMProxy calls callback.UdpateNode().

Context

yunikorn-k8shim cache.Context

Policy Group

Set in the scheduler when a new resource manager is registered.

Gang Scheduling

Gang scheduling style can be "hard" (the application will fail after placeholder timeout) or "soft" (after the timeout the application will be scheduled as a normal application).

Organizatorium

1

When a new pod annotated with schedulerName: yunikorn needs scheduling, the API server (admission controller (?)) calls the "admission-webhook.yunikorn.mutate-pods" webhook with a POST https://yunikorn-admission-controller-service.yunikorn.svc:443/mutate?timeout=10s.

Service "yunikorn-admission-controller-service"

When running locally, the service does not get deployed, yet the pods get scheduled. This is how: there's a Kubernetes mechanism involving "informers" that periodically updates the state of the resources is interested in. There are "update", "add" and "delete" notifications. When a new pod shows up, general.Manager.AddPod() is invoked, which creates and Application and Task using the pod metadata → PodEventHandler.addPod() → cache.Context.AddApplication(). At the same time, there's the main KubernetesShim scheduling loop that finds the new application and so the scheduling process begins.