Troubleshooting

Scheduler logs

Understanding the linkage between Pod UUID, Task TaskID, AllocationAsk AllocationKey and Allocation AllocationID

Pod is always submitted with UID, a unique identifier to differentiate various pods. When a pod is submitted, a Task gets created in the Shim. It uses UID of the POD as TaskID and passed as an AllocationAsk request to the core. AllocationAsk uses Task's TaskID as AllocationKey and passed onto core for further processing. On receiving the ask request, Core tries to find a suitable Allocation using AllocationAsk's AllocationKey as AllocationID. Understanding this flow and its linkage between different objects helps to debug the issues.

An example has been described below to explain how pod's UID is getting translated with different name and passed through different objects.

On task creation,

2023-10-05T10:00:02.224Z  INFO cache/context.go:832 task added  {"appID": "yunikorn-dex-app-mqgh4dw2-autogen", "taskID": "849b762d-68c7-4cce-96e1-5acb545a8620", "taskState": "New"}

On processing allocation,

2023-10-05T10:00:02.523Z  INFO  scheduler/partition.go:890  scheduler allocation processed  {"appID": "yunikorn-dex-app-mqgh4dw2-autogen", "allocationKey": "849b762d-68c7-4cce-96e1-5acb545a8620", "allocationID": "849b762d-68c7-4cce-96e1-5acb545a8620-0", "allocatedResource": "map[memory:343932928 vcore:200]", "placeholder": false, "targetNode": "ip-10-130-86-3.eu-west-1.compute.internal"}

On binding the pod to node,

2023-10-05T10:00:02.523Z  INFO  client/kubeclient.go:112  bind pod to node  {"podName": "dbtdmchicdbdmchicdbstagingdata-1f8b6b53321f4cee9da13a7ac1f2a60c", "podUID": "849b762d-68c7-4cce-96e1-5acb545a8620", "nodeID": "ip-10-130-86-3.eu-west-1.compute.internal"}

Retrieve scheduler logs

Currently, the scheduler writes its logs to stdout/stderr, docker container handles the redirection of these logs to a local location on the underneath node, you can read more document here. These logs can be retrieved by kubectl logs. Such as:

// get the scheduler pod
kubectl get pod -l component=yunikorn-scheduler -n yunikorn
NAME                                  READY   STATUS    RESTARTS   AGE
yunikorn-scheduler-766d7d6cdd-44b82   2/2     Running   0          33h

// retrieve logs
kubectl logs yunikorn-scheduler-766d7d6cdd-44b82 yunikorn-scheduler-k8s -n yunikorn

In most cases, this command cannot get all logs because the scheduler is rolling logs very fast. To retrieve more logs in the past, you will need to setup the cluster level logging. The recommended setup is to leverage fluentd to collect and persistent logs on an external storage, e.g s3.

Set Logging Level

Edit the yunikorn-configs configmap:

kubectl edit configmap/yunikorn-configs -n yunikorn

Add log.level to the data field of the configmap. For example setting log.level to DEBUG sets the logging level to DEBUG.

apiVersion: v1
data:
  log.level: DEBUG
  ...
kind: ConfigMap
metadata:
   ...

The log.level value can be either numeric (-1 through 5) or textual (DEBUG through FATAL).

Value	Logging Level
-1	DEBUG
0	INFO
1	WARN
2	ERROR
3	DPANIC
4	PANIC
5	FATAL

Pods are stuck at Pending state

If some pods are stuck at Pending state, that means the scheduler could not find a node to allocate the pod. There are several possibilities to cause this:

1. Non of the nodes satisfy pod placement requirement

A pod can be configured with some placement constraints, such as node-selector, affinity/anti-affinity, do not have certain toleration for node taints, etc. To debug such issues, you can describe the pod by:

kubectl describe pod <pod-name> -n <namespace>

the pod events will contain the predicate failures and that explains why nodes are not qualified for allocation.

2. The queue is running out of capacity

If the queue is running out of capacity, pods will be pending for available queue resources. To check if a queue is still having enough capacity for the pending pods, there are several approaches:

1. check the queue usage from yunikorn UI

If you do not know how to access the UI, you can refer the document here. Go to the Queues page, navigate to the queue where this job is submitted to. You will be able to see the available capacity left for the queue.

2. check the pod events

Run the kubectl describe pod to get the pod events. If you see some event like: Application <appID> does not fit into <queuePath> queue. That means the pod could not get allocated because the queue is running out of capacity.

The pod will be allocated if some other pods in this queue is completed or removed. If the pod remains pending even the queue has capacity, that may because it is waiting for the cluster to scale up.

Obtain full state dump

A Yunikorn state dump contains the every state object for every process which getting dumped. With endpoint to retrieve we can have many useful information in a single response for troubleshooting for example: list of partitions, list of applications which includes running, completed also historical application details, number of nodes, utilization of nodes, generic cluster information, cluster utilization details, container history and queues information.

The state dump is a valuable resource that Yunikorn offers for use while troubleshooting.

There are a few ways to obtain the full state dump.

1. Scheduler URL

STEPS:

• Open the Scheduler URL in your browser window/tab and edit the URL as follows:
• Replace /#/dashboard with /ws/v1/fullstatedump, (For example, http://localhost:9889/ws/v1/fullstatedump)
• Press Enter

That displays and provides an easy user experience to view live full state dump.

2. Scheduler REST API

With the below scheduler REST API returns information about full state dump used by the YuniKorn Scheduler.

curl -X 'GET' http://localhost:9889/ws/v1/fullstatedump -H 'accept: application/json'

For more details around the content of the state dump, please refer to the documentation on retrieve-full-state-dump

Restart the scheduler

note

In accordance with best practices for troubleshooting, restarting the scheduler should only be done as a last effort to get everything back up and running. It should never be done before gathering all logs and state dumps.

YuniKorn can recover its state upon a restart. YuniKorn scheduler pod is deployed as a deployment, restart the scheduler can be done by scale down and up the replica:

kubectl scale deployment yunikorn-scheduler -n yunikorn --replicas=0
kubectl scale deployment yunikorn-scheduler -n yunikorn --replicas=1

Gang Scheduling

1. No placeholders created, app's pods are pending

Reason: This is usually because the app is rejected by the scheduler, therefore non of the pods are scheduled. The common reasons caused the rejection are: 1) The taskGroups definition is invalid. The scheduler does the sanity check upon app submission, to ensure all the taskGroups are defined correctly, if these info are malformed, the scheduler rejects the app; 2) The total min resources defined in the taskGroups is bigger than the queues' max capacity, scheduler rejects the app because it won't fit into the queue's capacity. Check the pod event for relevant messages, and you will also be able to find more detail error messages from the schedulers' log.

Solution: Correct the taskGroups definition and retry submitting the app.

2. Not all placeholders can be allocated

Reason: The placeholders also consume resources, if not all of them can be allocated, that usually means either the queue or the cluster has no sufficient resources for them. In this case, the placeholders will be cleaned up after a certain amount of time, defined by the placeholderTimeoutInSeconds scheduling policy parameter.

Solution: Note, if the placeholder timeout reaches, currently the app will transit to failed state and can not be scheduled anymore. You can increase the placeholder timeout value if you are willing to wait for a longer time. In the future, a fallback policy might be added to provide some retry other than failing the app.

3. Not all placeholders are swapped

Reason: This usually means the actual app's pods are less than the minMembers defined in the taskGroups.

Solution: Check the minMember in the taskGroup field and ensure it is correctly set. The minMember can be less than the actual pods, setting it to bigger than the actual number of pods is invalid.

4.Placeholders are not cleaned up when the app terminated

Reason: All the placeholders are set an ownerReference to the first real pod of the app, or the controller reference. If the placeholder could not be cleaned up, that means the garbage collection is not working properly.

Solution: check the placeholder ownerReference and the garbage collector in Kubernetes.

Still got questions?

No problem! The Apache YuniKorn community will be happy to help. You can reach out to the community with the following options:

1. Post your questions to dev@yunikorn.apache.org
2. Join the YuniKorn slack channel and post your questions to the #yunikorn-user channel.
3. Join the community sync up meetings and directly talk to the community members.