Sequential Admission with Ready Pods

Simple implementation of the all-or-nothing scheduling

Some jobs need all pods to be running at the same time to operate; for example, synchronized distributed training or MPI-based jobs which require pod-to-pod communication. On a default Kueue configuration, a pair of such jobs may deadlock if the physical availability of resources do not match the configured quotas in Kueue. The same pair of jobs could run to completion if their pods were scheduled sequentially.

To address this requirement, in version 0.3.0 we introduced an opt-in mechanism configured via the flag waitForPodsReady that provides a simple implementation of the all-or-nothing scheduling. When enabled, admission of workloads is blocked not only on the availability of quota, but also until all previously admitted jobs have their pods scheduled (all pods are running or completed, up the the level of the job parallelism).

This page shows you how to configure Kueue to use waitForPodsReady, which is a simple implementation of the all-or-nothing scheduling. The intended audience for this page are batch administrators.

Before you begin

Make sure the following conditions are met:

  • A Kubernetes cluster is running.
  • The kubectl command-line tool has communication with your cluster.
  • Kueue is installed in version 0.3.0 or later.

Enabling waitForPodsReady

Follow the instructions described here to install a release version by extending the configuration with the following fields:

      enable: true
      timeout: 10m
      blockAdmission: true
        timestamp: Eviction | Creation
        backoffLimitCount: 5

The timeout (waitForPodsReady.timeout) is an optional parameter, defaulting to 5 minutes.

When the timeout expires for an admitted Workload, and the workload’s pods are not all scheduled yet (i.e., the Workload condition remains PodsReady=False), then the Workload’s admission is cancelled, the corresponding job is suspended and the Workload is re-queued.

The blockAdmission (waitForPodsReady.blockAdmission) is an optional parameter, defaulting to false. When the blockAdmission is set to true regardless of the enable (waitForPodsReady.enable) is set to false, the enable is overridden to true. When the blockAdmission is set to true, admitted Workload with not ready pods block admission for other Workloads.

Requeuing Strategy

The requeuingStrategy (waitForPodsReady.requeuingStrategy) contains optional parameters: timestamp (waitForPodsReady.requeuingStrategy.timestamp) and backoffLimitCount (waitForPodsReady.requeuingStrategy.backoffLimitCount).

The timestamp field defines which timestamp Kueue uses to order the Workloads in the queue:

  • Eviction (default): The lastTransitionTime of the Evicted=true condition with PodsReadyTimeout reason in a Workload.
  • Creation: The creationTimestamp in a Workload.

If you want to re-queue a Workload evicted by the PodsReadyTimeout back to its original place in the queue, you should set the timestamp to the Creation.

Kueue will re-queue a Workload evicted by the PodsReadyTimeout reason until the number of re-queues reaches backoffLimitCount. If you don’t specify any value for backoffLimitCount, a Workload is repeatedly and endlessly re-queued to the queue based on the timestamp. Once the number of re-queues reaches the limit, Kueue deactivates the Workload.


In this example we demonstrate the impact of enabling waitForPodsReady in Kueue. We create two jobs which both require all their pods to be running at the same time to complete. The cluster has enough resources to support running one of the jobs at the same time, but not both.

Note In this example we use a cluster with autoscaling disabled in order to simulate issues with resource provisioning to satisfy the configured cluster quota.

1. Preparation

First, check the amount of allocatable memory in your cluster. In many cases this can be done with this command:

TOTAL_ALLOCATABLE=$(kubectl get node --selector='!,!' -o jsonpath='{range .items[*]}{.status.allocatable.memory}{"\n"}{end}' | numfmt --from=auto | awk '{s+=$1} END {print s}')

In our case this outputs 8838569984 which, for the purpose of the example, can be approximated as 8429Mi.

Configure ClusterQueue quotas

We configure the memory flavor by doubling the total memory allocatable in our cluster, in order to simulate issues with provisioning.

Save the following cluster queues configuration as cluster-queues.yaml:

kind: ResourceFlavor
  name: "default-flavor"
kind: ClusterQueue
  name: "cluster-queue"
  namespaceSelector: {}
  - coveredResources: ["memory"]
    - name: "default-flavor"
      - name: "memory"
        nominalQuota: 16858Mi # double the value of allocatable memory in the cluster         
kind: LocalQueue
  namespace: "default"
  name: "user-queue"
  clusterQueue: "cluster-queue"

Then, apply the configuration by:

kubectl apply -f cluster-queues.yaml

Prepare the job template

Save the following job template in the job-template.yaml file. Note the _ID_ placeholders which will be replaced to create configurations for the two jobs. Also, pay attention to configure the memory field for the container as be 75% of the total allocatable memory per pod. In our case this is 75%*(8429Mi/20)=316Mi. In this scenario there is not enough resources to run all pods for both jobs at the same time, risking deadlock.

apiVersion: v1
kind: Service
  name: svc_ID_
  clusterIP: None
    job-name: job_ID_
  - name: http
    protocol: TCP
    port: 8080
apiVersion: v1
kind: ConfigMap
  name: script-code_ID_
data: |
    from http.server import BaseHTTPRequestHandler, HTTPServer
    from urllib.request import urlopen
    import sys, os, time, logging

    logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
    serverPort = 8080
    INDEX_COUNT = int(sys.argv[1])
    index = int(os.environ.get('JOB_COMPLETION_INDEX'))
    logger = logging.getLogger('LOG' + str(index))

    class WorkerServer(BaseHTTPRequestHandler):
        def do_GET(self):
            if "exit" in self.path:
              self.wfile.write(bytes("Exiting", "utf-8"))
              self.wfile.write(bytes("Running", "utf-8"))

    def call_until_success(url):
      while True:
"Calling URL: " + url)
          with urlopen(url) as response:
            response_content ='utf-8')
  "Response content from %s: %s" % (url, response_content))
        except Exception as e:
          logger.warning("Got exception when calling %s: %s" % (url, e))

    if __name__ == "__main__":
      if index == 0:
        for i in range(1, INDEX_COUNT):
          call_until_success("http://job_ID_-%d.svc_ID_:8080/ping" % i)"All workers running")

        time.sleep(10) # sleep 10s to simulate doing something

        for i in range(1, INDEX_COUNT):
          call_until_success("http://job_ID_-%d.svc_ID_:8080/exit" % i)"All workers stopped")
        webServer = HTTPServer(("", serverPort), WorkerServer)"Server started at port %s" % serverPort)

apiVersion: batch/v1
kind: Job
  name: job_ID_
  labels: user-queue
  parallelism: 20
  completions: 20
  completionMode: Indexed
  suspend: true
      subdomain: svc_ID_
      - name: script-volume
          name: script-code_ID_
      - name: main
        image: python:bullseye
        command: ["python"]
        - /script-path/
        - "20"
        - containerPort: 8080
        imagePullPolicy: IfNotPresent
            memory: "316Mi" # choose the value as 75% * (total allocatable memory / 20)
          - mountPath: /script-path
            name: script-volume
      restartPolicy: Never
  backoffLimit: 0

Additional quick job

We also prepare an additional job to increase the variance in the timings to make the deadlock more likely. Save the following yaml as quick-job.yaml:

apiVersion: batch/v1
kind: Job
  name: quick-job
  annotations: user-queue
  parallelism: 50
  completions: 50
  suspend: true
      restartPolicy: Never
      - name: sleep
        image: bash:5
        command: ["bash"]
        args: ["-c", 'echo "Hello world"']
            memory: "1"
  backoffLimit: 0

2. Induce a deadlock under the default configuration (optional)

Run the jobs

sed 's/_ID_/1/g' job-template.yaml > /tmp/job1.yaml
sed 's/_ID_/2/g' job-template.yaml > /tmp/job2.yaml
kubectl create -f quick-job.yaml
kubectl create -f /tmp/job1.yaml
kubectl create -f /tmp/job2.yaml

After a while check the status of the pods by

kubectl get pods

The output is like this (omitting the pods of the quick-job for brevity):

job1-0-9pvs8    1/1     Running     0          28m
job1-1-w9zht    1/1     Running     0          28m
job1-10-fg99v   1/1     Running     0          28m
job1-11-4gspm   1/1     Running     0          28m
job1-12-w5jft   1/1     Running     0          28m
job1-13-8d5jk   1/1     Running     0          28m
job1-14-h5q8x   1/1     Running     0          28m
job1-15-kkv4j   0/1     Pending     0          28m
job1-16-frs8k   0/1     Pending     0          28m
job1-17-g78g8   0/1     Pending     0          28m
job1-18-2ghmt   0/1     Pending     0          28m
job1-19-4w2j5   0/1     Pending     0          28m
job1-2-9s486    1/1     Running     0          28m
job1-3-s9kh4    1/1     Running     0          28m
job1-4-52mj9    1/1     Running     0          28m
job1-5-bpjv5    1/1     Running     0          28m
job1-6-7f7tj    1/1     Running     0          28m
job1-7-pnq7w    1/1     Running     0          28m
job1-8-7s894    1/1     Running     0          28m
job1-9-kz4gt    1/1     Running     0          28m
job2-0-x6xvg    1/1     Running     0          28m
job2-1-flkpj    1/1     Running     0          28m
job2-10-vf4j9   1/1     Running     0          28m
job2-11-ktbld   0/1     Pending     0          28m
job2-12-sf4xb   1/1     Running     0          28m
job2-13-9j7lp   0/1     Pending     0          28m
job2-14-czc6l   1/1     Running     0          28m
job2-15-m77zt   0/1     Pending     0          28m
job2-16-7p7fs   0/1     Pending     0          28m
job2-17-sfdmj   0/1     Pending     0          28m
job2-18-cs4lg   0/1     Pending     0          28m
job2-19-x66dt   0/1     Pending     0          28m
job2-2-hnqjv    1/1     Running     0          28m
job2-3-pkwhw    1/1     Running     0          28m
job2-4-gdtsh    1/1     Running     0          28m
job2-5-6swdc    1/1     Running     0          28m
job2-6-qb6sp    1/1     Running     0          28m
job2-7-grcg4    0/1     Pending     0          28m
job2-8-kg568    1/1     Running     0          28m
job2-9-hvwj8    0/1     Pending     0          28m

These jobs are now deadlock’ed and are not going to be able to make progress.


Clean up the jobs by:

kubectl delete -f quick-job.yaml
kubectl delete -f /tmp/job1.yaml
kubectl delete -f /tmp/job2.yaml

3. Run with waitForPodsReady enabled

Enable waitForPodsReady

Update the Kueue configuration following the instructions here.

Run the jobs

Run the script

sed 's/_ID_/1/g' job-template.yaml > /tmp/job1.yaml
sed 's/_ID_/2/g' job-template.yaml > /tmp/job2.yaml
kubectl create -f quick-job.yaml
kubectl create -f /tmp/job1.yaml
kubectl create -f /tmp/job2.yaml

Monitor the progress

Execute the following command in a couple of seconds internals to monitor the progress:

kubectl get pods

We omit the pods of the completed quick job for brevity.

Output when job1 is starting up, note that job2 remains suspended:

NAME            READY   STATUS              RESTARTS   AGE
job1-0-gc284    0/1     ContainerCreating   0          1s
job1-1-xz555    0/1     ContainerCreating   0          1s
job1-10-2ltws   0/1     Pending             0          1s
job1-11-r4778   0/1     ContainerCreating   0          1s
job1-12-xx8mn   0/1     Pending             0          1s
job1-13-glb8j   0/1     Pending             0          1s
job1-14-gnjpg   0/1     Pending             0          1s
job1-15-dzlqh   0/1     Pending             0          1s
job1-16-ljnj9   0/1     Pending             0          1s
job1-17-78tzv   0/1     Pending             0          1s
job1-18-4lhw2   0/1     Pending             0          1s
job1-19-hx6zv   0/1     Pending             0          1s
job1-2-hqlc6    0/1     ContainerCreating   0          1s
job1-3-zx55w    0/1     ContainerCreating   0          1s
job1-4-k2tb4    0/1     Pending             0          1s
job1-5-2zcw2    0/1     ContainerCreating   0          1s
job1-6-m2qzw    0/1     ContainerCreating   0          1s
job1-7-hgp9n    0/1     ContainerCreating   0          1s
job1-8-ss248    0/1     ContainerCreating   0          1s
job1-9-nwqmj    0/1     ContainerCreating   0          1s

Output when job1 is running and job2 is now unsuspended as job has all the required resources assigned:

job1-0-gc284    1/1     Running     0          9s
job1-1-xz555    1/1     Running     0          9s
job1-10-2ltws   1/1     Running     0          9s
job1-11-r4778   1/1     Running     0          9s
job1-12-xx8mn   1/1     Running     0          9s
job1-13-glb8j   1/1     Running     0          9s
job1-14-gnjpg   1/1     Running     0          9s
job1-15-dzlqh   1/1     Running     0          9s
job1-16-ljnj9   1/1     Running     0          9s
job1-17-78tzv   1/1     Running     0          9s
job1-18-4lhw2   1/1     Running     0          9s
job1-19-hx6zv   1/1     Running     0          9s
job1-2-hqlc6    1/1     Running     0          9s
job1-3-zx55w    1/1     Running     0          9s
job1-4-k2tb4    1/1     Running     0          9s
job1-5-2zcw2    1/1     Running     0          9s
job1-6-m2qzw    1/1     Running     0          9s
job1-7-hgp9n    1/1     Running     0          9s
job1-8-ss248    1/1     Running     0          9s
job1-9-nwqmj    1/1     Running     0          9s
job2-0-djnjd    1/1     Running     0          3s
job2-1-trw7b    0/1     Pending     0          2s
job2-10-228cc   0/1     Pending     0          2s
job2-11-2ct8m   0/1     Pending     0          2s
job2-12-sxkqm   0/1     Pending     0          2s
job2-13-md92n   0/1     Pending     0          2s
job2-14-4v2ww   0/1     Pending     0          2s
job2-15-sph8h   0/1     Pending     0          2s
job2-16-2nvk2   0/1     Pending     0          2s
job2-17-f7g6z   0/1     Pending     0          2s
job2-18-9t9xd   0/1     Pending     0          2s
job2-19-tgf5c   0/1     Pending     0          2s
job2-2-9hcsd    0/1     Pending     0          2s
job2-3-557lt    0/1     Pending     0          2s
job2-4-k2d6b    0/1     Pending     0          2s
job2-5-nkkhx    0/1     Pending     0          2s
job2-6-5r76n    0/1     Pending     0          2s
job2-7-pmzb5    0/1     Pending     0          2s
job2-8-xdqtp    0/1     Pending     0          2s
job2-9-c4rcl    0/1     Pending     0          2s

Once job1 completes, it frees the resources required by job2 to run its pods to make progress. Finally, all jobs complete.


Clean up the jobs by:

kubectl delete -f quick-job.yaml
kubectl delete -f /tmp/job1.yaml
kubectl delete -f /tmp/job2.yaml


When enabling waitForPodsReady, the admission of Workloads may be unnecessarily slowed down by sequencing in case the cluster has enough resources to support concurrent Workload startup.

Last modified March 25, 2024: typo fixes on site (#1901) (f982cda)