Troubleshooting

Finding and fixing problems in your Kubeflow deployment

TensorFlow and AVX

There are some instances where you may encounter a TensorFlow-related Python installation or a pod launch issue that results in a SIGILL (illegal instruction core dump). Kubeflow uses the pre-built binaries from the TensorFlow project which, beginning with version 1.6, are compiled to make use of the AVX CPU instruction. This is a recent feature and your CPU might not support it. Check the host environment for your node to determine whether it has this support.

Linux:

  1. grep -ci avx /proc/cpuinfo

AVX2

Some components requirement AVX2 for better performance, e.g. TF Serving.To ensure the nodes support AVX2, we addedminCpuPlatformarg in our deploymentconfig.

On GCP this will fail in regions (e.g. us-central1-a) that do not explicitly have IntelHaswell (even when there are other newer platforms in the region).In that case, please choose another region, or change the config to otherplatformnewer than Haswell.

Minikube

On Minikube the Virtualbox/VMware drivers for Minikube are recommended as there is a knownissue between the KVM/KVM2 driver and TensorFlow Serving. The issue is tracked in kubernetes/minikube#2377.

We recommend increasing the amount of resources Minikube allocates:

  1. minikube start --cpus 4 --memory 8096 --disk-size=40g
  • Minikube by default allocates 2048Mb of RAM for its VM which is not enoughfor JupyterHub.
  • The larger disk is needed to accommodate Kubeflow’s Jupyter images whichare 10s of GBs due to all the extra Python libraries we include.

If you just installed Minikube following instructions from the quick start guide, you most likelycreated a VM with the default resources. You would want to recreate your Minikube with the appropriate resource settings:

  1. minikube stop
  2. minikube delete
  3. minikube start --cpus 4 --memory 8096 --disk-size=40g

You might encounter a jupyter-xxxx pod in Pending status, described with the following warning message:

  1. Warning  FailedScheduling  8s (x22 over 5m)  default-scheduler  0/1 nodes are available: 1 Insufficient memory.
  • Then try recreating your Minikube cluster (and re-apply Kubeflow using kustomize) with more resources (as your environment allows):

RBAC clusters

If you are running on a Kubernetes cluster with RBAC enabled, you may get an error like the following when deploying Kubeflow:

  1. ERROR Error updating roles kubeflow-test-infra.jupyter-role: roles.rbac.authorization.k8s.io "jupyter-role" is forbidden: attempt to grant extra privileges: [PolicyRule{Resources:["*"], APIGroups:["*"], Verbs:["*"]}] user=&{your-user@acme.com  [system:authenticated] map[]} ownerrules=[PolicyRule{Resources:["selfsubjectaccessreviews"], APIGroups:["authorization.k8s.io"], Verbs:["create"]} PolicyRule{NonResourceURLs:["/api" "/api/*" "/apis" "/apis/*" "/healthz" "/swagger-2.0.0.pb-v1" "/swagger.json" "/swaggerapi" "/swaggerapi/*" "/version"], Verbs:["get"]}] ruleResolutionErrors=[]

This error indicates you do not have sufficient permissions. In many cases you can resolve this just by creating an appropriateclusterrole binding like so and then redeploying kubeflow:

  1. kubectl create clusterrolebinding default-admin --clusterrole=cluster-admin --user=your-user@acme.com
  • Replace your-user@acme.com with the user listed in the error message.

If you’re using GKE, you may want to refer to GKE’s RBAC docs to understandhow RBAC interacts with IAM on GCP.

Problems spawning Jupyter pods

This section has been moved to Jupyter Notebooks Troubleshooting Guide.

Pods stuck in Pending state

There are three pods that have Persistent Volume Claims (PVCs) that will get stuck in pending state if they are unable to bind their PVC. The three pods are minio, mysql, and katib-mysql.Check the status of the PVC requests:

  1. kubectl -n ${NAMESPACE} get pvc
  • Look for the status of “Bound”
  • PVC requests in “Pending” state indicate that the scheduler was unable to bind the required PVC.

If you have not configured dynamic provisioning for your cluster, including a default storage class, then you must create a persistent volume for each of the PVCs.

You can use the example below to create local persistent volumes:

  1. sudo mkdir /mnt/pv{1..3}
  3. kubectl create -f - <<EOF
  4. kind: PersistentVolume
  5. apiVersion: v1
  6. metadata:
  7.   name: pv-volume1
  8. spec:
  9.   storageClassName:
  10.   capacity:
  11.     storage: 10Gi
  12.   accessModes:
  13.     - ReadWriteOnce
  14.   hostPath:
  15.     path: "/mnt/pv1"
  16. ---
  17. kind: PersistentVolume
  18. apiVersion: v1
  19. metadata:
  20.   name: pv-volume2
  21. spec:
  22.   storageClassName:
  23.   capacity:
  24.     storage: 20Gi
  25.   accessModes:
  26.     - ReadWriteOnce
  27.   hostPath:
  28.     path: "/mnt/pv2"
  29. ---
  30. kind: PersistentVolume
  31. apiVersion: v1
  32. metadata:
  33.   name: pv-volume3
  34. spec:
  35.   storageClassName:
  36.   capacity:
  37.     storage: 20Gi
  38.   accessModes:
  39.     - ReadWriteOnce
  40.   hostPath:
  41.     path: "/mnt/pv3"
  42. EOF

Once created the scheduler will successfully start the remaining three pods. The PVs may also be created prior to running any of the kfctl commands.

OpenShift

If you are deploying Kubeflow in an OpenShift environment which encapsulates Kubernetes, you will need to adjust the security contexts for the ambassador and Jupyter-hub deployments in order to get the pods to run:

  1. oc adm policy add-scc-to-user anyuid -z ambassador
  2. oc adm policy add-scc-to-user anyuid -z jupyter-hub

Once the anyuid policy has been set, you must delete the failing pods and allow them to be recreated in the project deployment.

You will also need to adjust the privileges of the tf-job-operator service account for TFJobs to run. Do this in the project where you are running TFJobs:

  1. oc adm policy add-role-to-user cluster-admin -z tf-job-operator

403 API rate limit exceeded error

Because kubectl uses GitHub to pull kubeflow, unless user specifies GitHub API token, it will quickly consume maximum API call quota for anonymous.To fix this issue first create GitHub API token using this guide, and assign this token to GITHUB_TOKEN environment variable:

  1. export GITHUB_TOKEN=<< token >>