Troubleshooting guide

Tags: Troubleshoot, Basic

The content in this section will help Flyte users isolate the most probable causes for some of the common issues that could arise while getting started with the project.

Before getting started, collect the following information from the underlying infrastructure:

  • Capture the Status column from the output of:
kubectl describe pod <PodName> -n <namespace>

Where <PodName> will typically correspond to the node execution string that you can find in the UI.

  • Pay close attention to the Events section in the output.
  • Also, collect the logs from the Pod:
kubectl logs pods -n <namespace>

Where <namespace> will typically correspond to the Flyte <project>-<domain>, e.g., flytesnacks-development.

Depending on the contents of the logs or the Events, you can try different things:

Debugging Common Execution Errors

Error: Cannot connect to the Docker daemon at unix:///var/run/docker.sock. Is the docker daemon running?

This error will show if you are not running Docker with the native Docker engine in a Linux machine. Most probably you are running Docker via Docker Desktop.

  • If you are using Docker Desktop in macOS, run:
sudo ln -s ~/Library/Containers/com.docker.docker/Data/docker.raw.sock /var/run/docker.sock
  • If you are using Docker Desktop in Linux, run:
sudo ln -s ~$USER/.docker/desktop/docker.sock /var/run/docker.sock

  • If you are using another tool to run Docker, you need to make sure that /var/run/docker.sock is linked to the correct socket file.

    For example, if you are using Rancher Desktop on Linux, run:

sudo ln -s ~$USER/.rd/docker.sock /var/run/docker.sock

message: ‘0/1 nodes are available: 1 Insufficient cpu. preemption: 0/1 nodes are available: 1 No preemption victims found for incoming pod.’

This issue is more common on macOS devices. Make sure that your Docker daemon has allocated a minimum of 4 CPU cores and 3GB of RAM.

terminated with exit code (137). Reason [OOMKilled]

  • For single binary environment deployed with Helm chart, make sure you are using the most recent charts.

  • For EKS deployments, you can adjust resource limits and requests in the inline section of the eks-production.yaml file. Example:

inline:
  task_resources:
    defaults:
      cpu: 100m
      memory: 100Mi
      storage: 100Mi
    limits:
      memory: 1Gi
  • Also, the default container resource limits can be overridden from the task itself:
from flytekit import Resources, task

@task(limits=Resources(mem="256Mi"))
def your_task(...):
    ...

Error: ImagePullBackOff

  • If your environment requires the use of a network proxy, use the --env option when starting the sandbox and pass the proxy configuration:
flytectl demo start --env HTTP_PROXY=<your-proxy-IP>
  • If you’re building a custom Docker image, make sure to use a tag other than latest. Otherwise, the Kubernetes default pull policy will be changed from IfNotPresent to Always, forcing an image pull with every Pod deployment.

Issues Running Workloads

OPENSSL_internal:WRONG_VERSION_NUMBER

  • For flyte-binary: make sure that the endpoint name you have set in your config.yaml file is included in the DNS names of the SSL certificate installed (be it self-signed or issued by a Certificate Authority).
  • For sandbox: verify the FLYTECTL_CONFIG environment variable has the correct value by running:
export FLYTECTL_CONFIG=~/.flyte/config-sandbox.yaml

ModuleNotFoundError

  • If you’re using a custom container image and using Docker, make sure your Dockerfile is located at the same level of the flyte directory and that there is an empty __init__.py file in your project’s folder:
myflyteapp
├── Dockerfile
├── docker_build_and_tag.sh
├── flyte
│   ├── __init__.py
│   └── workflows
│       ├── __init__.py
│       └── example.py
└── requirements.txt

An error occurred (AccessDenied) when calling the PutObject operation in an EKS deployment

  • Make sure that the Kubernetes service account Flyte is using has the annotation that refers to the IAM Role connected to it:
kubectl describe sa <my-flyte-sa> -n <flyte-namespace>

Example output:

Name:                <my-flyte-sa>
Namespace:           flyte
Labels:              app.kubernetes.io/managed-by=eksctl
Annotations:         eks.amazonaws.com/role-arn: arn:aws:iam::<aws-account-id>:role/flyte-system-role
...
  • Otherwise, obtain your IAM role’s ARN and manually annotate the service account:
kubectl annotate serviceaccount -n <flyte-namespace> eks.amazonaws.com/role-arn=arn:aws:iam::xxxx:role/<flyte-iam-role>
  • Refer to this community-maintained guide for further information about Flyte deployment on EKS.

FlyteScopedUserException: ‘JavaPackage’ object is not callable when running a Spark task

Please add spark to the list of enabled-plugins in the config YAML file. For example:

tasks:
  task-plugins:
    enabled-plugins:
      - container
      - sidecar
      - K8S-ARRAY
      - spark
    default-for-task-types:
      - container: container
      - container_array: K8S-ARRAY

authentication handshake failed: x509: “Kubernetes Ingress Controller Fake Certificate” certificate is not trusted when deploying flyte-core to your own Kubernetes cluster

This issue is caused by TLS being disabled in your Kubernetes cluster. You can resolve the problem by following these steps:

  • Enable tls in the values.yaml ingress configuration of flyte-core in order to expose gRPC service at port 443:
ingress:
  host: example.com
  separateGrpcIngress: true
  separateGrpcIngressAnnotations:
    ingress.kubernetes.io/backend-protocol: "grpc"
  annotations:
    ingress.kubernetes.io/app-root: "/console"
    ingress.kubernetes.io/default-backend-redirect: "/console"
    kubernetes.io/ingress.class: haproxy
  tls:
    enabled: true
  • Disable insecure in your flytectl client config.yaml:
admin:
  endpoint: dns:///example.com
  authType: Pkce
  insecure: false
  insecureSkipVerify: true