acoustic-carpenter-78188
02/14/2023, 11:05 AMmap
function. The Spark Executors can only access the code from the image it pulled.
Expected behavior
The Spark Executors should use the same code as the Spark Driver - which is the code that was pushed using Fast Registration.
Additional context to reproduce
1. Following this exemple : use the code from https://docs.flyte.org/projects/cookbook/en/stable/auto/integrations/kubernetes/k8s_spark/pyspark_pi.html , building a Docker image like this one https://docs.flyte.org/projects/cookbook/en/stable/auto/integrations/kubernetes/k8s_spark/index.html#step-2-environment-setup
2. pyflyte --pkgs my_code --image my_image --force
3. flytectl register files --project my_project --domain development --archive flyte-package.tgz --outputLocationPrefix <s3://my_bucket> --k8sServiceAccount my_account --version 0.0.1
4. Go to the Flyte Console, run the workflow : it works as expected, the output of my_spark
is close to 3.14
5. Then, change the code from the function f
in step 1: now, f
always returns 1.
6. I updated my code but I don't need to update my environnement, so I can use Fast Registration :
pyflyte --pkgs my_code --image my_image --force --fast
7. flytectl register files --project my_project --domain development --archive flyte-package.tgz --outputLocationPrefix <s3://my_bucket> --k8sServiceAccount my_account --version 0.0.2
8. Go to the Flyte Console, run the workflow's new version : the output of my_spark
is close to 3.14, but I was expecting it to be 4.
Screenshots
No response
Are you sure this issue hasn't been raised already?
☑︎ Yes
Have you read the Code of Conduct?
☑︎ Yes
flyteorg/flyte