Module Not Found Error, while running workflows in FlyteConsole

[sakthi-tech4th]

Hi,

I created a simple worflow with one task, that reads a csv file into a pandas dataframe.
My code is something like below.
`
import os
import sys
import pandas as pd
sys.path.append(r'path_to_my_python_modules')

from data.data_ingestion import load_data # data is folder in path_to_my_models, which contains the file data_ingestion.py.

@task
def getdata(data_path: str='', some_int: int=0) -> pd.DataFrame:
df =load_data(data_path=data_path, some_int=some_int)
return df

@workflow
def my_wf(my_path: str='', my_int: int=0) -> pd.DataFrame:
df = getdata(data_path=my_path, some_int=my_int)
return df

if name == "main":
data_root = os.getenv('raw_data_folder')
data_path = os.path.join(data_root, 'my_data.csv')
print(data_path)
some_int= 6
data_df = my_wf(my_path=data_path, my_int=some_int)
print(data_df)
`
When i run the code locally using:
'python my_pipeline.py'
I have no errors, outputs are as expected.

However, when I run pyflyte:
pyflyte run --remote my_pipeline.py my_wf

I get Module Not Found Error.
from data.data_ingestion import load_data
ModuleNotFoundError: No module named 'data'

I believe, I have include my modules in the dockerimage/container where my code is running.
Could you please guide me on how to make my modules available to the workflow?

[samhita-alla]

@sakthi-tech4th, pyflyte run command operates on a single file. For you to be able to use code in a different module, you'll have to run pyflyte register command: pyflyte register <package-name>.

[laruokol]

@samhita-alla
I'm trying to do a similar thing. I have the following structure for my code:

project_root/
├── flyte/
   ├── __init__.py
   ├── tasks.py
   ├── utils.py
   └── workflow.py

I have run the ./docker_build_and_tag.sh script to build a container image.
Then packaged the workflow:
pyflyte --pkgs flyte package --image <image-name>
registered the workflow:
flytectl register files --project mlexample --domain development --archive flyte-package.tgz --version v1
registered the code (as instructed above):
pyflyte register flyte --project mlexample --domain development
created a launchplan:
flytectl get launchplan --project mlexample --domain development src.workflow.wf --latest --execFile exec_spec.yaml
and created an execution:
flytectl create execution --project mlexample --domain development --execFile exec_spec.yaml

I have also tried to pyflyte run --remote Flyte/workflow.py wf, but all attempts end with ModuleNotFoundError, when tasks are imported from flyte.tasks in workflow.py. I'm pretty much out of ideas here.

[samhita-alla]

@laruokol, you're trying to use code from a different Python module. pyflyte register is the preferred command for this case. pyflyte run is for single files. pyflyte package + flytectl register works as well but pyflyte register is the easiest to get started with and test. So let's zero in on the pyflyte register command.

How are you importing tasks in the workflow.py file? Have you tried doing a relative import: from .tasks import ...?

[samhita-alla]

Um, you can ignore the tar warning. So the issue now is that the pod's taking forever to get started. Let me ask my team to take a look at this. It seems like an issue with M2 chip cause the code works fine on intel.

[samhita-alla]

@laruokol, could you wait some time for it to get started -- 30 mins or so? Let's see if it works then.

[laruokol]

However,
I decided to give custom packaging a one more try:

pyflyte --pkgs flyte.workflows package --image "my_flyte_project:v1"
pyflyte register --image "my_flyte_project:v1" src/test.py

This did not work, but, from the error I deduced that the python version that the container was built on was too old. Setting that to a newer version (3.10) did the trick!

[laruokol]

Thank you @samhita-alla for your support on this issue :)

[samhita-alla]

Awesome! Glad that you got it working.

[highfly22]

Just add ENV PYTHONPATH=/root in your Dockerfile.