I'm running into a bit of a catch 22. I am trying to set up a CI process that registers the flows. Here is an example

import pandas as pd 
from prefect.storage.docker import Docker

# ... task definitions 

with Flow('test_flow', storage= Docker(
            registry_url=ecr_registry_url,
            image_name=a_repo_name,
            python_dependencies=["python==1.2"])
) as flow:
   
    # ... all the steps

I then to to execute

prefect build -p path/to/file.py

and it throws an error that pandas is not installed (which it isnt)

ModuleNotFoundError: No module named 'pandas'

Is there a way to register a flow, without having to install the flow's dependencies first?

Hi @Marc Lipoff! Can I have a bit more detail? Wouldn’t it be a matter of installing the dependencies in the Docker container?

Correct. I made my example a little more clear

I’ll try something and get back to you

Thats unfortunate. Appreciate you checking

@Marc Lipoff I was wrong. What you need to try is deferring the import inside a task. This will defer the execution.

ah ok...

Second, for a more accurate explanation, Prefect runs the flow script to serialize the flow metadata for registration.

I know some users also

try/except ImportErrors

at the top so they can get their flow to register without having the modules available. This is kind of a limitation of Python -- we need to be able to import the flow from that file which requires running the whole file.

Last thing is that you may need to change the Docker storage to

stored_as_script=True

in Docker storage so that the flow doesn’t get serialized.

I'll give it a try. So far, I'm having a few (circumventable) problems, with typing. For example, one function takes df: pd.DataFrame as an input

Oh ok yeah that’s gonna a bit too hard to circumvent