Prefect User Config not working with Prefect cloud

Description

A clear description of the bug Working with user config files locally works great, as described in this document

However, I’ve tried exporting a user config file (.toml) to prefect cloud (we’re running the Fargate agent). I do this by copying the file into docker storage using the files arg. I then set the environment variable PREFECT__USER_CONFIG_PATH to the path of this .toml file.

The flows get deployed successfully, the file gets copied over and the envar gets set. However, when the flow is executed on cloud, and we call prefect.config.SFK_USER we get the following error:

`DEBUG	Task ‘execute_snowflake_query[0]’: Calling task.run() method…
2019-11-9 3:38pm	prefect.CloudTaskRunner

As an alternative way of deploying the config, rather than setting PREFECT__USER_CONFIG_PATH, I tried parsing the .toml file during the deployment and prefixing every envar with PREFECT__ and just adding these envars to the env_vars arg in docker storage. This produced the same error (above).

Expected Behavior

What did you expect to happen instead? I would expect the value of prefect.config.SFK_USER to get returned, just as it does when I run this locally.

Reproduction

A minimal example that exhibits the behavior.

1. create a file called uat.toml. Inside of this file put SFK_USER = PREFECT_READ_ONLY
2. in your flow code, call prefect.config.SFK_USER within one of your tasks

Deploy Flow 3. when deploying your flow to docker storage, in the files arg, use the source and destination of this file to copy the file into the docker container 4. in the env_vars arg, set PREFECT__USER_CONFIG_PATH to the destination file path of uat.toml

Run the Flow 5. From prefect cloud trigger the newly deployed flow

Environment

Any additional information about your environment We’re running the Fargate Agent

[mhmcdonald]

Thanks, @cicdw. This makes sense. We will definitely use the prefect.* pattern going forward.

I’ll go ahead and close this issue.

[cicdw]

Hi @mhmcdonald ! I’ve been lurking and I think I know what’s going on. It’s a little complicated, but ultimately boils down to how Flows are serialized and stored (using cloudpickle).

Whenever you use the

from prefect import task, Flow, config

@task
def test():
    return config.MY_VAR

pattern, cloudpickle freezes the config object as-is. However, whenever you use

import prefect
from prefect import task, Flow

@task
def test():
    return prefect.config.MY_VAR

cloudpickle recognizes config as something it can access as a module attribute, and therefore at runtime it dynamically loads the configuration object (as you were expecting). In general we recommend the prefect.* pattern for both config and context for this reason.

You can see this in action if you call cloudpickle.dumps(...) on each of the above tasks – you’ll notice that one of them produces a large output with your entire config file, whereas the other one is very minimal.