Serving code that uses the NumPy package

[wigging]

Below is a Python script that imports the NumPy package to multiply two arrays. This script is part of a uv project which uses a pyproject.toml file to define the dependencies. This code is on Computer A and the Prefect server is running on Computer B which is on the same network.

# example.py

import numpy as np
from prefect import flow


@flow(log_prints=True)
def multiply_arrays():
    a = np.array([1, 2, 3, 4, 5])
    b = np.array([3, 4, 5, 6, 7])
    c = a * b
    print("Array c is", c)


if __name__ == "__main__":
    multiply_arrays.serve(
        name="my-array-deploy",
        tags=["compute"],
        interval=60
    )

On Computer A, I run this code with:

uv run example.py

and here is some of the terminal output on Computer A after running the script:

Your flow 'multiply-arrays' is being served and polling for scheduled runs!

To trigger a run for this flow, use the following command:

        $ prefect deployment run 'multiply-arrays/my-array-deploy'

You can also run your flow via the Prefect UI: http://x.x.x.x:4200/deployments/deployment/621f5d39-5d18-415c-98ba-24b4bab91ea9

20:33:51.258 | INFO    | prefect.flow_runs.runner - Runner 'my-array-deploy' submitting flow run '0198e3ca-cef7-7bd6-8b76-a4aa8d451c54'
20:33:51.298 | INFO    | prefect.flow_runs.runner - Opening process...
20:33:51.306 | INFO    | prefect.flow_runs.runner - Completed submission of flow run '0198e3ca-cef7-7bd6-8b76-a4aa8d451c54'
20:33:51.786 | INFO    | Flow run 'debonair-cockatoo' - Downloading flow code from storage at '.'
20:33:52.909 | INFO    | Flow run 'debonair-cockatoo' - Beginning flow run 'debonair-cockatoo' for flow 'multiply-arrays'
20:33:52.911 | INFO    | Flow run 'debonair-cockatoo' - View at http://x.x.x.x:4200/runs/flow-run/0198e3ca-cef7-7bd6-8b76-a4aa8d451c54
20:33:52.912 | INFO    | Flow run 'debonair-cockatoo' - Array c is [ 3  8 15 24 35]
20:33:52.946 | INFO    | Flow run 'debonair-cockatoo' - Finished in state Completed()
20:33:53.062 | INFO    | prefect.flow_runs.runner - Process for flow run 'debonair-cockatoo' exited cleanly.
20:34:46.263 | INFO    | prefect.flow_runs.runner - Runner 'my-array-deploy' submitting flow run '0198e3ca-cef7-762e-bb06-dabe0e28c892'
20:34:46.303 | INFO    | prefect.flow_runs.runner - Opening process...
20:34:46.326 | INFO    | prefect.flow_runs.runner - Completed submission of flow run '0198e3ca-cef7-762e-bb06-dabe0e28c892'
20:34:46.818 | INFO    | Flow run 'funny-centipede' - Downloading flow code from storage at '.'
20:34:52.942 | INFO    | Flow run 'funny-centipede' - Beginning flow run 'funny-centipede' for flow 'multiply-arrays'
20:34:52.944 | INFO    | Flow run 'funny-centipede' - View at http://x.x.x.x:4200/runs/flow-run/0198e3ca-cef7-762e-bb06-dabe0e28c892
20:34:52.945 | INFO    | Flow run 'funny-centipede' - Array c is [ 3  8 15 24 35]
20:34:52.980 | INFO    | Flow run 'funny-centipede' - Finished in state Completed()
20:34:53.078 | INFO    | prefect.flow_runs.runner - Process for flow run 'funny-centipede' exited cleanly.

Everything seems to be working fine but since Computer B does not have the NumPy package then how is it executing the code? I read the How to run flows in local processes example and it makes it sound like the flow is run on the server. And in the output I shared above, it displays Downloading flow code from storage at '.' which makes me think the code is download to the server. So is the virtual environment on Computer A being downloaded to the Prefect server on Computer B?

[desertaxle]

Hey @wigging! When you run a script with .serve, a couple of things happen:

1. A deployment is registered with the Prefect server. A deployment contains information about how to run your code, but doesn't include your code. In this case, the deployment is given a schedule that says your flow should run every 60 seconds.
2. A long-running process starts on machine A and starts polling the Prefect server for scheduled flow runs.

Each time the server tells the process to start a run, the process on machine A runs your code. Prefect offers the ability to store your code remotely and pull it onto a machine before executing, However, your code is already available on machine A, so the Downloading flow code from storage at '.' message is not entirely accurate. It'd be more accurate to say that the flow code is loaded from the current working directory.

All in all, your flow code is never sent to the server and runs in the same environment it's served in, so numpy and any other dependencies you installed will be available when your code runs.

Let me know if you have any other questions while trying out Prefect!

[wigging]

Ah, ok, so the flow code on Computer A is still run on Computer A. The Prefect server is just telling the flow code on Computer A how often to run. So if I have a function like

import numpy as np

def multiply_arrays():
    a = np.array([1, 2, 3, 4, 5])
    b = np.array([3, 4, 5, 6, 7])
    c = a * b
    print("Array c is", c)

then how would I use Prefect to run that function on a remote machine?

[desertaxle]

If you want to run a function on a remote machine, there are a couple of ways to do it.

1. You could build your script with .serve into a Docker image and use that Docker image to host an instance somewhere. This is a guide for how to build and host an image like that.
2. You could deploy your flow to a work pool. That will require creating a work pool, running a worker for that work pool, storing your code in a network accessible location (like GitHub repository, and S3 bucket, or a Docker image), and creating a deployment to tie all of that configuration together. How to do this depends on what type of infrastructure you want to use, but this guide of how to using a Docker workpool is a good way to see how these concepts fit together.