In Prefect 3.x, for a Managed work pool you pass extra pip packages via the

job_variables

argument to

Flow.deploy()

. The key is

pip_packages

and it expects a list of pip requirement strings. Example:

from prefect import flow

@flow
def my_flow():
    return "hello"

if __name__ == "__main__":
    my_flow.deploy(
        name="my-managed-deployment",
        work_pool_name="managed",
        job_variables={
            "pip_packages": [
                "pandas==2.2.2",
                "pyarrow>=15",
                "prefect-aws>=0.4.0",
            ],
        },
    )

Notes: -

pip_packages

entries can be any valid pip requirement specifier (pinned versions, ranges, extras). - These packages are installed into the managed runtime before your flow runs. Docs for more details: - Managed work pool - Customize job variables

Thanks for sharing that snippet — a couple of things in there can cause the managed run to fail before you see useful logs. What to fix first - Remove the angle brackets. Slack wraps URLs like <https://...>, but those literal

<

and

>

will break pip parsing if copied into your

job_variables

. - Use valid PEP 508 syntax for VCS requirements. It should be

name @ git+https://...

(note the space on both sides of

@

), or the legacy

#egg=

form. - Make sure your token is URL-encoded (tokens sometimes contain characters that must be percent-encoded) and has access to the repo. - Prefer pinning to a commit SHA instead of a branch for reliability. Working examples - PEP 508 style (recommended):

{
  "pip_packages": [
    "dataflows @ git+<https://x-access-token:<YOUR_TOKEN>@github.com/OWNER/REPO.git@main#subdirectory=path/to/pkg>"
  ]
}

- Legacy

#egg=

style (also valid):

{
  "pip_packages": [
    "git+<https://x-access-token:<YOUR_TOKEN>@github.com/OWNER/REPO.git@main#egg=dataflows&subdirectory=path/to/pkg>"
  ]
}

Notes - Do not include literal angle brackets in the actual value. Only replace

<YOUR_TOKEN>

with your real token (URL-encoded). - If you’re using a GitHub App installation token, the

x-access-token:<TOKEN>

pattern is fine. If you’re using a PAT,

https://<USERNAME>:<TOKEN>@github.com/...

also works. - If you continue to see failures, the most robust alternative is to build a custom image with this dependency pre-installed and pass it to

Flow.deploy(image=...)

. This avoids VCS installs at runtime entirely. If it still fails: - Could you paste the exact

pip_packages

entry you’re using (with the token redacted), making sure it’s without Slack’s

< >

markup? - Is the repo private? If so, confirm the token scopes (read access to that repo). - If you can, try swapping

@main

for a commit SHA to rule out branch-related changes. Related docs: - Managed work pool - Customize job variables