Great question. Assuming you’re running your flow on a Docker work pool in Prefect 3.x, you can pass Docker container creation options via the deployment’s

job_variables

, using the

container_create_kwargs

key. These values get forwarded to docker-py’s container create call. Example using a Python deployment with `Flow.from_source(...).deploy(...)`:

from prefect.flows import Flow

flow = Flow.from_source(
    source="<https://github.com/owner/repo>",
    entrypoint="path/to/flows.py:my_flow",
)

flow.deploy(
    name="my-docker-deployment",
    work_pool_name="my-docker-pool",
    image="myrepo/myimage:latest",
    job_variables={
        # Common top-level Docker worker variables
        "env": {"PREFECT_LOGGING_LEVEL": "DEBUG"},
        "volumes": ["/host/data:/container/data:ro"],  # prefer this over putting volumes in create kwargs
        "networks": ["my-network"],
        "auto_remove": True,
        "image_pull_policy": "ALWAYS",

        # Additional docker-py create kwargs passed through as-is
        "container_create_kwargs": {
            # examples — set what you need here
            "extra_hosts": {"host.docker.internal": "host-gateway"},
            "shm_size": "2g",
            "ulimits": [{"name": "nofile", "soft": 65536, "hard": 65536}],
            "network_mode": "bridge",
            # GPU example (docker-py accepts dicts here)
            "device_requests": [
                {"Driver": "nvidia", "Count": -1, "Capabilities": [["gpu"]]}
            ],
            # other supported docker create options can go here
        },
    },
)

Notes and tips: - Set pool-wide defaults in the Docker work pool’s base job template (via the UI). Any

job_variables

you pass in

deploy(...)

will override those defaults for this deployment. - Prefer the top-level

volumes

job variable instead of putting

volumes

inside

container_create_kwargs

(Prefect warns if you do both). See the related change note: PR #18233. - There is a

container_create_kwargs

field on the Docker worker job variables; it’s meant for anything not covered by the top-level keys. - Related discussion on customizing Docker worker container parameters: Issue #18544. - General guidance on customizing job variables: Customize job variables. If you’re not using the Docker worker (e.g., Kubernetes worker or Process), this field won’t apply. Can you confirm which worker type you’re using and what specific container options you need to set? I can tailor an example.

Thanks for the details. Short answer: I’m not seeing an open issue specifically about Docker workers ignoring `job_variables`/`container_create_kwargs` when using

Flow.from_source(...).deploy(...)

. Closest related items I can find: - General request to customize Docker worker container parameters (mentions

container_create_kwargs

and GPUs): #18544 - Note about overlap between

volumes

and

container_create_kwargs

(and a warning logged if volumes are set in both): PR #18233 - CLI-only bug where

--all

could drop

job_variables

(not your case since you’re deploying in Python): #17110 To help pinpoint why your args aren’t reaching the Docker worker, could you share: - Your Prefect version and how you’re running the worker (command and version) - The exact

deploy(...)

snippet (especially your

job_variables

) - Worker logs from a run where you expected the options to apply Quick checks you can try now: 1) Confirm the deployment actually has your job variables - In the UI: open the deployment and check the “Job variables/Run config” section - Note: deployment job variables override the work pool’s base job template; workers do not override deployments 2) Use a simple canary var to prove job vars are applied

from prefect import flow
import os

@flow
def my_flow():
    print("FOO=", os.environ.get("FOO"))

flow = flow.from_source(
    "<https://github.com/owner/repo>",
    "path/to/flows.py:my_flow",
)

flow.deploy(
    name="canary",
    work_pool_name="my-docker-pool",
    job_variables={
        "env": {"FOO": "BAR"},
        "container_create_kwargs": {"shm_size": "1g"},
    },
)

- Run it and check the log prints FOO=BAR. If not, the variables aren’t making it through. 3) Enable debug logs on the worker - Start the worker with

PREFECT_LOGGING_LEVEL=DEBUG

and watch for the rendered job payload the worker uses to create the container. It will show the values it’s sending to Docker. 4) Validate key usage and nesting - Top-level recommended keys for Docker worker include:

image

,

env

,

volumes

,

networks

,

network_mode

,

auto_remove

,

image_pull_policy

,

registry_credentials

,

privileged

, etc. - Use

volumes

at the top level (don’t also put volumes under `container_create_kwargs`; see PR #18233) - Use

container_create_kwargs

only for extras like

extra_hosts

,

shm_size

,

ulimits

,

device_requests

, etc. - Be aware that merges may be shallow: setting

container_create_kwargs

at the deployment may overwrite (not merge) a dict set at the pool level. Docs to help: - Customize job variables: Customize job variables If you share your version, deployment snippet, and a redacted worker log line showing the job payload, I can help confirm whether this is a bug or a key/structure issue—and open an issue if needed.