You’re running into this because

prefect deploy

loads/imports your flow at deploy time to derive metadata (e.g., parameter schema) and validate the entrypoint. That import happens on the machine running the CLI. So even if you specify a Docker image in

prefect.yaml

, the CI node still tries to import your flow and its dependencies locally — which explains the missing dependency traceback. Ways to fix it: Option 1 — Run the deploy inside your prebuilt image Since your image already has the code and dependencies, run the CLI from that image in CI. For example:

docker run --rm \
  -e PREFECT_API_URL=$PREFECT_API_URL \
  -e PREFECT_API_KEY=$PREFECT_API_KEY \
  -w /app \
  my-registry/my-image:tag \
  prefect deploy --prefect-file /app/prefect.yaml --name some_name

Notes: - Ensure

prefect.yaml

is in the image at

/app/prefect.yaml

(or mount it:

-v $PWD/prefect.yaml:/app/prefect.yaml

). - Set

-w

to the project root so relative paths in

prefect.yaml

resolve. - This keeps the deployment creation in the same environment your flow uses, avoiding missing deps on the CI host. Option 2 — Install your project deps on the CI node before deploy If you prefer to keep using the CI host, install the dependencies required to import your flow before running

prefect deploy

. This doesn’t need Docker; it just needs enough packages so the module import succeeds. Option 3 — Use a small deploy script inside the image (Python API) If you don’t want to rely on

prefect.yaml

, you can run a short script inside the image that imports the flow and calls

.deploy()

while reusing your prebuilt image (no build/push at deploy time):

python - << 'EOF'
import asyncio
from prefect import flow

# Import your flow from the code inside the image
from my_package.flows import my_flow  # adjust this import

async def main():
    await my_flow.deploy(
        name="some_name",
        work_pool_name="your-work-pool",
        image="my-registry/my-image:tag",
        build=False,  # prebuilt
        push=False,   # already pushed
        job_variables={"image": "my-registry/my-image:tag"},
    )

asyncio.run(main())
EOF

Then run it in CI via:

docker run --rm \
  -e PREFECT_API_URL=$PREFECT_API_URL \
  -e PREFECT_API_KEY=$PREFECT_API_KEY \
  my-registry/my-image:tag \
  bash -lc '<the heredoc above>'

Why the CLI behaves this way - The CLI validates the entrypoint and inspects the flow by importing it. If your module imports heavy libraries at module import time (e.g., pandas) and those aren’t installed on CI, the import fails. - Specifying a Docker image in

prefect.yaml

tells workers what to run at execution time, but it does not change how deployment creation imports the flow. If you can share: - Your Prefect version (assuming 3.x), - The top of the traceback, - The

deployments:

section of your

prefect.yaml

(especially

entrypoint

,

work_pool

, and any

pull

steps), …I can confirm exactly where the import is failing and tailor the command/YAML accordingly.