Short answer: yes—keep your scheduler “thin” and move all flow dependencies into the Kubernetes pods that actually run the work. In Prefect 3.x you can do this by: - Creating deployments that pull code from a remote source (e.g., Git) and run in per-deployment Docker images - Having your scheduler only trigger deployments (no importing flow code, no heavy libs installed) Here’s a recommended setup. 1) Package dependencies in per-flow images - Build a Docker image per flow (or per group of flows) with all needed Python libs. - Your Kubernetes work pool will launch pods with the image you specify on each deployment. 2) Create deployments from remote source (no local imports on the scheduler) - Point deployments to your repo and specify the image and work pool. This keeps the scheduler free of flow dependencies. - Example (run this in CI/CD or a thin admin box that just has Prefect installed):

from prefect import flow
from prefect.runner.storage import GitRepository
# If the repo is private, also import GitCredentials

etl = flow.from_source(
    source=GitRepository(
        url="<https://github.com/acme/data-flows.git>",
        # credentials=GitCredentials(username="...", access_token="..."),
        branch="main",
    ),
    entrypoint="flows/etl.py:etl",
)

etl.deploy(
    name="etl-prod",
    work_pool_name="k8s-pool",
    image="<http://ghcr.io/acme/etl:1.2.3|ghcr.io/acme/etl:1.2.3>",
    job_variables={
        "namespace": "prefect",
        "image_pull_policy": "IfNotPresent",
        # add other k8s vars as needed: env, service_account_name, image_pull_secrets, etc.
    },
    # For prebuilt images: build=False, push=False
)

Docs: - Flow code storage with from_source - Job variables (per-deployment K8s settings) - Kubernetes work pools 3) Have your scheduler only trigger deployments - Your scheduler receives events and calls `run_deployment`—no need to import flow code or install its libs.

from prefect.deployments import run_deployment

# Trigger an existing deployment by its "flow-name/deployment-name"
run = run_deployment(
    name="etl/etl-prod",
    parameters={"date": "2025-11-07"},
    # Optional: override the image at trigger-time if you must
    # job_variables={"image": "<http://ghcr.io/acme/etl:1.2.4|ghcr.io/acme/etl:1.2.4>"}
)

- Alternatively, use the Prefect REST API or client to create a flow run from a deployment (still no imports in your scheduler). Docs: Run deployments programmatically API ref: Create flow run from deployment (REST) Notes and gotchas - Avoid removed 2.x patterns in 3.x: don’t use

Deployment.build_from_flow()

or `prefect deployment build`; use

flow.from_source(...).deploy(...)

or

prefect deploy

. - If you must create deployments dynamically on events, you can still do so from your scheduler as long as you: - Reference flows via

from_source(...)

(so you don’t import code locally) - Set

image=...

(and, if the image is prebuilt,

build=False, push=False

) - For private repos, use

GitCredentials

in

GitRepository

. Questions to tailor the best approach for you: - Do you want to create deployments on-the-fly per event, or just trigger pre-created deployments? - Where is your flow code hosted (GitHub/GitLab, private repo)? - Do you already have per-flow Docker images, or should we outline a simple CI step to build/push them?