TopolearnThe point of packaging a model is to run it somewhere else with the same behaviour.
Any pyfunc-compatible model can be served without writing a web app:
mlflow models serve -m "models:/churn@champion" -p 5001This starts a server (using an environment built from the model's requirements, or the current one with --env-manager local) with endpoints:
POST /invocationsfor predictions, accepting JSON such as{"dataframe_split": {"columns": [...], "data": [[...]]}},{"dataframe_records": [...]},{"inputs": ...}for tensors, plus optional"params";GET /pingand/healthfor health checks;/version.
For containers, mlflow models build-docker -m <uri> -n churn-model builds an image that serves the model on port 8080, ready for Kubernetes or a cloud container service. mlflow models generate-dockerfile emits the Dockerfile instead, so you can customise it.
Other deployment paths:
- Batch: load with
mlflow.pyfunc.load_modelin a job, ormlflow.pyfunc.spark_udfto score Spark DataFrames in parallel. - Cloud targets: deployment plugins for SageMaker, Azure ML, Databricks and others via
mlflow deployments.
Before deploying, check the model in isolation with mlflow models predict -m <uri> -i input.json or mlflow.models.predict(...), which runs it in a fresh environment and catches missing dependencies early.
Serve by alias URI so redeploying picks up the new champion; pin the version in the image tag for traceability.
Example
# Shell:
# mlflow models serve -m "models:/iris-clf@champion" -p 5001 --env-manager local
# mlflow models build-docker -m "models:/iris-clf@champion" -n iris-clf:v3
import json
import urllib.request
payload = {
"dataframe_split": {
"columns": ["sepal length (cm)", "sepal width (cm)",
"petal length (cm)", "petal width (cm)"],
"data": [[5.1, 3.5, 1.4, 0.2], [6.7, 3.0, 5.2, 2.3]],
}
}
req = urllib.request.Request(
"http://127.0.0.1:5001/invocations",
data=json.dumps(payload).encode(),
headers={"Content-Type": "application/json"},
)
# with urllib.request.urlopen(req) as resp:
# print(json.loads(resp.read())) # {'predictions': [0, 2]}
print(json.dumps(payload)[:40])Remember
mlflow models serveexposes/invocationsfor any pyfunc modelbuild-dockerpackages model plus server into an image- Validate with
mlflow models predictin a fresh environment before deploying
Watch out for
- Sending a bare list of rows to /invocations; the payload must use a key such as dataframe_split, dataframe_records or inputs.