Log directly with MLflow

The platform uses MLflow behind experiment tracking and the model registry. Training run outside the platform (laptops, in-house GPU servers, CI) also records to the platform with the standard MLflow Python client. You need two things.

• Tracking address: https://<platform address>/mlflow
• An MLflow token — a token bound to a tenant. The token itself is the tenant boundary, so your code needs no tenant setting and no X-Tenant header

Experiments, runs and registered models recorded with a token are visible only in that token's tenant, and appear on the platform's Experiments (실험) and Model Registry (모델 레지스트리) screens as soon as they are recorded.

1. Issue a token — screen#

In the list below, see when each token was issued, when it expires and when it was last used, and Revoke (회수) any token that may have leaked. Requests with a revoked token are rejected with 401. An ADMIN sees every token of the tenant; other members see only their own. A token is revoked by its issuer or an ADMIN. The training:… tokens in the list are ones the platform issued for each training run and revoked when the run ended.

1-b. Issue a token — API#

Where you cannot use the screen, such as CI, issue one through the API.

API=https://mlops.example.com
JWT=$(curl -sS -X POST "$API/auth/jwt/login" \
  --data-urlencode "[email protected]" \
  --data-urlencode "password=<your-password>" | jq -r .access_token)

curl -sS -X POST "$API/api/v1/mlflow/tokens" \
  -H "Authorization: Bearer $JWT" -H "X-Tenant: DEMO" \
  -H 'Content-Type: application/json' \
  -d '{"name": "ci-train", "expires_in_days": 30}'

{"id": "9d32c814-…", "token": "<secret>", "tenant": "DEMO", "workspace": "demo",
 "tracking_uri": "https://mlops.example.com/mlflow"}

Leaving out expires_in_days gives a token with no expiry. Such a token stays valid until revoked, so use it only where you must.

2. Environment variables#

export MLFLOW_TRACKING_URI=https://mlops.example.com/mlflow
export MLFLOW_TRACKING_TOKEN=<issued token>

Do not write the token in code or a repository; pass it through an environment variable or a secret management tool. To check the connection:

import mlflow
print(mlflow.get_tracking_uri())
for exp in mlflow.search_experiments():     # only this tenant's experiments are visible
    print(exp.experiment_id, exp.name)

3. Record#

The code below is for MLflow 3.13.0. Outside the platform, you may name the experiment with set_experiment() — the only place you must not call it is inside a platform training container (the experiment no longer matches the injected run and MLflow raises an exception).

import mlflow
import pandas as pd
from mlflow.models import infer_signature

mlflow.set_experiment("docs-mlflow-direct")          # created if it does not exist

with mlflow.start_run(run_name="baseline") as run:
    mlflow.log_params({"epochs": 3, "lr": 0.001})
    for epoch in range(3):
        mlflow.log_metrics({"train/loss": 1.0 / (epoch + 1),
                            "eval/accuracy": 0.5 + 0.1 * epoch}, step=epoch)
    mlflow.log_text("hello", "notes/summary.txt")     # any file as an artifact

    class Echo(mlflow.pyfunc.PythonModel):
        def predict(self, context, model_input, params=None):
            return model_input

    df = pd.DataFrame({"x": [1.0]})
    info = mlflow.pyfunc.log_model(name="model", python_model=Echo(),
                                   signature=infer_signature(df, df))

• If you pass step, curves are drawn in the metrics tab of the run detail.
• A recorded run shows up right away on the platform's Experiments (실험) screen.

4. Register in the model registry#

mv = mlflow.register_model(info.model_uri, "docs-mlflow-direct")
print(mv.name, mv.version)          # docs-mlflow-direct 1

• Pass info.model_uri (models:/m-…) returned by log_model() to register_model() as is. MLflow 3 keeps logged models outside the run artifacts, so guessing runs:/<run>/model can fail.
• The old way — uploading only the weight file as an artifact and registering it with MlflowClient().create_model_version(source="runs:/…/weights/best.pt") — does register. But without MLmodel, no serving image can be built. A model headed for deployment must be logged with pyfunc.log_model.

Stage#

A version registered directly with the MLflow client has the stage None. A version the platform registers (automatic registration by training, or Register model (모델 등록) in the run detail) goes straight to Staging. Promote to Production through the promotion request → approval procedure on the platform screens, not through the MLflow API. Changing the stage through the MLflow API is not blocked, but it leaves no approval or audit record.

Troubleshooting#