Documentation Index
Fetch the complete documentation index at: https://mintlify.com/characat0/mlops-fundamentals-homework/llms.txt
Use this file to discover all available pages before exploring further.
MLflow is the connective tissue between your training pipeline and your deployed container. Every time src/train.py runs, it opens an MLflow run, records the hyperparameters and accuracy for that model, and stores the serialised model artifact in the tracking server’s backend store. Once all runs are complete, src/evaluate.py queries those runs, picks the winner, and publishes it to the MLflow Model Registry under the @champion alias. The Dockerfile then downloads exactly that alias at build time — so the container is always running whichever model your pipeline has decided is the current best, with no manual intervention required.
Starting the MLflow Server
All components expect a running MLflow tracking server. Start it in a dedicated terminal before running the pipeline or building the container.
mlflow server --host 0.0.0.0 --port 5000
MLFLOW_TRACKING_URI in your .env file:
# .env
MLFLOW_TRACKING_URI=http://localhost:5000
What Gets Logged in train.py
The training script iterates over every model configuration in params.yaml and opens a separate MLflow run for each one. The pseudocode in the skeleton file illustrates exactly what each run must log:
# data_pipeline/src/train.py (illustrative implementation)
with mlflow.start_run(run_name=model_name):
# 1. Hyperparameters from params.yaml
mlflow.log_params(model_params)
# e.g. {"C": 1.0, "max_iter": 1000} for logistic_regression
# e.g. {"max_depth": 6, "learning_rate": 0.1, "n_estimators": 100} for xgboost
# 2. Fit the model
model.fit(X_to_use, y_encoded)
# 3. Accuracy metric
accuracy = model.score(X_to_use, y_encoded)
mlflow.log_metric("accuracy", accuracy)
# 4. Model artifact — use the flavour-appropriate function
mlflow.sklearn.log_model(model, artifact_path="model")
# OR for XGBoost:
# mlflow.xgboost.log_model(model, artifact_path="model")
params.yaml produce two separate runs in the default MLflow experiment, each visible in the tracking UI with their respective parameters and metrics.
The Champion Alias in evaluate.py
After training, the evaluate stage finds the run with the highest logged accuracy and registers it in the MLflow Model Registry under a fixed model name. Two MLflow client calls implement this:
# data_pipeline/src/evaluate.py
client = mlflow.tracking.MlflowClient()
# 1. Search all runs, ordered by accuracy descending
runs = client.search_runs(
experiment_ids=[experiment.experiment_id],
order_by=["metrics.accuracy DESC"],
max_results=100
)
best_run = runs[0]
model_uri = f"runs:/{best_run.info.run_id}/model"
model_name = "spotify-genre-classifier"
# 2. Register the model artifact as a new version
model_version = client.create_model_version(
name=model_name,
source=model_uri,
run_id=best_run.info.run_id
)
# 3. Tag that version with the @champion alias
client.set_registered_model_alias(
name=model_name,
alias="champion",
version=model_version.version
)
spotify-genre-classifier with version 1 tagged @champion. Any subsequent run of dvc repro that produces a better model will bump the alias to the new version automatically.
Loading the Champion in Docker
The Dockerfile resolves the @champion alias at build time using the mlflow models download command. This bakes the model into the image so the container needs no network access to MLflow at runtime.
mlflow models download -m "models:/spotify-genre-classifier@champion"
RUN instruction:
ARG MLFLOW_TRACKING_URI=http://localhost:5000
ENV MLFLOW_TRACKING_URI=${MLFLOW_TRACKING_URI}
RUN mlflow models download \
-m "models:/spotify-genre-classifier@champion" \
-d ./models
docker build \
--build-arg MLFLOW_TRACKING_URI=http://mlflow-host:5000 \
-t spotify-api:latest .
Open the MLflow UI at http://localhost:5000 after running dvc repro to verify your runs. Navigate to Models → spotify-genre-classifier and confirm that the @champion alias appears next to the version you expect. If the alias is missing, the Docker build will fail with a “Model not found” error. 