Skip to main content

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.

This page documents the most common mistakes students make in the MLOps Fundamentals Homework and exactly how to fix them. Work through these before opening your PR — each pitfall below has caused students to lose points that were otherwise within reach.
Year 2010 must go to the training set (year <= 2010), not to the production set. The single most common mistake is using a strict less-than (year < 2010), which silently misassigns every 2010 row to the production split.
# ❌ Wrong — year 2010 rows end up in prod_sim instead of train
train_df = df[df['year'] < year_threshold]
prod_df  = df[df['year'] >= year_threshold]

# ✅ Right — year 2010 rows are included in training
train_df = df[df['year'] <= year_threshold]
prod_df  = df[df['year'] > year_threshold]
The test test_process_data_year_boundary_condition specifically checks this boundary condition. If it fails, this is the first thing to verify.
scikit-learn models require numeric targets. Passing raw string genre values (e.g., 'Rock', 'Pop') directly to .fit() raises a ValueError. You must encode the labels with LabelEncoder before fitting any model.
from sklearn.preprocessing import LabelEncoder

le = LabelEncoder()
y_encoded = le.fit_transform(y)  # Converts 'Rock' → 0, 'Pop' → 1, etc.
Keep the fitted le object — you will need it to decode numeric predictions back to genre strings when the API returns results.
If you train a model without an active mlflow.start_run() context, the run is recorded in the default experiment but carries no parameters, no metrics, and no model artifact. evaluate.py may still find the run, but graders check the MLflow UI explicitly for proper run naming and logged values.Open the UI and confirm each run shows:
  • The model’s hyperparameters (e.g., C for Logistic Regression, max_depth for XGBoost)
  • An accuracy metric
  • A logged model artifact
# Start the MLflow tracking server if it's not already running
mlflow server --host 0.0.0.0 --port 5000
Then navigate to http://localhost:5000 and confirm both the Logistic Regression and XGBoost runs appear with all expected values before running evaluate.py.
The test payload in test_api.py sends exactly 12 fields. If SpotifyFeatures is missing any field, has an extra field with a different name, or uses a wrong type (e.g., duration_ms as float instead of int), the endpoint returns 422 Unprocessable Entity instead of 200 OK and the test fails.The exact fields and types that must match:
FieldType
danceabilityfloat
energyfloat
keyint
loudnessfloat
modeint
speechinessfloat
acousticnessfloat
instrumentalnessfloat
livenessfloat
valencefloat
tempofloat
duration_msint
Pay special attention to key, mode, and duration_ms — these are integers, not floats.
The RUN mlflow models download step in the Dockerfile requires two things to be true at build time:
  1. The MLflow tracking server must be reachable from the machine running docker build.
  2. The spotify-genre-classifier model must already have a @champion alias registered in the Model Registry.
The correct build order is:
  1. Run dvc repro in data_pipeline/ to execute the full training pipeline.
  2. Open the MLflow UI at http://localhost:5000 and verify the champion alias is assigned.
  3. Only then run docker build.
# Ensure MLflow is running before building
mlflow server --host 0.0.0.0 --port 5000

# Override the tracking URI if your MLflow server is on a non-default host
docker build --build-arg MLFLOW_TRACKING_URI=http://localhost:5000 -t spotify-api:latest .
If the build fails with “Connection refused”, the MLflow server is not running or is not accessible from the Docker build context.
If you create new files (e.g., a utility module or helper script) but forget to git add them, those files exist locally but are invisible to CI. The result is an ImportError or ModuleNotFoundError in the Actions run even though all tests pass on your machine.Always run git status before pushing to verify every changed or new file is staged:
git status          # Confirm no untracked files relevant to your implementation
git add .
git commit -m "feat: add helper utilities"
git push origin solution/<your-name>
By default, flake8 flags lines longer than 79 characters as E501. Long lines most commonly appear in MLflow logging calls, dictionary literals, and URL strings. There are two clean ways to handle this:Option 1 — use Python’s implicit line continuation inside parentheses:
# Break the call across multiple lines inside parentheses
mlflow.log_params({
    "C": params["C"],
    "max_iter": params["max_iter"],
})
Option 2 — suppress the warning on a specific line where breaking would hurt readability:
mlflow.set_tracking_uri("http://localhost:5000")  # noqa: E501
Run flake8 . locally before every push — it runs in under a second and shows you exactly which lines to fix.
evaluate.py calls client.get_experiment_by_name(experiment_name) to find the runs from training. This returns None if no experiment with that name exists — which happens when train.py uses a different experiment name or never calls mlflow.set_experiment() at all and falls back to the default experiment (ID "0").To fix this, ensure the experiment name is consistent between train.py and evaluate.py:
# In train.py — set a named experiment before starting runs
mlflow.set_experiment("spotify-genre-classifier")

# In evaluate.py — search using the same name
experiment = client.get_experiment_by_name("spotify-genre-classifier")
If you intentionally use the default experiment, update evaluate.py to fall back to client.get_experiment("0") when the named lookup returns None.
If CI is failing but you cannot reproduce the error locally, paste the full GitHub Actions log into your PR description when asking for help. Sharing only the final error line makes it much harder for instructors to diagnose the root cause.

Build docs developers (and LLMs) love

Get started for freeTalk to us