DVC and MLflow Solve Different Problems

Every MLOps listicle names both, usually in the same bullet. They address genuinely different failures.

DVC answers: what data produced this?

Git is excellent for code and bad for a 4 GB parquet file. DVC keeps a small pointer file in Git while the real artifact lives in remote storage.

dvc add data/raw/train.parquet
git add data/raw/train.parquet.dvc

Checking out an old commit now retrieves the exact dataset that commit was trained against. Without it, "reproduce last month's result" is a request you cannot honour.

It also defines the pipeline as a DAG, so changing the cleaning stage re-runs everything downstream and nothing upstream.

MLflow answers: which run was that?

MLflow logs parameters, metrics and artifacts per run.

with mlflow.start_run():
    mlflow.log_param("learning_rate", 3e-4)
    mlflow.log_metric("val_kappa", 0.82)

The value is not visible on day one. It shows up three weeks later when you need the hyperparameters behind one good number, and the alternative is scrolling shell history.

The split

They overlap slightly and complement each other more than they compete.