What must you track to make an ML experiment reproducible?

Reproducibility requires capturing everything that influences the result. Track: the code version (git commit), the exact dataset version or snapshot (via hashing or tools like DVC or LakeFS), all hyperparameters and config, the training environment (library versions, container image, hardware/GPU type), and random seeds. Also log the resulting metrics, the model artifact, and its lineage linking model to code and data. Nondeterminism is a common pitfall: GPU operations, parallel data loading, and some library kernels can produce different results even with fixed seeds, so you may need deterministic flags and to accept small tolerances. Tools like MLflow, Weights & Biases, and Neptune automate much of this tracking. Good reproducibility lets you rerun an experiment, compare runs fairly, audit a production model, and debug regressions by isolating which change caused a metric shift.