Tornare a una versione precedente del dataset

Sometimes a new dataset hurts your model, or you need the exact data from last month's experiment. DVC lets you travel back to any past version.

Each data version is captured by a committed .dvc file. So your data history lives right inside your normal Git commit history.

First use git checkout to bring back the old .dvc pointer. This rewinds which data version your project expects.

The pointer changed but the file on disk has not yet. Run dvc checkout to update the actual dataset to match the restored pointer.

Rolling back data is always two moves: git checkout changes the pointer, then dvc checkout swaps the data. Skip the second and your data stays stale.

If that old version is not in your local cache, dvc checkout cannot find it. Run dvc pull first to fetch it from the remote, then check out.

To restore an entire past state, checkout a full commit, then dvc checkout. Code and data jump back to that moment together. ⏪

Give meaningful snapshots a Git tag so they are easy to return to later. Tags turn a cryptic hash into a memorable label.

Done experimenting in the past? Switch your branch forward again and run dvc checkout to restore the newest data. Nothing is lost.

Because any commit maps to an exact dataset, you can reproduce any past result precisely. This is the whole promise of data versioning.

Want one old file without changing branches? dvc get downloads a specific version from a repo into a chosen path, leaving your project untouched.

You ran git checkout on an old .dvc file. What must you run next to restore the actual data?

Rolling back is a two-step dance: git checkout restores the .dvc pointer, then dvc checkout swaps the data to match. Pull first if the old version is not cached.