Edit on GitHub

Upgrading from DVC 2.x to 3.0

DVC 3.0 introduced changes to how DVC hashes files and to where DVC-tracked data is stored in the cache. DVC 3.0 remains compatible with pre-existing data tracked by DVC 2.0, but there are a few important points that users should note when upgrading to DVC 3.0.

For a full list of breaking changes in DVC 3.0, please refer to the release notes.

File hashing changes

Previously, DVC would attempt to identify whether a DVC-tracked file contained text content, and would convert Windows-style CRLF line endings to Unix-style LF line endings before hashing the file content (i.e. a dos2unix conversion). This behavior was intended to simplify usage in cross-platform scenarios (where a DVC repository was used on both Unix and Windows machines). However, even though DVC would convert line endings when computing hashes, DVC would still store the original native content in both local DVC cache and in remote storage. This would lead to unintended side effects in situations where a given file was a binary file misidentified as text by DVC or where a text file was not intended to be cross platform (and CRLF should not have been considered equivalent to LF).

In DVC 3.0, the line ending conversion behavior has been removed, and DVC treats all files as if they contain binary data. This means that a text file with CRLF line endings will always be identified as completely separate from a file containing LF line endings, even if all other text content in the two files is identical.

When upgrading to DVC 3.0, users with pipelines that may be run in both Unix and Windows environments should ensure that any pipeline stages with text outputs (such as .csv or .tsv files) generate files with consistent line endings, regardless of the platform where a stage is run.

For example, Python stages should explicitly generate files with either Unix-style \n or Windows-style \r\n line endings, rather than relying on the default platform specific os.linesep behavior.

Optional local cache migration

In order to avoid hash collisions between files tracked in DVC 3.0 and older releases, files tracked in DVC 3.0 are stored separately from files tracked in older releases. By default, DVC does not automatically de-duplicate any data between files tracked in DVC 3.0 and files tracked in older releases. DVC will still read cached files from DVC 2.0 and will only duplicate for new or modified data.

Users can manually migrate existing local DVC cache data to the DVC 3.0 location by running the dvc cache migrate command. On most local filesystems, dvc cache migrate is equivalent to forcing the de-duplication of files tracked in DVC 3.0 and files tracked in older releases. Files from the old cache location will be re-hashed using the DVC 3.0 hash algorithm, atomically moved to the new cache location, and then a link will be created from the old location to the new one. This may take a long time.

On filesystems that do not support any type of linking, data will be copied from the old cache location into the DVC 3.0 location (resulting in no de-duplication).

By default, dvc cache migrate only migrates cache data and does not modify DVC files in the DVC repository. dvc cache migrate --dvc-files will migrate entries in all DVC files in the repository so that DVC will only use data from the DVC 3.0 cache location.

Note that when using --dvc-files option, DVC will only migrate DVC files in workspace (and Git history will not be re-written).

For DVC remotes, there is no equivalent migration command since it is not possible to link between old and new locations on many remote filesystems. Instead, once you have migrated data locally and pushed to the remote, you may use dvc gc -c commands to remove outdated data from the remote.


๐Ÿ› Found an issue? Let us know! Or fix it:

Edit on GitHub

โ“ Have a question? Join our chat, we will help you:

Discord Chat