Datalad

DataLad

Overview

DataLad is a data management layer over Git and git-annex. Git tracks the dataset structure, small text files, and the history. git-annex tracks the content of large files, storing each file as a key and keeping the bytes somewhere that is not necessarily the local repository.

That split is the single most important thing to internalise, because it means a freshly cloned dataset contains the full history and the full file listing while containing almost none of the data. A 100 TB dataset clones in seconds and occupies a few megabytes. The bytes arrive only when asked for, per file, with datalad get.

The second thing DataLad adds is provenance. datalad run executes a command and commits the result together with a machine-readable record of the command, its inputs, and its outputs. datalad rerun reads that record back and re-executes it. This turns "how was this figure produced" from an archaeology problem into a command.

When to use DataLad instead of plain Git

Use DataLad when any of the following holds:

• Files are too large for Git to handle comfortably, or the total exceeds what every collaborator wants on disk.
• Data lives in more than one place (a lab server, a cluster scratch, S3, a supercomputer) and you need to know which copies exist.
• The analysis must be re-executable, and a plain commit message is not enough evidence.
• You are consuming published datasets from OpenNeuro, DANDI, or datasets.datalad.org, which are distributed as DataLad datasets.
• The project nests other datasets inside it and you want each one to keep its own independent history.

Use plain Git when the repository is code and text only, everything fits comfortably in Git, and nobody needs partial checkouts. DataLad on top of a small pure-code repository adds indirection without buying anything.

Installation

# git-annex is NOT written in Python but is available from PyPI if you already
# have git itself installed:
uv pip install git-annex
# You can also install it first from the system
# (Debian/Ubuntu: apt install git-annex; macOS: brew install git-annex;
#  conda-forge: conda install -c conda-forge git-annex)
uv pip install datalad
uv pip install datalad-container   # only for containers-run

datalad wtf --section dependencies   # confirm git-annex version is visible

The PyPI git-annex package ships the prebuilt binary as a wheel for Linux, macOS, and Windows rather than building the Haskell sources, so it installs like any other Python dependency and can be pinned in the same environment as DataLad. It does not bring git along with it.

datalad wtf prints the resolved environment and is the first thing to run when behaviour looks impossible. An old or missing git-annex is behind a large share of confusing errors.

DataLad itself is MIT licensed. git-annex is a separate tool under the AGPL, which matters only if you redistribute a modified git-annex rather than call it.

The failure that bites first: pointers are not data

After datalad clone, annexed files exist as symlinks into .git/annex/objects/ (or as small pointer files where symlinks are unavailable, such as on Windows or a crippled filesystem). Nothing has downloaded the content yet.

datalad clone https://github.com/OpenNeuroDatasets/ds000001.git
cd ds000001
ls sub-01/anat/            # the file is listed
python -c "import nibabel; nibabel.load('sub-01/anat/sub-01_T1w.nii.gz')"   # fails
datalad get sub-01/anat/sub-01_T1w.nii.gz                                   # now it works

The failure mode to recognise: a tool reports the file as empty, truncated, corrupt, "not a gzip file", or a broken symlink, and the file size on disk is a few hundred bytes. That is a pointer, not a corrupted download. Run datalad get before reading data, and treat "file exists" as insufficient evidence that its content is present.

Before an analysis touches a directory, fetch it explicitly:

datalad get sub-01/                  # everything under a path
datalad get -r .                     # everything, including subdatasets
datalad get -n -r .                  # subdataset structure only, no file content

datalad status --annex reports how much content is present locally, and git annex whereis <path> reports which repositories hold a given file. whereis reads recorded state and does not contact the remotes, so it tells you what git-annex last learned rather than what is true right now.

See data-access.md for finding datasets, subdataset behaviour, dropping content safely, and repairing a dataset.

Recording provenance with datalad run

datalad run is the reason to reach for DataLad in a methods context. It saves the command alongside its effect, in the same commit:

datalad run -m "extract brain mask" \
  --input "sub-01/anat/sub-01_T1w.nii.gz" \
  --output "derivatives/sub-01_brain.nii.gz" \
  "bet {inputs} {outputs} -m"

What each part does, and why skipping it hurts:

• --input retrieves the content before running, so the command does not fail on a pointer. It also records the dependency, which is what lets rerun fetch the same inputs on a different machine.
• --output unlocks or removes the target first, so git-annex does not refuse to write over content it is protecting. Without it, a second run of the same command commonly fails with a permission error on an annexed file that looks read-only.
• {inputs} and {outputs} expand to those values. {pwd}, {dspath}, and {tmpdir} are also available, and {inputs[0]} indexes individual entries.
• The commit message carries a JSON run record between === Do not change lines below === and ^^^ Do not change lines above ^^^. Do not hand-edit that block; rerun parses it.

datalad run refuses to start when the dataset has unsaved modifications, because an unclean starting state makes the record unreliable. Save or discard first, or pass --explicit to declare that the listed inputs and outputs are the complete story. Check a command before committing to it with --dry-run basic or --dry-run command.

A run that changes nothing produces no commit, exactly as datalad save does.

Re-executing

datalad rerun                       # redo the run recorded at HEAD
datalad rerun --report              # show what would be done, change nothing
datalad rerun --script recompute.sh # extract the commands instead of running them
datalad rerun --since <commit> -b check <revision>   # replay a range onto a new branch

Rerunning onto a branch (-b) is the safe way to test reproducibility: the replay lands somewhere else, and a diff against the original branch answers whether the outputs came back identical.

Containers

With the datalad-container extension, register an image once and every subsequent run records which image produced the outputs:

datalad containers-add fsl --url docker://brainlife/fsl:6.0.4
datalad containers-run -n fsl -m "brain mask in container" \
  --input "sub-01/anat/sub-01_T1w.nii.gz" \
  --output "derivatives/sub-01_brain.nii.gz" \
  "bet {inputs} {outputs} -m"

The image itself is tracked in the dataset, so the software environment travels with the data and the provenance record rather than living in someone's shell history. When only one container is configured, -n may be omitted.

See provenance.md for the STAMPED principles and the YODA project layout, the run record format, --explicit and --assume-ready semantics, and exporting provenance toward W3C PROV.

Saving and inspecting changes

datalad status                 # what changed, including subdataset state
datalad save -m "add QC report" path/to/file
datalad save -m "checkpoint" -r                 # recurse into subdatasets
datalad save -m "small text file" --to-git notes.md

datalad save decides per file whether content goes to Git or to git-annex, following the dataset's .gitattributes. Force a file into Git with --to-git, which is the right call for code and small text files that should stay directly readable. The yoda procedure (datalad create -c yoda) sets this up for code/, README.md, and CHANGELOG.md automatically.

Creating a dataset

datalad create my_dataset               # plain dataset
datalad create -c yoda my_analysis      # analysis layout (code/ tracked in Git,
                                        # README.md and CHANGELOG.md preconfigured)
datalad create -d . inputs/raw          # register a new subdataset under an existing one

-c yoda applies the analysis project layout described in provenance.md. -d . is what registers a new dataset as a subdataset of the parent rather than leaving an unrelated repository inside it.

Publishing

A DataLad dataset is usually published to two places at once: a Git hosting service for the history, and a storage remote for the annexed content.

datalad create-sibling-github myaccount/mydataset
git annex initremote store type=S3 bucket=my-bucket encryption=none autoenable=true
datalad siblings configure -s github --publish-depends store
datalad push --to github

The Git sibling and the storage sibling are created by different tools on purpose. A Git sibling is a Git remote, and datalad create-sibling-* handles the hosting-service ones. An S3 bucket (or WebDAV, or an SSH directory) is a git-annex special remote, not a Git remote, so it is created with git annex initremote. datalad siblings picks the special remote up afterwards and treats it like any other. Using datalad siblings add --url s3://... here is the mistake this section exists to prevent: --url is a Git remote URL, S3 is not, and the push --to github below then fails on the --publish-depends hop.

--publish-depends is what stops the common broken publication: a Git repository whose history references content that was never uploaded, so collaborators clone successfully and then find every datalad get failing. Declaring the dependency makes the storage sibling publish first, every time.

datalad push sends both the Git history and, by default (--data auto-if-wanted), the annexed content the target is configured to want. Pass --data anything to push all content regardless of the target's preferences.

See publishing.md for RIA stores, special remotes, credential handling, and configuring which sibling holds what.

Freeing disk space

git annex whereis sub-01/                 # confirm another copy exists first
datalad drop sub-01/                      # remove local content, keep the pointer
datalad drop --what all --reckless kill <path>   # last resort, destroys data

datalad drop refuses by default when it cannot verify another copy of the content exists, which is a safety check rather than an obstacle. --nocheck and --if-dirty are deprecated; the current spelling is --reckless availability, and it means what it says. --what selects between filecontent (the default), allkeys, datasets, and all.

Failure modes worth knowing

Detailed references

• data-access.md: finding published datasets (registry.datalad.org, OpenNeuro, DANDI, datasets.datalad.org and the /// shortcut), clone and get options, subdataset handling, annex content states, dropping and removing, and fsck repair.
• provenance.md: the STAMPED principles and the YODA layout, the run record format, run and rerun options in full, containers-run, and the current state of exporting DataLad provenance toward W3C PROV.
• publishing.md: siblings and their actions, create-sibling-* variants, RIA stores, special remotes, push semantics, and credential handling.

Related skills

The bids skill covers the Brain Imaging Data Structure that most of the neuroimaging datasets distributed through DataLad are organised in. A typical workflow clones a BIDS dataset with DataLad, validates it with the BIDS tooling, then runs a BIDS-App under datalad containers-run so the derivatives carry provenance.

Primary sources

• DataLad documentation: https://docs.datalad.org/en/stable/
• DataLad Handbook: https://handbook.datalad.org/en/latest/
• datalad run chapter: https://handbook.datalad.org/en/latest/basics/101-108-run.html
• YODA principles: https://handbook.datalad.org/en/latest/basics/101-127-yoda.html
• STAMPED principles (operationalized from YODA): https://stamped-principles.org
• datalad-container: https://docs.datalad.org/projects/container/en/stable/
• git-annex: https://git-annex.branchable.com/
• Dataset registry: https://registry.datalad.org

Acknowledgment

Topic scope for this skill was informed in part by @bcmcpher's MIT-licensed datalad-cli plugin (nineteen per-command slash-command skills). The text here is written independently and grounded in the upstream DataLad documentation; overlap is unavoidable because both cover DataLad, but the structure, style, and specific technical claims are different.