datalad – skill Dữ liệu & Khoa học cho agent AI

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.

A normal clone retrieves Git history and the top-level file listing while leaving

annexed bytes unfetched. Installed subdatasets have their own histories; a clone does not

automatically populate them. Clone cost depends on Git history and file count, not just

the data volume. Retrieve annexed bytes selectively 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==1.6.5"
uv pip install "datalad-container==1.2.6"   # only for containers-run

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

The PyPI git-annex package supplies platform-specific binaries. The reviewed

10.20260901.post1 wheels cover Linux glibc 2.34+ (x86_64/ARM64), macOS ARM64 14+ and

x86_64 15+, and Windows x86_64. Use a system package when no wheel matches. Keep its

environment on PATH and verify the executable; the wheel does not supply Git itself.

Configure Git author name/email before creating or saving a dataset.

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 is MIT licensed; git-annex has a separate AGPL license. Consult the upstream

license when redistributing either tool.

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.

Illustrative remote-data example; inspect the selected revision for the exact path and

install NiBabel before the Python read. The refresh tested equivalent local pointer/get

behavior without downloading imaging data.

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. These symptoms can indicate an unfetched annex pointer; confirm with annex status

before diagnosing corruption. **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 availability checks which 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](references/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:

Illustrative FSL example (requires bet and an existing derivatives/ directory):

datalad run -m "extract brain and mask" \
  --input "sub-01/anat/sub-01_T1w.nii.gz" \
  --output "derivatives/sub-01_brain.nii.gz" \
  --output "derivatives/sub-01_brain_mask.nii.gz" \
  "bet {inputs[0]} {outputs[0]} -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 save only declared outputs. This does not capture unsaved input changes;

save all dependencies before claiming the run is reproducible. 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.

run records the command and dataset state; it does not freeze arbitrary host-installed software or external services. Version an environment lockfile and scripts as declared inputs, or use a tracked container image with containers-run. Record random seeds and relevant runtime settings, then test rerun from a fresh environment before claiming computational reproducibility.

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

--report only inspects the plan; it does not execute or validate the result. A branch

(-b) preserves the original commits, but uses the same worktree. See the reference for

a --since/--onto replay that starts before the first run, and compare annex keys or

content checksums as well as scientific outputs.

Containers

With the datalad-container extension, register an image once and every subsequent run

records which image produced the outputs:

Illustrative container workflow using a previously built local SIF image (not executed

in this refresh; the runtime and image must be available):

datalad containers-add fsl --url /path/to/fsl.sif \
  --call-fmt 'apptainer exec {img} {cmd}'
datalad containers-run -n fsl -m "brain and mask in container" \
  --input "sub-01/anat/sub-01_T1w.nii.gz" \
  --output "derivatives/sub-01_brain.nii.gz" \
  --output "derivatives/sub-01_brain_mask.nii.gz" \
  "bet {inputs[0]} {outputs[0]} -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](references/provenance.md) for the STAMPED principles and the YODA

project layout, the run record format, --explicit and --assume-ready semantics, and

exporting provenance toward W

… (nội dung đầy đủ có trong gói tải về)