- Data Discovery
- Data Consumption
- Data Sharing
- For Git Users
- Reproducible Science
- Automation and Control of Data Acquisition
With the growing quantities and variety of data, it is getting progressively harder to determine whether and where given data are available. DataLad aids data discovery through a built-in, extensible system for automated metadata extraction and homogenization.
Being a tool for decentralized data management, DataLad does not require a central service to perform metadata queries. Instead, any DataLad installation provides all the necessary functionality to automatically index and search any data it was made aware of — whether it's one's personal data, the data of an entire lab, all the data of an institution, or a complete public data portal.
The following demo shows how one can quickly find desired datasets, install them, and selectively obtain necessary files across multiple datasets.
For Data Consumers
Data are key to many applications, and while an enormous amount of data is available, it is all-too-often impossible for interested parties to consume the relevant pieces efficiently, become aware when any part of it changes, and why it changed.
For a growing number of data portals, the DataLad team provides native DataLad datasets that offer fine-grained access to current and historic version of many datasets.
However, the vast majority of all available data are scattered around the web using an incomprehensible variety of sites, access methods, and formats. In order to help mitigate this complexity, DataLad provides tools to monitor non-DataLad data sources on the web. With these tools, DataLad users are able to track their changes over time, curate them, and even re-publish them as native DataLad datasets.
The following demo shows how one can create a readily usable dataset from data and code posted on a textbook's companion website.
Sharing data, code, and results is a key element of the scientific process. Sharing comes in many flavors, for example:
- sharing with yourself, and having certainty that the copy on your laptop matches what you have been working with on your workstation;
- sharing with colleagues, and making sure they all know when data changed and which results have to be recomputed; and
- sharing with the world, to make sure that your work has maximum impact.
DataLad aids many types of sharing efforts. It supports synchronization of multiple instances of a dataset belonging to a single person. It provides on-demand updates of datasets shared with local or remote collaborators. It offers a wide variety of publications methods, ranging from a (cloud) server, to services such as GitHub, DropBox, or box.com.
Here is a demo of using DataLad with a cloud storage service. This combination allows for convenient data exchange between colleagues, or simple data synchronization between the machines of a single person.
Most cloud storage services are not ideal for data sharing with the general public, because they require a file or directory to be shared with particular accounts of that service provider. An SSH-accessible web server does not have these limitation, and is equally well supported by DataLad. Here is a demo that shows how one can publish large data "on GitHub" for maximum visibility, and configure DataLad to transparently obtain all data files from a different web server.
For Git and git-annex Users
DataLad is built atop of Git and git-annex, and if you are familiar with these tools, you should feel at home with DataLad too. DataLad aims to be a unifying top-level layer that makes working with Git and git-annex easier — without getting in the way of power-users.
DataLad's most unique feature, compared to Git and git-annex, is its ability to seamlessly work with an entire tree of nested repositories (i.e. Git submodules). This is best explained in a brief demo:
The following table helps to illustrate which features are contributed by which technology layer.
|Version control (text, code)|
|Version control (binary data)||(not advised)|
|Auto-crawling available resources||RSS feeds||flexible|
|Unified dataset handling|
|Unified authentication interface|
DataLad is an ideal tool for conducting reproducible science. It can track and obtain shared data and publish results. Importantly, it jointly manages both analysis input data and the associated analysis code --- critical to reproduce any analysis. Lastly, DataLad is able to temporally capture the exact commands used to produce the results.
All together, DataLad can automatically create an extensive provenance record, with all the information necessary to rerun an analysis from beginning to end and demonstrate reproducibility.
Here is a quick demo of how DataLad captures the results derived from input data. Scroll further down for a more in-depth demo of a full analysis.
The next demo shows the complete setup, execution, and replication of a brain imaging analysis on real data — in just five minutes. You'll see how easy it is to automatically yield a reproducible outcome. It also demonstrates how DataLad helps to maintain a strict separation of inputs and outputs to clearly identify what needs to be archived alongside a publication.
Automation and Control of Data Acquisition
In MRI research, data typically comes as a set of DICOM files which need to be first converted to a data format convenient for visualization and analysis, typically NIfTI. Another step further is to layout and organize the data according to BIDS (Brain Imaging Data Structure). One of the tools available to assist with such conversions is HeuDiConv (Heuristic DICOM Converter), which also has an option to place converted data and pre-generated templates under DataLad's control. This makes it immediately possible to distribute collected data across processing infrastructure, track provenance of derived data, and also updating datasets with more of freshly acquired data while relying on git's powerful merge mechanisms.
Whenever the data are ready for public sharing, it is a datalad publish away, while also allowing to to easily control and restrict the public release to only data files which do not carry any possibly subject identifying information (e.g., non-defaced high-resolution anatomicals).