8.3. Overview: Publishing datasets

The sections YODA-compliant data analysis projects, Beyond shared infrastructure, and Dataset hosting on GIN have each shown you crucial aspects of the functions of dataset publishing with datalad push. This section wraps them all together.

push availability

datalad push requires DataLad version 0.13.0 or higher. Older DataLad versions need to use the datalad publish command. For details into datalad publish, please check out the find-out-more on the difference between the two commands at the end of this page.

8.3.1. The general overview

datalad push is the command to turn to when you want to publish datasets. It is capable of publishing all dataset content, i.e., files stored in Git, and files stored with git-annex, to a known dataset sibling.

Push internals

The datalad push uses git push, and git annex copy under the hood. Publication targets need to either be configured remote Git repositories, or git-annex special remotes (if they support data upload).

In order to publish a dataset, the dataset needs to have a sibling to push to. This, for instance, can be a GitHub, GitLab, or Gin repository, but it can also be a Remote Indexed Archive (RIA) store for backup or storage of datasets1, or a regular clone.

all of the ways to configure siblings

  • Add an existing repository as a sibling with the datalad siblings command. Here are common examples:

    # to a remote repository
    $ datalad siblings add --name github-repo --url <url.to.github>
    # to a local path
    $ datalad siblings add --name local-sibling --url /path/to/sibling/ds
    # to a clone on an SSH-accessible machine
    $ datalad siblings add --name server-sibling --url [user@]hostname:/path/to/sibling/ds
    
  • Create a sibling on an external hosting service from scratch, right from within your repository: This can be done with the commands create-sibling-github (for GitHub) or create-siblings-gitlab (for GitLab), or create-sibling-ria (for a remote indexed archive dataset store1). Note that create-sibling-ria can add an existing store as a sibling or create a new one from scratch.

  • Create a sibling on a local or SSH accessible Unix machine with datalad create-sibling (datalad-create-sibling manual).

In order to publish dataset content, DataLad needs to know to which sibling content shall be pushed. This can be specified with the --to option directly from the command line:

$ datalad push --to <sibling>

If you have more than one branch in your dataset, note that a datalad push command will by default update all branches that both the sibling and the dataset share. If such advanced aspects of pushing are relevant for your workflow, please check out the find-out-more about this.

By default, push will make the last saved state of the dataset available. Consequently, if the sibling is in the same state as the dataset, no push is attempted. Additionally, push will attempt to automatically decide what type of dataset contents are going to be published. With a sibling that has a special remote configured as a publication dependency, or a sibling that contains an annex (such as a Gin repository or a Remote Indexed Archive (RIA) store), both the contents stored in Git (i.e., a dataset’s history) as well as file contents stored in git-annex will be published unless dataset configurations overrule this. Alternatively, one can enforce particular operations or push a subset of dataset contents. For one, when specifying a path in the datalad push command, only data or changes for those paths are considered for a push. Additionally, one can select a particular mode of operation with the -data option. Several different modes are possible:

  • nothing: With this option, annexed contents are not published. This means that the sibling will have information on the annexed files’ names, but file contents will not be available, and thus datalad get calls in the sibling would fail.

  • anything: Transfer all annexed contents.

  • auto: With this option, the decision which data is transferred is based on configurations that can determine rules on a per-file and per-sibling level. On a technical level, the git annex copy call to publish file contents is called with its --auto option. With this option, only data that satisfies specific git-annex configurations gets transferred. Those configurations could be numcopies settings (the number of copies available at different remotes), or wanted settings (preferred contents for a specific remote), and need to be created by a user2 with git-annex commands. If you have files you want to keep private, or do not need published, these configurations are very useful.

  • auto-if-wanted (Default): Unless a wanted or numcopies configuration exists in the dataset, all content are published. Should a wanted or numcopies configuration exist, the command enables --auto in the underlying git annex copy call.

Beyond different modes of transferring data, the -f/--force option allows to force specific publishing operations with three different modes. Be careful when using it, as its modes possibly overrule safety protections or optimizations:

  • checkdatapresent: With this option, the underlying git annex copy call to publish file contents is invoked without a --fast option. Usually, the --fast option increases the speed of the operation, as it disables a check whether the sibling already has content. This however, might skip copying content in some cases. Therefore, --force datatransfer is a slower, but more fail-safe option to publish annexed file contents.

  • gitpush: This option triggers a git push --force. Be very careful using this option - it will push all branches that are known to the sibling, and if the changes on these branches are conflicting with the changes that exist in the sibling, the changes that exist in the sibling will be overwritten.

  • all: The final mode, all, combines all force modes – thus attempting to really get your dataset contents published by any means.

datalad push can publish available subdatasets recursively if the -r/--recursive flag is specified. Note that this requires that all subdatasets that should be published have sibling names identical to the sibling specified in the top-level push command, or that appropriate default publication targets are configured throughout the dataset hierarchy.

Pushing more than the current branch

If you have more than one branch in your dataset, a datalad push --to <sibling> will by default only push the current branch, unless you provide configurations that alter this default. Here are two ways in which this can be achieved:

Option 1: Setting the push.default configuration variable from simple (the default) to matching will configure the dataset such that push pushes all branches to the sibling. A concrete example: On a dataset level, this can be done using

$ git config --local push.default matching

Option 2: Tweaking the default push refspec for the dataset allows to select a range of branches that should be pushed. The link above gives a thorough introduction into the refspec. For a hands-on example, consider how it is done for the published DataLad-101 dataset:

The published version of the handbook is known to the local handbook dataset as a remote called public, and each section of the book is identified with a custom branch name that corresponds to the section name. Whenever an update to the public dataset is pushed, apart from pushing only the master branch, all branches starting with the section identifier sct are pushed automatically as well. This configuration was achieved by specifying these branches (using globbing with *) in the push specification of this remote:

$ git config --local remote.public.push 'refs/heads/sct*'

8.3.2. Pushing errors

If you are unfamiliar with Git, please be aware that cloning a dataset to a different place and subsequently pushing to it can lead to Git error messages if changes are pushed to a currently checked out branch of the sibling (in technical Git terms: When pushing to a checked-out branch of a non-bare repository remote). As an example, consider what happens if we attempt a datalad push to the sibling roommate that we created in the chapter Collaboration:

$ datalad push --to roommate
[INFO] Determine push target
[INFO] Push refspecs
[INFO] Transfer data
copy(ok): books/TLCL.pdf (file) [to roommate...]
copy(ok): books/bash_guide.pdf (file) [to roommate...]
copy(ok): books/byte-of-python.pdf (file) [to roommate...]
[INFO] Update availability information
[INFO] Start enumerating objects 
[INFO] Start counting objects 
[INFO] Start compressing objects 
[INFO] Start writing objects 
[INFO] Start resolving deltas
[INFO] Finished
publish(ok): . (dataset) [refs/heads/git-annex->roommate:refs/heads/git-annex 6b28a11..e7bb084]
[ERROR] refs/heads/master->roommate:refs/heads/master [remote rejected] (branch is currently checked out) [publish(/home/me/dl-101/DataLad-101)] 
publish(error): . (dataset) [refs/heads/master->roommate:refs/heads/master [remote rejected] (branch is currently checked out)]
[INFO] Finished push of Dataset(/home/me/dl-101/DataLad-101)

Publishing fails with the error message [remote rejected] (branch is currently checked out). This can be prevented with configuration settings in Git versions 2.3 or higher, or by pushing to a branch of the sibling that is currently not checked-out. For more information on this, and other error messages during push, please checkout the section How to get help.

8.3.3. Setting access control via publishing

There are a number of ways to restrict access to your dataset or individual files of your dataset. One is via choice of (third party) hosting service for annexed file contents. If you chose a service only selected people have access to, and publish annexed contents exclusively there, then only those selected people can perform a successful datalad get. On shared file systems you may achieve this via permissions for certain groups or users, and for third party infrastructure you may achieve this by invitations/permissions/… options of the respective service.

If it is individual files that you do not want to share, you can selectively publish the contents of all files you want others to have, and withhold the data of the files you do not want to share. This can be done by publishing only selected files by providing paths, or overriding default push behavior with the --data option and git-annex wanted configurations. Specifying --data nothing would for example not push any annexed contents.

Let’s say you have a dataset with three files:

  • experiment.txt

  • subject_1.dat

  • subject_2.dat

Consider that all of these files are annexed. While the information in experiment.txt is fine for everyone to see, subject_1.dat and subject_2.dat contain personal and potentially identifying data that can not be shared. Nevertheless, you want collaborators (with whom data can not be shared) to know that these files exist, for example such that they can develop scripts against the directory structure and file names of a dataset, submit those scripts to the data owners, and thus still perform an analysis despite not having access to the data.

By publishing only the file contents of experiment.txt with

$ datalad push --to github experiment.txt

only meta data about file availability of subject_1.dat and subject_2.dat exists, but as these files’ annexed data is not published, a datalad get will fail. Note, though, that push will publish the complete dataset history (unless you specify a commit range with the --since option – see the manual for more information).

On the datalad publish command

Starting with DataLad version 0.13.0, datalad push was introduced and became an alternative to datalad publish, which will be removed in a future DataLad release.

By default, datalad publish publishes the last saved state of the dataset (i.e., its Git history) to a specified sibling:

$ datalad publish --to <sibling>

Like push, it supports recursive publishing across dataset hierarchies (if all datasets have appropriately configured default publication targets or identical sibling names) with the -r/--recursive flag, and it supports the --since option.

Main differences to push lie in publishs --transfer-data option that can be specified with either all, auto or none and determines whether and how annexed contents should be published if the sibling carries an annex: none will transfer only Git history and no annexed data, auto relies on configurations of the sibling, and all will publish all annexed contents.

By default, when using a plain datalad publish --to <sibling> with no path specification or --transfer-data option, publish will be used in auto mode. In practice, this default will most likely lead to the same outcome as when specifying none: only your datasets history, but no annexed contents will be published. On a technical level, the auto option leads to adding auto to the underlying git annex copy command, which in turn publishes annexed contents based on the git-annex preferred content configuration of the sibling.

In order to publish all annexed contents, one needs to specify --transfer-data all. Alternatively, adding paths to the publish call will publish the specified annexed content (unless --transfer-data none is explicitly added). As yet another alternative, one needs to add the same configuration for git-annex that the option --auto of push need.

Footnotes

1(1,2)

RIA siblings are filesystem-based, scalable storage solutions for DataLad datasets. You can find out more about them in the section Remote Indexed Archives for dataset storage and backup.

2

For information on the numcopies and wanted settings of git-annex see its documentation at git-annex.branchable.com/git-annex-wanted/ and git-annex.branchable.com/git-annex-numcopies/.