Greg Wilson1, Jennifer Bryan2, Karen Cranston3, Justin Kitzes4, Lex Nederbragt5, Tracy K. Teal6
1) Software Carpentry Foundation / firstname.lastname@example.org
2) University of British Columbia / email@example.com
3) Duke University / firstname.lastname@example.org
4) University of California, Berkeley / email@example.com
5) University of Oslo / firstname.lastname@example.org
6) Data Carpentry / email@example.com
A PDF of this paper can be found at http://arxiv.org/abs/1609.00037. We would appreciate feedback of all kinds: please email comments to the authors.
We present a set of computing tools and techniques that every researcher can and should consider adopting. These recommendations synthesize inspiration from our own work, from the experiences of the thousands of people who have taken part in Software Carpentry and Data Carpentry workshops over the past six years, and from a variety of other guides. Our recommendations are aimed specifically at people who are new to research computing.
Computers are now essential in all branches of science, but most researchers are never taught the equivalent of basic lab skills for research computing. As a result, data can get lost, analyses can take much longer than necessary, and researchers are limited in how effectively and they can work with software and data. Computing workflows need to follow the same practices as lab projects and notebooks, with organized data, documented steps and the project structured for reproducibility, but researchers new to computing often don't know where to start.
This paper presents a set of good computing practices that every researcher can adopt regardless of their current level of computational skill. These practices, which encompass data management, programming, collaborating with colleagues, organizing projects, tracking work, and writing manuscripts, are drawn from a wide variety of published sources, from our daily lives, and from our work with volunteer organizations that have delivered workshops to over 11,000 people since 2010.
Three years ago a group of researchers involved in Software Carpentry1 and Data Carpentry2 wrote a paper called "Best Practices for Scientific Computing" [wilson2014]. That paper provided recommendations for people who were already doing significant amounts of computation in their research. However, as computing has become an essential part of science for all researchers, there is a larger group of people newer to scientific computing, and the question then becomes "where to start?".
This paper focuses on these first, accessible skills and perspectives - the "good enough" practices - for scientific computing, a minimum set of tools and techniques that we believe every researcher can and should consider adopting. It draws inspiration from many sources [gentzkow2014, noble2009, brown2015, wickham2014, kitzes2016, sandve2013, hart2016], from our personal experience, and from the experiences of the thousands of people who have taken part in Software Carpentry and Data Carpentry workshops over the past six years.
Our intended audience is researchers who are working alone or with a handful of collaborators on projects lasting a few days to several months. A practice is included in our list if large numbers of researchers use it, and large numbers of people are still using it months after first trying it out. We include the second criterion because there is no point recommending something that people won't actually adopt.
Many of our recommendations are for the benefit of the collaborator every researcher cares about most: their future self3. Change is hard and if researchers don't see those benefits quickly enough to justify the pain, they will almost certainly switch back to their old way of doing things. This rules out many practices, such as code review, that we feel are essential for larger-scale development (Section [sec:omitted]).
We organize our recommendations into the following topics:
Data Management: saving both raw and intermediate forms; documenting all steps; creating tidy data amenable to analysis.
Software: writing, organizing, and sharing scripts and programs used in an analysis.
Collaboration: making it easy for existing and new collaborators to understand and contribute to a project.
Project Organization: organizing the digital artifacts of a project to ease discovery and understanding.
Tracking Changes: recording how various components of your project change over time.
Manuscripts: writing manuscripts in a way that leaves an audit trail and minimizes manual merging of conflict.
We are grateful to Arjun Raj (University of Pennsylvania), Steven Haddock (Monterey Bay Aquarium Research Institute), Stephen Turner (University of Virginia), Elizabeth Wickes (University of Illinois), and Garrett Grolemund (RStudio) for their feedback on early versions of this paper, to those who contributed during the outlining of the manuscript, and to everyone involved in Data Carpentry and Software Carpentry for everything they have taught us.
Data within a project may need to exist in various forms, ranging from what first arrives to what is actually used for the primary analyses. Our recommendations have two main themes. One is to work towards ready-to-analyze data incrementally, documenting both the intermediate data and the process. We also describe the key features of "tidy data", which can be a powerful accelerator for analysis [wickham2014, hart2016].
Save the raw data (1a). Where possible, save data as originally generated (i.e. by an instrument or from a survey). It is tempting to overwrite raw data files with cleaned-up versions, but faithful retention is essential for re-running analyses from start to finish; for recovery from analytical mishaps; and for experimenting without fear. Consider changing file permissions to read-only or using spreadsheet protection features, so it is harder to damage raw data by accident or to hand edit it in a moment of weakness.
Some data will be impractical to manage in this way. For example, you should avoid making local copies of large, stable databases. In that case, record the exact procedure used to obtain the raw data, as well as any other pertinent information, such as an official version number or the date of download.
Ensure that raw data is backed up in more than one location. (1b) If external hard drives are used, store them off-site of the original location. Universities often have their own data storage solutions, so it is worthwhile to consult with your local Information Technology (IT) group or library. Alternatively cloud computing resources, like Amazon Simple Storage Service (Amazon S3)4, Google Cloud Storage5 or https://azure.microsoft.com/en-us/services/storage/ are reasonably priced and reliable. For large data sets, where storage and transfer can be expensive and time-consuming, you may need to use incremental backup or specialized storage systems, and people in your local IT group or library can often provide advice and assistance on options at your university or organization as well.
Create the data you wish to see in the world (1b). Create the dataset you wish you had received. The goal here is to improve machine and human readability, but not to do vigorous data filtering or add external information. Machine readability allows automatic processing using computer programs, which is important when others want to reuse your data. Specific examples of non-destructive transformations that we recommend at the beginning of analysis:
File formats: Convert data from closed, proprietary formats to open, non-proprietary formats that ensure machine readability across time and computing setups [UIllinois]. Good options include CSV for tabular data, JSON, YAML, or XML for non-tabular data such as graphs6, and HDF5 for certain kinds of structured data.
Variable names: Replace inscrutable variable names and artificial
data codes with self-explaining alternatives, e.g., rename variables
recode the treatment variable from
treated, and replace artificial codes for missing data, such as
NA, a code used in most programming languages to
indicate that data is "Not Available" [white2013].
Filenames: Store especially useful metadata as part of the
filename itself, while keeping the filename regular enough for easy
pattern matching. For example, a filename like
2016-05-alaska-b.csv makes it easy for both people and programs to
select by year or by location.
Create analysis-friendly data (1c): Analysis can be much easier if you are working with so-called "tidy" data [wickham2014]. Two key principles are:
Make each column a variable: Don't cram two variables into one, e.g., "male_treated" should be split into separate variables for sex and treatment status. Store units in their own variable or in metadata, e.g., "3.4" instead of "3.4kg".
Make each row an observation: Data often comes in a wide format, because that facilitated data entry or human inspection. Imagine one row per field site and then columns for measurements made at each of several time points. Be prepared to gather such columns into a variable of measurements, plus a new variable for time point. Figure [fig:tidy] presents an example of such a transformation.
Record all the steps used to process data (1d): Data manipulation is as integral to your analysis as statistical modeling and inference. If you do not document this step thoroughly, it is impossible for you, or anyone else, to repeat the analysis.
The best way to do this is to write scripts for every stage of data processing. This might feel frustratingly slow, but you will get faster with practice. The immediate payoff will be the ease with which you can re-do data preparation when new data arrives. You can also re-use data preparation steps in the future for related projects. For very large data sets, data preparation may also include writing and saving scripts to obtain the data or subsets of the data from remote storage.
Some data cleaning tools, such as OpenRefine7, provide a graphical user interface, but also automatically keep track of each step in the process. When tools like these or scripting is not feasible, it's important to clearly document every manual action (what menu was used, what column was copied and pasted, what link was clicked, etc.). Often you can at least capture what action was taken, if not the complete why. For example, choosing a region of interest in an image is inherently interactive, but you can save the region chosen as a set of boundary coordinates.
Anticipate the need to use multiple tables, and use a unique identifier for every record (1e): Raw data, even if tidy, is not necessarily complete. For example, the primary data table might hold the heart rate for individual subjects at rest and after a physical challenge, identified via a subject ID. Demographic variables, such as subject age and sex, are stored in a second table and will need to be brought in via merging or lookup. This will go more smoothly if subject ID is represented in a common format in both tables, e.g., always as "14025" versus "14,025" in one table and "014025" in another. It is generally wise to give each record or unit a unique, persistent key and to use the same names and codes when variables in two datasets refer to the same thing.
Submit data to a reputable DOI-issuing repository so that others can access and cite it. (1f) Your data is as much a product of your research as the papers you write, and just as likely to be useful to others (if not more so). Sites such as Figshare8, Dryad9, and Zenodo10 allow others to find your work, use it, and cite it; we discuss licensing in Section [sec:collaboration] below. Follow your research community's standards for how to provide metadata. Note that there are two types of metadata: metadata about the dataset as a whole and metadata about the content within the dataset. If the audience is humans, write the metadata (the README file) for humans. If the audience includes automatic metadata harvesters, fill out the formal metadata and write a good README file for the humans [wickes2015].
Taken in order, the recommendations above will produce intermediate data files with increasing levels of cleanliness and task-specificity. An alternative approach to data management would be to fold all data management tasks into a monolithic procedure for data analysis, so that intermediate data products are created "on the fly" and stored only in memory, not saved as distinct files.
While the latter approach may be appropriate for projects in which very little data cleaning or processing is needed, we recommend the explicit creation and retention of intermediate products. Saving intermediate files makes it easy to re-run parts of a data analysis pipeline, which in turn makes it less onerous to revisit and improve specific data processing tasks. Breaking a lengthy workflow into pieces makes it easier to understand, share, describe, and modify. This is particularly true when working with large data sets, where storage and transfer of the entire data set is not trivial or inexpensive.
If you or your group are creating tens of thousands of lines of software for use by hundreds of people you have never met, you are doing software engineering. If you're writing a few dozen lines now and again, and are probably going to be its only user, you may not be doing engineering, but you can still make things easier on yourself by adopting a few key engineering practices. What's more, adopting these practices will make it easier for people to understand and (re)use your code.
The core realization in these practices is that readable, reusable, and testable are all side effects of writing modular code, i.e., of building programs out of short, single-purpose functions with clearly-defined inputs and outputs [hunt1999]. Much has been written on this topic (refs!), and this section focuses on practices that best balance ease of use with benefit for you and collaborators.
Place a brief explanatory comment at the start of every program (2a), no matter how short it is. That comment should include at least one example of how the program is used: remember, a good example is worth a thousand words. Where possible, the comment should also indicate reasonable values for parameters. An example of such a comment is show below.
Synthesize image files for testing circularity estimation algorithm. Usage: make_images.py -f fuzzing -n flaws -o output -s seed -v -w size where: -f fuzzing = fuzzing range of blobs (typically 0.0-0.2) -n flaws = p(success) for geometric distribution of # flaws/sample (e.g. 0.5-0.8) -o output = name of output file -s seed = random number generator seed (large integer) -v = verbose -w size = image width/height in pixels (typically 480-800) -h = show help message
Decompose programs into functions (2b) that are no more than one page (about 60 lines) long. A function is a reusable section of software that can be treated as a black box by the rest of the program. The syntax for creating functions depends on programming language, but generally you name the function, list its input parameters, and describe what information it produces. Functions should take no more than five or six input parameters and should not reference outside information.
The key motivation here is to fit the program into the most limited memory of all: ours. Human short-term memory is famously incapable of holding more than about seven items at once [miller1956]. If we are to understand what our software is doing, we must break it into chunks that obey this limit, then create programs by combining these chunks. Putting code into functions also makes it easier to test and troubleshoot when things go wrong.
Be ruthless about eliminating
duplication (2c). Write and re-use
functions instead of copying and pasting code, and use data
structures like lists instead of creating many closely-related
variables, e.g. create
score = (1, 2, 3) rather than
Also look for well-maintained libraries that already do what you're trying to do. All programming languages have libraries that you can import and use in your code. This is code that people have already written and made available for distribution that have a particular function. For instances there are libraries for statistics, modeling, mapping and many more. Many languages catalog the libraries in a centralized source, for instance R has CRAN11, Python has PyPI12, and so on. So always search for well-maintained software libraries that do what you need (2d) before writing new code yourself, but test libraries before relying on them (2e).
Give functions and variables meaningful
names (2f), both to document their
purpose and to make the program easier to read. As a rule of thumb,
the greater the scope of a variable, the more informative its name
should be: while it's acceptable to call the counter variable in a
j, things that are re-used often, such as the major
data structures in a program should not have one-letter names.
Remember to follow each language's conventions for names, such as
net_charge for Python and
NetCharge for Java.
Almost all modern text editors provide tab completion, so that typing the first part of a variable name and then pressing the tab key inserts the completed name of the variable. Employing this means that meaningful longer variable names are no harder to type than terse abbreviations.
Make dependencies and requirements
explicit. (2g) This is usually done on
a per-project rather than per-program basis, i.e., by adding a file
called something like
requirements.txt to the root directory of
the project, or by adding a "Getting Started" section to the
Do not comment and uncomment sections of code to control a program's behavior (2h), since this is error prone and makes it difficult or impossible to automate analyses. Instead, put if/else statements in the program to control what it does.
Provide a simple example or test data set (2i) that users (including yourself) can run to determine whether the program is working and whether it gives a known correct output for a simple known input. Such a "build and smoke test" is particularly helpful when supposedly-innocent changes are being made to the program, or when it has to run on several different machines, e.g., the developer's laptop and the department's cluster.
Submit code to a reputable DOI-issuing repository (2j) upon submission of paper, just as you do with data. Your software is as much a product of your research as your papers, and should be as easy for people to credit. DOIs for software are provided by Figshare13 and Zenodo14. Zenodo integrates directly with GitHub.
You may start working on projects by yourself or with a small group of collaborators you already know, but you should design it to make it easy for new collaborators to join. These collaborators might be new grad students or postdocs in the lab, or they might be you returning to a project that has been idle for some time. As summarized in [steinmacher2015], you want to make it easy for people to set up a local workspace so that they can contribute, help them find tasks so that they know what to contribute, and make the contribution process clear so that they know how to contribute. You also want to make it easy for people to give you credit for your work.
Create an overview of your
project. (3a) Have a short file in the
project's home directory that explains the purpose of the project.
This file (generally called
README.txt, or something
similar) should contain the project's title, a brief description,
up-to-date contact information, and an example or two of how to run
various cleaning or analysis tasks. It is often the first thing
users and collaborators on your project will look at, so make it
explicit how you want people to engage with the project. If you are
looking for more contributors, make it explicit that you welcome
contributors and point them to the license (more below) and ways
they can help.
You should also create a
CONTRIBUTING file that describes what
people need to do in order to get the project going and use or
contribute to it, i.e., dependencies that need to be installed,
tests that can be run to ensure that software has been installed
correctly, and guidelines or checklists that your project adheres
Create a shared "to-do"
list (3b). This can be a plain text
file called something like
todo.txt, or you can use
sites such as GitHub or Bitbucket to create a new issue for each
to-do item. (You can even add labels such as "low hanging fruit" to
point newcomers at issues that are good starting points.) Whatever
you choose, describe the items clearly so that they make sense to
Decide on communication strategies. (3c) Make explicit decisions about (and publicize where appropriate) how members of the project will communicate with each other and with externals users / collaborators. This includes the location and technology for email lists, chat channels, voice / video conferencing, documentation, and meeting notes, as well as which of these channels will be public or private.
Make the license
explicit. (3d) Have a
in the project's home directory that clearly states what license(s)
apply to the project's software, data, and manuscripts. Lack of an
explicit license does not mean there isn't one; rather, it implies
the author is keeping all rights and others are not allowed to
re-use or modify the material.
We recommend Creative Commons licenses for data and text, either CC-015 (the "No Rights Reserved" license) or CC-BY16 (the "Attribution" license, which permits sharing and reuse but requires people to give appropriate credit to the creators). For software, we recommend a permissive open source license such as the MIT, BSD, or Apache license [laurent2004].
What Not To Do
We recommend against the "no commercial use" variations of the Creative Commons licenses because they may impede some forms of re-use. For example, if a researcher in a developing country is being paid by her government to compile a public health report, she will be unable to include your data if the license says "non-commercial". We recommend permissive software licenses rather than the GNU General Public License (GPL) because it is easier to integrate permissively-licensed software into other projects, see chapter three in [laurent2004].
Make the project
citable (3e) by including a
file in the project's home directory that describes how to cite this
project as a whole, and where to find (and how to cite) any data
sets, code, figures, and other artifacts that have their own DOIs.
The example below shows the
CITATION file for the
Ecodata Retriever17; for an example of
a more detailed
CITATION file, see the one for the
Please cite this work as: Morris, B.D. and E.P. White. 2013. "The EcoData Retriever: improving access to existing ecological data." PLOS ONE 8:e65848. http://doi.org/doi:10.1371/journal.pone.0065848
Organizing the files that make up a project in a logical and consistent directory structure will help you and others keep track of them. Our recommendations for doing this are drawn primarily from [noble2009, gentzkow2014].
Put each project in its own directory, which is named after the project. (4a) Like deciding when a chunk of code should be made a function, the ultimate goal of dividing research into distinct projects is to help you and others best understand your work. Some researchers create a separate project for each manuscript they are working on, while others group all research on a common theme, data set, or algorithm into a single project.
As a rule of thumb, divide work into projects based on the overlap in data and code files. If two research efforts share no data or code, they will probably be easiest to manage independently. If they share more than half of their data and code, they are probably best managed together, while if you are building tools that are used in several projects, the common code should probably be in a project of its own. Projects do often require their own organizational model, but the below are recommendations on how you can structure data, code, analysis outputs and other files. The important concept is that is useful to organize the project by the types of files and that consistency helps you effectively find and use things later.
Put text documents associated with the project in the
doc directory. (4b) This includes
files for manuscripts, documentation for source code, and/or an
electronic lab notebook recording your experiments. Subdirectories
may be created for these different classes of files in large
Put raw data and metadata in a
data directory, and
files generated during cleanup and analysis in a
directory (4c), where "generated
files" includes intermediate results, such as cleaned data sets or
simulated data, as well as final results such as figures and tables.
results directory will usually require additional
subdirectories for all but the simplest projects. Intermediate files
such as cleaned data, statistical tables, and final
publication-ready figures or tables should be separated clearly by
file naming conventions or placed into different subdirectories;
those belonging to different papers or other publications should be
grouped together. Similarly, the
data directory might require
subdirectories to organize raw data based on time, method of
collection, or other metadata most relevant to your analysis.
Put project source code in the
src contains all of
the code written for the project. This includes programs written in
interpreted languages such as R or Python; those written compiled
languages like Fortran, C++, or Java; as well as shell scripts,
snippets of SQL used to pull information from databases; and other
code needed to regenerate the results.
This directory may contain two conceptually distinct types of files that should be distinguished either by clear file names or by additional subdirectories. The first type are files or groups of files that perform the core analysis of the research, such as data cleaning or statistical analyses. These files can be thought of as the "scientific guts" of the project.
The second type of file in
src is controller or driver scripts
that that contains all the analysis steps for the entire project
from start to finish, with particular parameters and data
input/output commands. A controller script for a simple project, for
example, may read a raw data table, import and apply several cleanup
and analysis functions from the other files in this directory, and
create and save a numeric result. For a small project with one main
output, a single controller script should be placed in the main
src directory and distinguished clearly by a name such as
"runall". The short example in Figure [fig:script] is typical of
scripts of this kind; note how it uses one variable,
avoid repeating the name of a particular directory four times.
TEMP_DIR = ./temp_zip_files echo "Packaging zip files required by analysis tool..." mkdir $(TEMP_DIR) ./src/make-zip-files.py $(TEMP_DIR) *.dat echo "Analyzing..." ./bin/sqr_mean_analyze -i $(TEMP_DIR) -b "temp" echo "Cleaning up..." rm -rf $(TEMP_DIR)
Put compiled programs in the
executable programs compiled from code in the
Projects that do not have any will not require
Scripts vs. Programs
We use the term "script" to mean "something that is executed directly as-is", and "program" to mean "something that is explicitly compiled before being used". The distinction is more one of degree than kind—libraries written in Python are actually compiled to bytecode as they are loaded, for example—so one other way to think of it is "things that are edited directly" and "things that are not".
srcis for human-readable source code, and
binis for compiled binaries, where should projects put scripts that are executed directly—particularly ones that are brought in from outside the project? On the one hand, these are written in the same languages as the project-specific scripts in
src; on the other, they are executable, like the programs in
bin. The answer is that it doesn't matter, as long as each team's projects follow the same rule. As with many of our other recommendations, consistency and predictability are more important than hair-splitting.
Name all files to reflect their content or
function. (4f) For example, use names
sightings_analysis.py. Do not using sequential numbers (e.g.,
result2.csv) or a location in a final manuscript
fig_3_a.png), since those numbers will almost certainly
change as the project evolves.
The diagram in Figure [fig:project] provides a concrete example of how a simple project might be organized following these recommendations:
. |-- CITATION |-- README |-- LICENSE |-- requirements.txt |-- data | -- birds_count_table.csv |-- doc | -- notebook.md | -- manuscript.md | -- changelog.txt |-- results | -- summarized_results.csv |-- src | -- sightings_analysis.py | -- runall.py
The root directory contains a
README file that provides an overview of
the project as a whole, a
CITATION file that explains how to reference
it, and a
LICENSE file that states the licensing. The
contains a single CSV file with tabular data on bird counts
(machine-readable metadata could also be included here). The
sightings_analysis.py, a Python file containing
functions to summarize the tabular data, and a controller script
runall.py that loads the data table, applies functions imported from
sightings_analysis.py, and saves a table of summarized results in the
This project doesn't have a
bin directory, since it does not rely on
any compiled software. The
doc directory contains two text files
written in Markdown, one containing a running lab notebook describing
various ideas for the project and how these were implemented and the
other containing a running draft of a manuscript describing the project
Keeping track of changes that you or your collaborators make to data and software is a critical part of research. Being able to reference or retrieve a specific version of the entire project aids in reproducibility for you leading up to publication, when responding to reviewer comments, and when providing supporting information for reviewers, editors, and readers.
We believe that the best tools for tracking changes are the version control systems that are used in software development, such as Git, Mercurial, and Subversion. They keep track of what was changed in a file when and by whom, and synchronize changes to a central server so that many users can manage changes to the same set of files.
While these version control tools make tracking changes easier, they can have a steep learning curve. So, we provide two sets of recommendations, 1 a systematic manual approach for managing changes and 2 version control in its full glory, and you can use the first while working towards the second, or just jump in to version control.
Whatever system you chose, we recommend that you:
Back up (almost) everything created by a human being as soon as it is created. (5a) This includes scripts and programs of all kinds, software packages that your project depends on, and documentation. A few exceptions to this rule are discussed below.
Keep changes small. (5b) Each change should not be so large as to make the change tracking irrelevant. For example, a single change such as "Revise script file" that adds or changes several hundred lines is likely too large, as it will not allow changes to different components of an analysis to be investigated separately. Similarly, changes should not be broken up into pieces that are too small. As a rule of thumb, a good size for a single change is a group of edits that you could imagine wanting to undo in one step at some point in the future.
Share changes frequently. (5c) Everyone working on the project should share and incorporate changes from others on a regular basis. Do not allow individual investigator's versions of the project repository to drift apart, as the effort required to merge differences goes up faster than the size of the difference. This is particularly important for the manual versioning procedure describe below, which does not provide any assistance for merging simultaneous, possibly conflicting, changes.
Create, maintain, and use a checklist for saving and sharing changes to the project. (5d) The list should include writing log messages that clearly explain any changes, the size and content of individual changes, style guidelines for code, updating to-do lists, and bans on committing half-done work or broken code. See [gawande2011] for more on the proven value of checklists.
Store each project in a folder that is mirrored off the researcher's working machine (5e) using a system such as Dropbox20 or a remote version control repository such as GitHub21. Synchronize that folder at least daily. It may take a few minutes, but that time is repaid the moment a laptop is stolen or its hard drive fails.
Our first suggested approach, in which everything is done by hand, has two additional parts:
Add a file called
CHANGELOG.txt to the project's
docs subfolder (5f), and make dated
notes about changes to the project in this file in reverse
chronological order (i.e., most recent first). This file is the
equivalent of a lab notebook, and should contain entries like those
## 2016-04-08 * Switched to cubic interpolation as default. * Moved question about family's TB history to end of questionnaire. ## 2016-04-06 * Added option for cubic interpolation. * Removed question about staph exposure (can be inferred from blood test results).
Copy the entire project whenever a significant change has been made (5g) (i.e., one that materially affects the results), and store that copy in a sub-folder whose name reflects the date in the area that's being synchronized. This approach results in projects being organized as shown below:
. |-- project_name | -- current | -- ...project content as described earlier... | -- 2016-03-01 | -- ...content of 'current' on Mar 1, 2016 | -- 2016-02-19 | -- ...content of 'current' on Feb 19, 2016
project_name folder is mapped to external storage (such
current is where development is done, and other
project_name are old versions.
Data is Cheap, Time is Expensive
Copying everything like this may seem wasteful, since many files won't have changed, but consider: a terabyte hard drive costs about $50 retail, which means that 50 GByte costs less than $5. Provided large data files are kept out of the backed-up area (discussed below), this approach costs less than the time it would take to select files by hand for copying.
This manual procedure satisfies the requirements outlined above without needing any new tools. If multiple researchers are working on the same project, though, they will need to coordinate so that only a single person is working on specific files at any time. In particular, they may wish to create one change log file per contributor, and to merge those files whenever a backup copy is made.
What the manual process described above requires most is self-discipline. The version control tools that underpin our second approach—the one we use in our own projects–don't just accelerate the manual process: they also automate some steps while enforcing others, and thereby require less self-discipline for more reliable results.
Box 2 briefly explains how version control systems work. It's hard to know what version control tool is most widely used in research today, but the one that's most talked about is undoubtedly Git22. This is largely because of GitHub23, a popular hosting site that combines the technical infrastructure for collaboration via Git with a modern web interface. GitHub is free for public and open source projects and for users in academia and nonprofits. GitLab24 is a well-regarded alternative that some prefer, because the GitLab platform itself is free and open source. Bitbucket25 provides free hosting for both Git and Mercurial repositories, but does not have nearly as many scientific users.
The benefits of version control systems don't apply equally to all file
types. In particular, version control can be more or less rewarding
depending on file size and format. First, file comparison in version
control systems is optimized for plain text files, such as source code.
The ability to see so-called "diffs" is one of the great joys of version
control. Unfortunately, Microsoft^^ Office files (like the
used by Word) or other binary files, e.g., PDFs, can be stored in a
version control system, but it is not possible to pinpoint specific
changes from one version to the next. Tabular data (such as CSV files)
can be put in version control, but changing the order of the rows or
columns will create a big change for the version control system, even if
the data itself has not changed.
Second, raw data should not change, and therefore should not require version tracking. Keeping intermediate data files and other results under version control is also not necessary if you can re-generate them from raw data and software. However, if data and results are small, we still recommend versioning them for ease of access by collaborators and for comparison across versions.
Third, today's version control systems are not designed to handle megabyte-sized files, never mind gigabytes, so large data or results files should not be included. (As a benchmark for "large", the limit for an individual file on GitHub is 100MB.) Some emerging hybrid systems such as Git LFS26 put textual notes under version control, while storing the large data itself in a remote server, but these are not yet mature enough for us to recommend.
Researchers dealing with data subject to legal restrictions that prohibit sharing (such as medical data) should be careful not to put data in public version control systems. Some institutions may provide access to private version control systems, so it is worth checking with your IT department.
Additionally, be sure not to unintentionally place security credentials, such as passwords and private keys, in a version control system where it may be accessed by others.
An old joke says that doing the research is the first 90% of any project; writing up is the other 90%. While writing is rarely addressed in discussions of scientific computing, computing has changed scientific writing just as much as it has changed research.
A common practice in academic writing is for the lead author to send successive versions of a manuscript to coauthors to collect feedback, which is returned as changes to the document, comments on the document, plain text in email, or a mix of all three. This allows co-authors to use familiar tools, but results in a lot of files to keep track of, and a lot of tedious manual labor to merge comments to create the next master version.
Instead of an email-based workflow, we recommend mirroring good practices for managing software and data to make writing scalable, collaborative, and reproducible. As with our recommendations for version control in general, we suggest that groups choose one of two different approaches for managing manuscripts. The goals of both are to:
Ensure that text is accessible to yourself and others now and in the future by making a single master document that is available to all coauthors at all times.
Reduce the chances of work being lost or people overwriting each other's work.
Make it easy to track and combine contributions from multiple collaborators.
Avoid duplication and manual entry of information, particularly in constructing bibliographies, tables of contents, and lists.
Make it easy to regenerate the final published form (e.g., a PDF) and to tell if it is up to date.
Make it easy to share that final version with collaborators and to submit it to a journal.
The First Rule Is…
The workflow you choose is less important than having all authors agree on the workflow before writing starts. Make sure to also agree on a single method to provide feedback, be it an email thread or mailing list, an issue tracker (like the ones provided by GitHub and Bitbucket), or some sort of shared online to-do list.
Our first alternative will already be familiar to many researchers:
We realize that in many cases, even this solution is asking too much
from collaborators who see no reason to move forward from desktop GUI
tools. To satisfy them, the manuscript can be converted to a desktop
editor file format (e.g., Microsoft^^ Word
.docx or LibreOffice
.odt) after major changes, then downloaded and saved in the
folder. Unfortunately, this means merging some changes and suggestions
manually, as existing tools cannot always do this automatically when
switching from a desktop file format to text and back (although
Pandoc27 can go a long way).
The second approach treats papers exactly like software, and has been used by researchers in mathematics, astronomy, physics, and related disciplines for decades:
Using a version control system provides good support for finding and merging differences resulting from concurrent changes. It also provides a convenient platform for making comments and performing review.
This approach re-uses the version control tools and skills used to manage data and software, and is a good starting point for fully-reproducible research. However, it requires all contributors to understand a much larger set of tools, including markdown or LaTeX, make, BiBTeX, and Git/GitHub.
We initially recommended that researchers should always using plain text in version control to manage manuscripts. However, several members of our community felt strongly that the learning curve associated with this recommendation was a significant barrier to entry. For example, Stephen Turner wrote:
…try to explain the notion of compiling a document to an overworked physician you collaborate with. Oh, but before that, you have to explain the difference between plain text and word processing. And text editors. And markdown/LaTeX compilers. And BiBTeX. And Git. And GitHub. Etc. Meanwhile he/she is getting paged from the OR…
…as much as we want to convince ourselves otherwise, when you have to collaborate with those outside the scientific computing bubble, the barrier to collaborating on papers in this framework is simply too high to overcome. Good intentions aside, it always comes down to "just give me a Word document with tracked changes," or similar.
Similarly, Arjun Raj said in a blog post31:
Google Docs excels at easy sharing, collaboration, simultaneous editing, commenting and reply-to-commenting. Sure, one can approximate these using text-based systems and version control. The question is why anyone would like to…
The goal of reproducible research is to make sure one can… reproduce… computational analyses. The goal of version control is to track changes to source code. These are fundamentally distinct goals, and while there is some overlap, version control is merely a tool to help achieve that, and comes with so much overhead and baggage that it is often not worth the effort.
Collaborative editing in something like Google Docs does not have all the benefits of text-based formats (notably, being able to store manuscripts in the same place, and in the same way, as other materials). However, it does meet the requirements that we initially outlined. We still recommend against using desktop tools like LibreOffice and MicrosoftWord with either email or file-sharing services like Dropbox, as workflows based on these do not scale beyond a small number of participants.
Supplementary materials often contain much of the work that went into the project, such as tables and figures or more elaborate descriptions of the algorithms, software, methods, and analyses. In order to make these materials as accessible to others as possible, do not rely solely on the PDF format, since extracting data from PDFs is notoriously hard. Instead, we recommend separating the results that you may expect others to reuse (e.g., data in tables, data behind figures) into separate, text-format files in formats such as CSV, JSON, YAML, XML, or HDF532. The same holds for any commands or code you want to include as supplementary material: use the format that most easily enables reuse (source code files, Unix shell scripts etc).
We have deliberately left many good tools and practices off our list, including some that we use daily, because they only make sense on top of the core practices described above, or because it takes a larger investment before they start to pay off.
A branch is a "parallel universe" within a version control repository. Developers create branches so that they can make multiple changes to a project independently. They are central to the way that experienced developers use systems like Git, but they add an extra layer of complexity to version control for newcomers. Programmers got along fine in the days of CVS and Subversion without relying heavily on branching, and branching can be adopted without significant disruption after people have mastered a basic edit-commit workflow.
Tools like Make33 were originally developed to recompile pieces of software that had fallen out of date. They are now used to regenerate data and entire papers: when one or more raw input files change, Make can automatically re-run those parts of the analysis that are affected, regenerate tables and plots, and then regenerate the human-readable PDF that depends on them. However, newcomers can achieve the same behavior by writing shell scripts that re-run everything; these may do unnecessary work, but given the speed of today's machines, that is unimportant for small projects.
A unit test is a small test of one particular feature of a piece
of software. Projects rely on unit tests to prevent regression,
i.e., to ensure that a change to one part of the software doesn't
break other parts. While unit tests are essential to the health of
large libraries and programs, we have found that they usually aren't
compelling for solo exploratory work. (Note, for example, the lack
test directory in Noble's rules [noble2009].) Rather than
advocating something which people are unlikely to adopt, we have
left unit testing off this list.
Tools like Travis-CI34 automatically run a set of user-defined commands whenever changes are made to a version control repository. These commands typically execute tests to make sure that software hasn't regressed, i.e., that things which used to work still do. These tests can be run either before changes are saved (in which case the changes can be rejected if something fails) or after (in which case the project's contributors can be notified of the breakage). CI systems are invaluable in large projects with many contributors, but pay fewer dividends in smaller projects where code is being written to do specific analyses.
Profiling is the act of measuring where a program spends its time, and is an essential first step in tuning the program (i.e., making it run faster). Both are worth doing, but only when the program's performance is actually a bottleneck: in our experience, most users spend more time getting the program right in the first place.
Every modern programming language comes with tools to report the coverage of a set of test cases, i.e., the set of lines that are and aren't actually executed when those tests are run. But as with unit testing, this only starts to pay off once the project grows larger, and is therefore not recommended here.
Ontologies and other formal definitions of data are useful, but in our experience, even simplified things like Dublin Core35 are rarely encountered in the wild.
Good documentation is a key factor in software adoption, but in practice, people won't write comprehensive documentation until they have collaborators who will use it. They will, however, quickly see the point of a brief explanatory comment at the start of each script, so we have recommended that as a first step.
Researchers should use a reference manager of some sort, such as Zotero36, and should also obtain and use an ORCID37 to identify themselves in their publications, but discussion of those is outside the scope of this paper.
These practices are valuable in projects with multiple people making large software contributions, which is not typical for the intended audience for this paper [petre2014].
One important observation about this list is that many experienced programmers actually do some or all of these things even for small projects. It makes sense for them to do so because (a) they've already paid the learning cost of the tool, so the time required to implement for the "next" project is small, and (b) they understand that their project will need some or all of these things as it scales, so they might as well put it in place now.
The problem comes when those experienced developers give advice to people who haven't already mastered the tools, and don't realize (yet) that they will save time if and when their project grows. In that situation, advocating unit testing with coverage checking and continuous integration is more likely to overwhelm newcomers rather than aid them.
We have outlined a series of practices for scientific computing based on our collective experience, and the experience of the thousands of researchers we have met through Software Carpentry, Data Carpentry, and similar organizations. These practices are pragmatic, accessible to people who consider themselves new to computing, and can be applied by both individuals and groups. Most importantly, these practices make researchers more productive individually by enabling them to get more done in less time and with less pain. They also accelerate research as a whole by making computational work (which increasingly means all work) more reproducible.
But progress will not happen by itself. The practices are being increasingly incentivized through requirements from funding agencies and journals, but the time and skills required to put these approaches into practice are still not being valued.
At a local level, PIs can have the most impact, requiring that the research their lab produces follow these recommendations. Even if a PI doesn't have a background in computation, they can require that students show and share their code in lab meetings and with lab mates, that data is available and accessible to all in the lab and that computational methods sections are comprehensive. PIs can also value the time it takes to do these things effectively and provide opportunities for training.
Universities can also support such efforts. While this is often provided by IT or High Performance Computing (HPC) groups, research librarians are an often under-appreciated resource. Librarians have thought about and worked with data and provenance even before these computational challenges and increasingly have dedicated data librarians on staff who have an explicit service role.
Many campuses also have self-organized groups led by students who wish to learn from each other, which may operate independently or in concert with organizations like Software Carpentry, Data Carpentry, or The Hacker Within38.
Finally, many funding agencies now require data management plans and education and outreach activities. The true cost of implementing these plans includes training: it is unfair as well as counter-productive to insist that researchers do things without teaching them how. We believe it is now time for funders to invest in such training; we hope that our recommendations will help shape consensus on what "good enough" looks like and how to achieve it.
[brown2015]: Brown CT. How to grow a sustainable software development process; 2015. Available from: http://ivory.idyll.org/blog/2015-growing-sustainable-software-development-process.html.
[gentzkow2014]: Gentzkow M, Shapiro JM. Code and Data for the Social Sciences: A Practitioner's Guide; 2014. Available from: http://web.stanford.edu/~gentzkow/research/CodeAndData.pdf.
[kitzes2016]: Kitzes J. Reproducible Workflows; 2016. Available from: http://datasci.kitzes.com/lessons/python/reproducible_workflow.html.
[stlaurent2004]: St Laurent AM. Understanding Open Source and Free Software Licensing. O'Reilly Media; 2004. Available from: http://www.oreilly.com/openbook/osfreesoft/book/.
[steinmacher2015]: Steinmacher I, Graciotto Silva M, Gerosa M, Redmiles DF. A Systematic Literature Review on the Barriers Faced by Newcomers to Open Source Software Projects. Information and Software Technology. 2015;59(C). doi:10.1016/j.infsof.2014.11.001.
[UIllinois]: U. of Illinois Library. File Formats and Organization. Available from: http://www.library.illinois.edu/sc/services/data_management/file_formats.html.
[white2013]: White EP, Baldridge E, Brym ZT, Locey KJ, McGlinn DJ, Supp SR. Nine Simple Ways to Make It Easier to (Re)use Your Data. Ideas in Ecology and Evolution. 2013;6(2). doi:doi:10.4033/iee.2013.6b.6.f.
[wickes2015]: Wickes E. Comment on "Metadata"; 2015. Available from: https://github.com/swcarpentry/good-enough-practices-in-scientific-computing/issues/3#issuecomment-157410442.
A version control system stores snapshots of a project's files in a repository. Users modify their working copy of the project, and then save changes to the repository when they wish to make a permanent record and/or share their work with colleagues. The version control system automatically records when the change was made and by whom along with the changes themselves.
Crucially, if several people have edited files simultaneously, the version control system will detect the collision and require them to resolve any conflicts before recording the changes. Modern version control systems also allow repositories to be synchronized with each other, so that no one repository becomes a single point of failure. Tool-based version control has several benefits over manual version control:
Instead of requiring users to make backup copies of the whole project, version control safely stores just enough information to allow old versions of files to be re-created on demand.
Instead of relying on users to choose sensible names for backup copies, the version control system timestamps all saved changes automatically.
Instead of requiring users to be disciplined about completing the changelog, version control systems prompt them every time a change is saved. They also keep a 100% accurate record of what was actually changed, as opposed to what the user thought they changed, which can be invaluable when problems crop up later.
Instead of simply copying files to remote storage, version control checks to see whether doing that would overwrite anyone else's work. If so, they facilitate identifying conflict and merging changes.
As the joke goes, yourself from three months ago doesn't answer email… ↩
The node-and-arc kind. ↩
bin is an old Unix convention, and comes from the term
We recommend against more innovative formats in deference to an old saying: "What's oldest lasts longest." ↩