Archive for January, 2020

January 17th, 2020

Feel free to discuss and contribute to this article over at the corresponding GitHub repo.

Many people suggest that you should use version control as part of your scientifc workflow. This is usually quickly followed up by recommendations to learn git and to put your project on GitHub. Learning and doing all of this for the first time takes a lot of effort. Alongside all of the recommendations to learn these technologies are horror stories telling how difficult it can be and memes saying that no one really knows what they are doing!

There are a lot of reasons to not embrace the git but there are even more to go ahead and do it. This is an attempt to convince you that it’s all going to be worth it alongside a bunch of resources that make it easy to get started and academic papers discussing the issues that version control can help resolve.

This document will not address how to do version control but will instead try to answer the questions what you can do with it and why you should bother. It was inspired by a conversation on twitter.

Improvements to individual workflow

Ways that git and GitHub can help your personal computational workflow – even if your project is just one or two files and you are the only person working on it.

Fixing filename hell

Is this a familiar sight in your working directory?

mycode.py
mycode_jane.py
mycode_ver1b.py
mycode_ver1c.py
mycode_ver1b_january.py
mycode_ver1b_january_BROKEN.py
mycode_ver1b_january_FIXED.py
mycode_ver1b_january_FIXED_for_supervisor.py

For many people, this is just the beginning. For a project that has existed long enough there might be dozens or even hundreds of these simple scripts that somehow define all of part of your computational workflow. Version control isn’t being used because ‘The code is just a simple script developed by one person’ and yet this situation is already becoming the breeding ground for future problems.

  • Which one of these files is the most up to date?
  • Which one produced the results in your latest paper or report?
  • Which one contains the new work that will lead to your next paper?
  • Which ones contain deep flaws that should never be used as part of the research?
  • Which ones contain possibly useful ideas that have since been removed from the most recent version?

Applying version control to this situation would lead you to a folder containing just one file

mycode.py

All of the other versions will still be available via the commit history. Nothing is ever lost and you’ll be able to effectively go back in time to any version of mycode.py you like.

git_resolution

A single point of truth

I’ve even seen folders like the one above passed down generations of PhD students like some sort of family heirloom. I’ve seen labs where multple such folders exist across a dozen machines, each one with a mixture of duplicated and unique files. That is, not only is there a confusing mess of files in a folder but there is a confusing mess of these folders!

This can even be true when only one person is working on a project. Perhaps you have one version of your folder on your University HPC cluster, one on your home laptop and one on your work machine. Perhaps you email zipped versions to yourself from time to time. There are many everyday events that can lead to this state of affairs.

By using a GitHub repository you have a single point of truth for your project. The latest version is there. All old versions are there. All discussion about it is there.

Everything…one place.

The power of this simple idea cannot be overstated. Whenever you (or anyone else) wants to use or continue working on your project, it is always obvious where to go. Never again will you waste several days work only to realise that you weren’t working on the latest version.

Keeping track of everything that changed

The latest version of your analysis or simulation is different from the previous one. Thanks to this, it may now give different results today compared to yesterday. Version control allows you to keep track of everything that changed between two versions. Every line of code you added, deleted or changed is highlighted. Combined with your commit messages where you explain why you made each set of changes, this forms a useful record of the evolution of your project.

commit_example

It is possible to compare the differences between any two commits, not just two consecutive ones which allows you to track the evolution of your project over time.

Always having a working version of your project

Ever noticed how your collaborator turns up unnanounced just as you are in the middle of hacking on your code. They want you to show them your simulation running but right now its broken! You frantically try some of the other files in your folder but none of them seem to be the version that was working last week when you sent the report that moved your collaborator to come to see you.

If you were using version control you could easily stash your current work, revert to the last good commit and show off your work.

Tracking down what went wrong

You are always changing that script and you test it as much as you can but the fact is that the version from last year is giving correct results in some edge case while your current version is not. There are 100 versions between the two and there’s a lot of code in each version! When did this edge case start to go wrong?

With git you can use git bisect to help you track down which commit started causing the problem which is the first step towards fixing it.

fire_gif

Providing a back up of your project

Try this thought experiment: Your laptop/PC has gone! Fire, theft, dead hard disk or crazed panda attack.

angry_panda

It, and all of it’s contents have vanished forever. How do you feel? What’s running through your mind? If you feel the icy cold fingers of dread crawling up your spine as you realise Everything related to my PhD/project/life’s work is lost then you have made bad life choices. In particular, you made a terrible choice when you neglected to take back ups.

Of course there are many ways to back up a project but if you are using the standard version control workflow, your code is automatically backed up as a matter of course. You don’t have to remember to back things up, back-ups happen as a natural result of your everyday way of doing things.

Making your project easier to find and install

There are dozens of ways to distribute your software to someone else. You could (HORRORS!) email the latest version to a colleuage or you could have a .zip file on your web site and so on.

Each of these methods has a small cognitive load for both recipient and sender. You need to make sure that you remember to update that .zip file on your website and your user needs to find it. I don’t want to talk about the email case, it makes me too sad. If you and your collaborator are emailing code to each other, please stop. Think of the children!

One great thing about using GitHub is that it is a standardised way of obtaining software. When someone asks for your code, you send them the URL of the repo. Assuming that the world is a better place and everyone knows how to use git, you don’t need to do anything else since the repo URL is all they need to get your code. a git clone later and they are in business.

Additionally, you don’t need to worry abut remembering to turn your working directory into a .zip file and uploading it to your website. The code is naturally available for download as part of the standard workflow. No extra thought needed!

In addition to this, some popular computational environments now allow you to install packages directly from GitHub. If, for example, you are following standard good practice for building an R package then a user can install it directly from your GitHub repo from within R using the devtools::install_github() function.

Automatically run all of your tests

You’ve sipped of the KoolAid and you’ve been writing unit tests like a pro. GitHub allows you to link your repo with something called Continuous Integration (CI) that helps maximise the utility of those tests.

Once its all set up the CI service runs every time you, or anyone else, makes a commit to your project. Every time the CI service runs, a virtual machine is created from scratch, your project is installed into it and all of your tests are run with any failures reported.

This gives you increased confidence that everything is OK with your latest version and you can choose to only accept commits that do not break your testing framework.

jdcook_tweet

Collaboration and Community

How git and GitHub can make it easier to collaborate with others on computational projects.

Control exactly who can see your work

‘I don’t want to use GitHub because I want to keep my project private’ is a common reason given to me for not using the service. The ability to create private repositories has been free for some time now (Price plans are available here https://github.com/pricing) and you can have up to 3 collaborators on any of your private repos before you need to start paying. This is probably enough for most small academic projects.

This means that you can control exactly who sees your code. In the early stages it can be just you. At some point you let a couple of trusted collaborators in and when the time is right you can make the repo public so everyone can enjoy and use your work alongside the paper(s) it supports.

Faciliate discussion about your work

Every GitHub repo comes with an Issues section which is effectively a discussion forum for the project. You can use it to keep track of your project To-Do list, bugs, documentation discussions and so on. The issues log can also be integrated with your commit history. This allows you to do things like git commit -m "Improve the foo algorithm according to the discussion in #34" where #34 refers to the Issue discussion where your collaborator pointed out

Allow others to contribute to your work

You have absolute control over external contributions! No one can make any modifications to your project without your explicit say-so.

I start with the above statement because I’ve found that when explaining how easy it is to collaborate on GitHub, the first question is almost always ‘How do I keep control of all of this?’

What happens is that anyone can ‘fork’ your project into their account. That is, they have an independent copy of your work that is clearly linked back to your original. They can happily work away on their copy as much as they like – with no involvement from you. If and when they want to suggest that some of their modifications should go into your original version, they make a ‘Pull Request’.

I emphasised the word ‘Request’ because that’s exactly what it is. You can completely ignore it if you want and your project will remain unchanged. Alternatively you might choose to discuss it with the contributor and make modifications of your own before accepting it. At the other end of the spectrum you might simply say ‘looks cool’ and accept it immediately.

Congratulations, you’ve just found a contributing collaborator.

Reproducible research

How git and GitHub can contribute to improved reproducible research.

Simply making your software available

A paper published without the supporting software and data is (much!) harder to reproduce than one that has both.

Making your software citable

Most modern research cannot be done without some software element. Even if all you did was run a simple statistical test on 20 small samples, your paper has a data and software dependency. Organisations such as the Software Sustainability Institute and the UK Research Software Engineering Association (among many others) have been arguing for many years that such software and data dependencies should be part of the scholarly record alongside the papers that discuss them. That is, they should be archived and referenced with a permanent Digital Object Identifier (DOI).

Once your code is in GitHub, it is straightforward to archive the version that goes with your latest paper and get it its own DOI using services such as Zenodo. Your University may also have its own archival system. For example, The University of Sheffield in the UK has built a system called ORDA which is based on an institutional Figshare instance which allows Sheffield academics to deposit code and data for long term archival.

Which version gave these results?

Anyone who has worked with software long enough knows that simply stating the name of the software you used is often insufficient to ensure that someone else could reproduce your results. To help improve the odds, you should state exactly which version of the software you used and one way to do this is to refer to the git commit hash. Alternatively, you could go one step better and make a GitHub release of the version of your project used for your latest paper, get it a DOI and cite it.

This doesn’t guarentee reproducibility but its a step in the right direction. For extra points, you may consider making the computational environment reproducible too (e.g. all of the dependencies used by your script – Python modules, R packages and so on) using technologies such as Docker, Conda and MRAN but further discussion of these is out of scope for this article.

Building a computational environment based on your repository

Once your project is on GitHub, it is possible to integrate it with many other online services. One such service is mybinder which allows the generation of an executable environment based on the contents of your repository. This makes your code immediately reproducible by anyone, anywhere.

Similar projects are popping up elsewhere such as The Littlest JupyterHub deploy to Azure button which allows you to add a button to your GitHub repo that, when pressed by a user, builds a server in their Azure cloud account complete with your code and a computational environment specified by you along with a JupterHub instance that allows them to run Jupyter notebooks. This allows you to write interactive papers based on your software and data that can be used by anyone.

Complying with funding and journal guidelines

When I started teaching and advocating the use of technologies such as git I used to make a prediction These practices are so obviously good for computational research that they will one day be mandated by journal editors and funding providers. As such, you may as well get ahead of the curve and start using them now before the day comes when your funding is cut off because you don’t. The resulting debate was usually good fun.

My prediction is yet to come true across the board but it is increasingly becoming the case where eyebrows are raised when papers that rely on software are published don’t come with the supporting software and data. Research Software Engineers (RSEs) are increasingly being added to funding review panels and they may be Reviewer 2 for your latest paper submission.

Other uses of git and GitHub for busy academics

It’s not just about code…..

  • Build your own websites using GitHub pasges. Every repo can have its own website served directly from GitHub
  • Put your presentations on GitHub. I use reveal.js combined with GitHub pages to build and serve my presentations. That way, whenever I turn up at an event to speak I can use whatever computer is plugged into the projector. No more ‘I don’t have the right adaptor’ hell for me.
  • Write your next grant proposal. Use Markdown, LaTex or some other git-friendly text format and use git and GitHub to collaboratively write your next grant proposal

The movie below is a visualisation showing how a large H2020 grant proposal called OpenDreamKit was built on GitHub. Can you guess when the deadline was based on the activity?

Further Resources

Further discussions from scientific computing practitioners that discuss using version control as part of a healthy approach to scientific computing

Learning version control

Convinced? Want to start learning? Let’s begin!

Graphical User Interfaces to git

If you prefer not to use the command line, try these

January 6th, 2020

My stepchildren are pretty good at mathematics for their age and have recently learned about Pythagora’s theorem

$c=\sqrt{a^2+b^2}$

The fact that they have learned about this so early in their mathematical lives is testament to its importance. Pythagoras is everywhere in computational science and it may well be the case that you’ll need to compute the hypotenuse to a triangle some day.

Fortunately for you, this important computation is implemented in every computational environment I can think of!
It’s almost always called hypot so it will be easy to find.
Here it is in action using Python’s numpy module

import numpy as np
a = 3
b = 4
np.hypot(3,4)

5

When I’m out and about giving talks and tutorials about Research Software Engineering, High Performance Computing and so on, I often get the chance to mention the hypot function and it turns out that fewer people know about this routine than you might expect.

Trivial Calculation? Do it Yourself!

Such a trivial calculation, so easy to code up yourself! Here’s a one-line implementation

def mike_hypot(a,b):
    return(np.sqrt(a*a+b*b))

In use it looks fine

mike_hypot(3,4)

5.0

Overflow and Underflow

I could probably work for quite some time before I found that my implementation was flawed in several places. Here’s one

mike_hypot(1e154,1e154)

inf

You would, of course, expect the result to be large but not infinity. Numpy doesn’t have this problem

np.hypot(1e154,1e154)

1.414213562373095e+154

My function also doesn’t do well when things are small.

a = mike_hypot(1e-200,1e-200)

0.0

but again, the more carefully implemented hypot function in numpy does fine.

np.hypot(1e-200,1e-200)

1.414213562373095e-200

Standards Compliance

Next up — standards compliance. It turns out that there is a an official standard for how hypot implementations should behave in certain edge cases. The IEEE-754 standard for floating point arithmetic has something to say about how any implementation of hypot handles NaNs (Not a Number) and inf (Infinity).

It states that any implementation of hypot should behave as follows (Here’s a human readable summary https://www.agner.org/optimize/nan_propagation.pdf)

hypot(nan,inf) = hypot(inf,nan) = inf

numpy behaves well!

np.hypot(np.nan,np.inf)

inf

np.hypot(np.inf,np.nan)

inf

My implementation does not

mike_hypot(np.inf,np.nan)

nan

So in summary, my implementation is

  • Wrong for very large numbers
  • Wrong for very small numbers
  • Not standards compliant

That’s a lot of mistakes for one line of code! Of course, we can do better with a small number of extra lines of code as John D Cook demonstrates in the blog post What’s so hard about finding a hypotenuse?

Hypot implementations in production

Production versions of the hypot function, however, are much more complex than you might imagine. The source code for the implementation used in openlibm (used by Julia for example) was 132 lines long last time I checked. Here’s a screenshot of part of the implementation I saw for prosterity. At the time of writing the code is at https://github.com/JuliaMath/openlibm/blob/master/src/e_hypot.c

 

openlibm_hypot

 

That’s what bullet-proof, bug checked, has been compiled on every platform you can imagine and survived code looks like.

There’s more!

Active Research

When I learned how complex production versions of hypot could be, I shouted out about it on twitter and learned that the story of hypot was far from over!

The implementation of the hypot function is still a matter of active research! See the paper here https://arxiv.org/abs/1904.09481

hypot_twitter

Is Your Research Software Correct?

Given that such a ‘simple’ computation is so complicated to implement well, consider your own code and ask Is Your Research Software Correct?.

TOP