1 of 13

Notebooks are not enough

Israel Herraiz

@herraiz

2 of 13

Data Science needs Software Engineering

Data Scientists are software writers.

Software Engineering is a mature discipline built on top of decades of bad experiences.

Let’s learn from history!

Let’s adopt Software Engineering best practices!

3 of 13

Have ever people died because Software Engineering best practices were ignored?

4 of 13

Yes

https://en.wikipedia.org/wiki/Therac-25

http://hackaday.com/2015/10/26/killed-by-a-machine-the-therac-25

5 of 13

The Therac-25

Radiotherapy system.

Killed 4 people, and severely injured 2 more

6 of 13

The investigation concluded that (some of) the root causes were

No code reviews
No tests were conducted
Code copy-pasted from past projects
Bugs happened in corner cases

7 of 13

Notebooks in Data Science are popular

Straightforward creation of experiments
Easy sharing of experiments
Embedded graphics
Widgets for users to interact with the notebook

8 of 13

Notebooks are designed to to copy-paste share code, but not to be tested, and do not work well with Software Engineering tools

9 of 13

Think of the kittens

For every notebook with no tests, a kitten may die

10 of 13

I am a data scientist

Your testimony is a serious concern for my conscience.

Should I stop using notebooks?

How can I save more kittens?

11 of 13

You can keep using notebooks, for what they are good for.

Excellent for interactive computing

~100 kernels available

12 of 13

You are using version control and continuous integration (CI)
All code is reviewed and approved by another data scientist
No code is written without the corresponding test
Your project can be deployed from your repo
Feature decisions and bugs are logged in an issue tracking system
Your repo and the issue tracking are connected
Agree on coding guidelines, follow and measure as part of CI

The 10 7 commands of Software Engineering in Data Science

13 of 13

Can you do that with a notebook?

I find it hard to write substantial code in a notebook

http://mike.place/2016/notebook/

Jupyter Notebook is a fantastic tool that can be used in many different ways. [...] working with the Notebook on data science problems in a team setting can be challenging

https://svds.com/jupyter-notebook-best-practices-for-data-science/

version-control [..] can handle it quite well, but diff-comparing is a chore of unintelligible gibberish.

http://timstaley.co.uk/posts/making-git-and-jupyter-notebooks-play-nice/