A logical, reasonably standardized, but flexible project structure for doing and sharing data science work.
We’re not talking about bikeshedding the indentation aesthetics or pedantic formatting standards — ultimately, data science code quality is about correctness and reproducibility.
When we think about data analysis, we often think just about the resulting reports, insights, or visualizations. While these end products are generally the main event, it’s easy to focus on making the products look nice and ignore the quality of the code that generates them. Because these end products are created programmatically, code quality is still important! And we’re not talking about bikeshedding the indentation aesthetics or pedantic formatting standards — ultimately, data science code quality is about correctness and reproducibility.
It’s no secret that good analyses are often the result of very scattershot and serendipitous explorations. Tentative experiments and rapidly testing approaches that might not work out are all part of the process for getting to the good stuff, and there is no magic bullet to turn data exploration into a simple, linear progression.
That being said, once started it is not a process that lends itself to thinking carefully about the structure of your code or project layout, so it’s best to start with a clean, logical structure and stick to it throughout. We think it’s a pretty big win all around to use a fairly standardized setup like this one. Here’s why:
Nobody sits around before creating a new Rails project to figure out where they want to put their views; they just run
rails newto get a standard project skeleton like everybody else.
A well-defined, standard project structure means that a newcomer can begin to understand an analysis without digging in to extensive documentation. It also means that they don’t necessarily have to read 100% of the code before knowing where to look for very specific things.
Well organized code tends to be self-documenting in that the organization itself provides context for your code without much overhead. People will thank you for this because they can:
A good example of this can be found in any of the major web development frameworks like Django or Ruby on Rails. Nobody sits around before creating a new Rails project to figure out where they want to put their views; they just run rails new to get a standard project skeleton like everybody else. Because that default project structure is logical and reasonably standard across most projects, it is much easier for somebody who has never seen a particular project to figure out where they would find the various moving parts.
Another great example is the Filesystem Hierarchy Standard for Unix-like systems. The /etc directory has a very specific purpose, as does the /tmp folder, and everybody (more or less) agrees to honor that social contract. That means a Red Hat user and an Ubuntu user both know roughly where to look for certain types of files, even when using each other’s system — or any other standards-compliant system for that matter!
Ideally, that’s how it should be when a colleague opens up your data science project.
Ever tried to reproduce an analysis that you did a few months ago or even a few years ago? You may have written the code, but it’s now impossible to decipher whether you should use make_figures.py.old, make_figures_working.py or new_make_figures01.py to get things done. Here are some questions we’ve learned to ask with a sense of existential dread:
These types of questions are painful and are symptoms of a disorganized project. A good project structure encourages practices that make it easier to come back to old work, for example separation of concerns, abstracting analysis as a DAG, and engineering best practices like version control.
“A foolish consistency is the hobgoblin of little minds” — Ralph Waldo Emerson (and PEP 8!)
Disagree with a couple of the default folder names? Working on a project that’s a little nonstandard and doesn’t exactly fit with the current structure? Prefer to use a different package than one of the (few) defaults?
Go for it! This is a lightweight structure, and is intended to be a good starting point for many projects. Or, as PEP 8 put it:
Consistency within a project is more important. Consistency within one module or function is the most important. … However, know when to be inconsistent – sometimes style guide recommendations just aren’t applicable. When in doubt, use your best judgment. Look at other examples and decide what looks best. And don’t hesitate to ask!
With this in mind, we’ve created a data science cookiecutter template for projects in Python. Your analysis doesn’t have to be in Python, but the template does provide some Python boilerplate that you’d want to remove (in the src folder for example, and the Sphinx documentation skeleton in docs).
pip install cookiecutterStarting a new project is as easy as running this command at the command line. No need to create a directory first, the cookiecutter will do it for you.
cookiecutter https://github.com/crplab/cdst
The initial project structure will be created. Open your project directory and run:
make init
After this command, you’ll get:
.conda-env.yml file.├── LICENSE
├── Makefile <- Makefile with commands like `make init` or `make clean`
├── make.bat <- It's like the Makefile, but for Windows
├── README.md <- The top-level README for developers using this project.
├── data
│ ├── external <- Data from third party sources.
│ ├── interim <- Intermediate data that has been transformed.
│ ├── processed <- The final, canonical data sets for modeling.
│ ├── features <- Features may be stored here
│ ├── inference <- Inference stages may be stored here
│ └── raw <- The original, immutable data dump.
│
├── docs <- A default Sphinx project; see sphinx-doc.org for details
│
├── models <- Trained and serialized models, model predictions, or model summaries
│
├── notebooks <- Jupyter notebooks. Naming convention is a number (for ordering),
│ the creator's initials, and a short `-` delimited description, e.g.
│ `1.0-jqp-initial-data-exploration`.
│
├── references <- Data dictionaries, manuals, and all other explanatory materials.
│
├── reports <- Generated analysis as HTML, PDF, LaTeX, etc.
│ └── figures <- Generated graphics and figures to be used in reporting
│
├── .conda-env.yml <- conda environment definition
│
├── .pre-commit-config.yaml <- pre-commit configuration
│
├── requirements.txt <- The requirements file for reproducing the analysis environment, e.g.
│ generated with `pip freeze > requirements.txt`
│
├── __init__.py
│
└──
├── __init__.py <- Makes a Python module
├── settings.py <- illustrates how to use .env file
├── data <- Scripts to download or generate data
│ └── make_dataset.py
├── features <- Scripts to turn raw data into features for modeling
│ └── featurize.py
└── models <- Scripts to train models and then use trained models to make
│ predictions
└── train.py
There are some opinions implicit in the project structure that have grown out of our experience with what works and what doesn’t when collaborating on data science projects. Some of the opinions are about workflows, and some of the opinions are about tools that make life easier. Here are some of the beliefs which this project is built on—if you’ve got thoughts, please contribute or share them.
Don’t ever edit your raw data, especially not manually, and especially not in Excel. Don’t overwrite your raw data. Don’t save multiple versions of the raw data. Treat the data (and its format) as immutable. The code you write should move the raw data through a pipeline to your final analysis. You shouldn’t have to run all of the steps every time you want to make a new figure (see Analysis is a DAG), but anyone should be able to reproduce the final products with only the code in src and the data in data/raw.
Also, if data is immutable, it doesn’t need source control in the same way that code does. Therefore, by default, the data folder is included in the .gitignore file. If you have a small amount of data that rarely changes, you may want to include the data in the repository. Github currently warns if files are over 50MB and rejects files over 100MB. Some other options for storing/syncing large data include AWS S3 with a syncing tool (e.g., s3cmd), Git Large File Storage, Git Annex, and dat. Currently by default, we ask for an S3 bucket and use AWS CLI to sync data in the data folder with the server.
Notebook packages like the Jupyter notebook, Beaker notebook, Zeppelin, and other literate programming tools are very effective for exploratory data analysis. However, these tools can be less effective for reproducing an analysis. When we use notebooks in our work, we often subdivide the notebooks folder. For example, notebooks/exploratory contains initial explorations, whereas notebooks/reports is more polished work that can be exported as html to the reports directory.
Since notebooks are challenging objects for source control (e.g., diffs of the json are often not human-readable and merging is near impossible), we recommended not collaborating directly with others on Jupyter notebooks. There are two steps we recommend for using notebooks effectively:
Follow a naming convention that shows the owner and the order the analysis was done in. We use the format <step>-<ghuser>-<description>.ipynb (e.g., 0.3-bull-visualize-distributions.ipynb).
Refactor the good parts. Don’t write code to do the same task in multiple notebooks. If it’s a data preprocessing task, put it in the pipeline at src/data/make_dataset.py and load data from data/interim. If it’s useful utility code, refactor it to src.
Now by default we turn the project into a Python package (see the setup.py file). You can import your code and use it in notebooks with a cell like the following:
# OPTIONAL: Load the "autoreload" extension so that code can change
%load_ext autoreload
# OPTIONAL: always reload modules so that as you change code in src, it gets loaded
%autoreload 2
from src.data import make_dataset
Often in an analysis you have long-running steps that preprocess data or train models. If these steps have been run already (and you have stored the output somewhere like the data/interim directory), you don’t want to wait to rerun them every time. We prefer make for managing steps that depend on each other, especially the long-running ones. Make is a common tool on Unix-based platforms (and is available for Windows). Following the make documentation, Makefile conventions, and portability guide will help ensure your Makefiles work effectively across systems. Here are some examples to get started. A number of data folks use make as their tool of choice, including Mike Bostock.
There are other tools for managing DAGs that are written in Python instead of a DSL (e.g., Paver, Luigi, Airflow, Snakemake, Ruffus, or Joblib). Feel free to use these if they are more appropriate for your analysis.
You really don’t want to leak your AWS secret key or Postgres username and password on Github. Enough said — see the Twelve Factor App principles on this point. Here’s one way to do this:
Create a .env file in the project root folder. Thanks to the .gitignore, this file should never get committed into the version control repository. Here’s an example:
# example .env file
DATABASE_URL=postgres://username:password@localhost:5432/dbname
AWS_ACCESS_KEY=myaccesskey
AWS_SECRET_ACCESS_KEY=mysecretkey
OTHER_VARIABLE=something
If you look at the stub script in src/data/make_dataset.py, it uses a package called python-dotenv to load up all the entries in this file as environment variables so they are accessible with os.environ.get. Here’s an example snippet adapted from the python-dotenv documentation:
# src/data/dotenv_example.py
import os
from dotenv import load_dotenv, find_dotenv
# find .env automagically by walking up directories until it's found
dotenv_path = find_dotenv()
# load up the entries as environment variables
load_dotenv(dotenv_path)
database_url = os.environ.get("DATABASE_URL")
other_variable = os.environ.get("OTHER_VARIABLE")
When using Amazon S3 to store data, a simple method of managing AWS access is to set your access keys to environment variables. However, managing mutiple sets of keys on a single machine (e.g. when working on multiple projects) it is best to use a credentials file, typically located in ~/.aws/credentials. A typical file might look like:
[default]
aws_access_key_id=myaccesskey
aws_secret_access_key=mysecretkey
[another_project]
aws_access_key_id=myprojectaccesskey
aws_secret_access_key=myprojectsecretkey
You can add the profile name when initialising a project; assuming no applicable environment variables are set, the profile credentials will be used be default.
Project structure and reproducibility is talked about more in the R research community. Here are some projects and blog posts if you’re working in R that may help you out.
Finally, a huge thanks to the Cookiecutter project (github), which is helping us all spend less time thinking about and writing boilerplate and more time getting things done.