pdpp 😁

Principled Data Processing with Python

• Installation Prerequisites
• Installation
• Example
  • Installing from source with Poetry
  • Initializing a new project
  • Adding tasks to the project pipeline
  • Executing the pipeline
  • Exporting finished work from any task
  • Visualizing task dependencies (i.e., the pipeline)
  • Running tasks in different virtual environments
  • Storing task parameters / configuration files alongside source code in src
• Command Line Usage
  • pdpp init
  • pdpp new
  • pdpp custom
  • pdpp sub
  • pdpp rig
  • pdpp run or doit
  • pdpp graph
  • pdpp extant
  • pdpp enable

pdpp is a command-line interface enabling transparent and reproducible data science workflows. It's designed around the principles espoused by Patrick Ball (Human Rights Data Analysis Group) in his 2016 Data & Society talk 'Principled Data Processing' (PDP). pdpp can be used to create a data processing and modelling pipeline consisting of modular 'tasks' with linked inputs and outputs. pdpp run executes these tasks as needed by using the doit suite of automation tools. pdpp also creates visualizations of task dependencies in a project pipeline with varying levels of details. For example,

In the principled data processing framework, the task is a quantum of workflow. Each task takes the form of a directory responsible for a discrete data operation, such as extracting records from plaintext documents and storing the results as a .csv file. Ideally, each task should be simple and conceptually unified enough that a short 3- or 4-word description in the directory name is enough to convey what the task accomplishes.

Each task directory contains at minimum three subdirectories:

1. input, which contains all of the task's local data dependencies;
2. output, which contains all of the task's local data outputs (also referred to as 'targets'); and
3. src, which contains all of the task's source code, ideally contained within a single script file.

The pdpp package adds two additional constraints to Patrick Ball's original formulation of PDP:

1. All local data files needed by the workflow but which are not generated by any of the workflow's tasks must be included in the _import_ directory, which pdpp places at the same directory level as the overall workflow during project initialization.
2. All local data files produced by the workflow as project outputs must be routed into the _export_ directory, which pdpp places at the same directory level as the overall workflow during project initialization.

These additional constraints disambiguate the input and output of the overall workflow, which permits pdpp workflows to be embedded within one another. Note that these directories, as well as all import and export directories contain .gitignore files by default.

Installation Prerequisites

The pdpp package depends on graphviz, which must be installed before attempting to install pdpp. Installation instructions for graphviz can be found at the GraphViz installation instructions page.

Installation

To install pdpp, use the following terminal command:

pip install pdpp

Installing from source with Poetry

It's best to install poetry using pipx.

On macOS:

brew install pipx
pipx ensurepath
pipx install poetry

Or on Debian/Ubuntu Linux:

sudo apt install pipx
pipx ensurepath
pipx install poetry

Once poetry is installed, clone and cd into the repo.

git clone https://github.com/UWNETLAB/pdpp.git
cd pdpp

Use poetry to create a virtual environment and install pdpp and it's dependencies.

poetry install

Activate the virtual environment with poetry shell and work within the virtual environment, or append poetry run to your commands to execute them inside the environment, e.g.:

poetry run pdpp init

Example

You can find a simple pdpp example in examples/.

Initializing a new project

The first step when using pdpp is to initialize a new project directory, which must be empty.

cd examples
mkdir example_1
cd example_1
poetry shell # if using pdpp in the package poetry environment

To initialize the new project directory, use the init command:

pdpp init

Doing so should produce a directory tree similar to this one:

For the purposes of this example, we added .csv file containing some toy data has been added to the _import_ directory.

Adding tasks to the project pipeline

At this point, we're ready to add our first task to the project. To do this, we'll use the new command:

pdpp new

Upon executing the command, pdpp will request a name for the new task. We'll call it task_1. After supplying the name, pdpp will display an interactive menu that allows you to easily specify other project tasks containing files that the new task -- task_1 -- depends upon. At the moment, there's only one other task in the project that can files can be imported from: _import_. Select it (using spacebar) and press the enter/return key. pdpp will then display a nested list of all the files available to be nominated as a dependency of task_1:

Select example.csv and press enter. pdpp will inform you that your new task has been created. At this point, the project's tree diagram should appear similar to this:

The tree diagram shows that task_1 exists and that it's input directory contains example.csv (which is hard linked to the example.csv file in _import_, meaning that any changes to one will be reflected in the other). The src directory also contains a Python script titled task_1.py -- for the sake of convenience, pdpp populates new tasks with a synonymous Python script. This file can be deleted or renamed if the task requires source code by a different name or using a different programming language. For this example, we'll populate the Python script with a simple set of instructions for loading example.csv, adding one to each of the items in it, and then saving it as a new .csv file in the output subdirectory of the task_1 directory:

import pandas as pd

df = pd.read_csv('../input/example.csv')
df_plus_one = df + 1

print(df)
print(df_plus_one)

df_plus_one.to_csv('../output/df_plus_one.csv', index=False)

Since each task contains it's own inputs, source code, and outputs, you can run a task by running the source in src.

python task_1.py

But in most cases you'll want to simply run the full pipeline and have pdpp execute other tasks if necessary.

Executing the pipeline

You can execute the entire pipeline -- which is only a single task so far -- by using one of the two following commands (both are functionally identical):

doit

pdpp run

Running this example should result in the following console message:

.  task_1

When a workflow is run, the doit automation suite -- atop which pdpp is built -- lists the name of each task in the workflow. If the task needed to be executed, a period is displayed before the task. If the task did not need to be executed, two dashes are displayed before the task name, like so:

-- task_1

This is because doit checks the relative ages of each tasks' inputs and outputs at runtime; if a given task has any outputs (or 'targets', in doit nomenclature) that are older than one or more of the task's inputs (or 'dependencies,' in doit nomenclature), then that task must be re-run. If all of a task's inputs are older than its outputs, the task does not need to be run.

This means that a pdpp/doit pipeline can be run as often as the user desires without running the risk of needlessly wasting time or computing power: tasks will only be re-run if changes to 'upstream' files necessitate it. You can read more about this feature of the doit suite here.

After executing pdpp run (or running the source code from the src subdirectory), a new file called df_plus_one.csv will be saved in task_1's output subdirectory.

Building out the pipeline

We can continue to build our our data processing and modelling pipeline by following this simple process of creating and linking tasks. For example, we can use pdpp new to create task_2 and add the output of task_1 as a dependency.

task_1/src/task_1.py added 1. It's a bit ridiculous, of course, but let's just have task_2/src/task_2.py add 2. Why not?

import pandas as pd

df = pd.read_csv('../input/df_plus_one.csv')
df_plus_two = df + 2

print(df)
print(df_plus_two)

df_plus_two.to_csv('../output/df_plus_two.csv', index=False)

Exporting finished work from any task

You can continue this process until your pipeline is complete. At any point in the pipeline, a task might produce an output that is ready to be shared (e.g., a table, figure, model, report). We can link that output to the _export_ directory to make it easily accessible and sharable. We can do this using the rig command, which is used to (re-)configure dependencies for tasks that already exit (such as _export_, which is a special kind of task). Run the command:

pdpp rig

Select _export_ from the list of tasks available, then select task_2 (and not _import_); finally, select df_plus_two.csv as the only dependency for _export_.

Once _export_ has been rigged, this simple project example_1 is a complete pdpp pipeline. The workflow imports a simple .csv file, executes task_1 to add 1 to each number in the file, then executes task_2 to add 2 to each number in the file produced by task_1, and then exports the .csv file produced by task_2.

Visualizing task dependencies (i.e., the pipeline)

pdpp pipelines can be visualized using the built-in visualization suite like so:

pdpp graph

The above command will prompt you for two pieces of information:

• the output format for the visualization (defaults to .png), and
• the colour scheme to be used (defaults to colour, can be set to greyscale).

You can accept both defaults by pressing enter twice. Alternatively, you can skip the interaction by running

pdpp graph --files png --style default

Whenever pdpp graphs a pipeline, it produces four visualizations with different amounts of information.

• Nodes with box-like nodes represent tasks
• Nodes with folder corners represent data files
• Nodes with two tabs on the left-hand side represent source code

Here are all four visualized for example_1, from least amount of information (dependencies_sparse.png) to most (dependencies_all.png).

dependencies_sparse.png:

dependencies_source.png:

dependencies_file.png:

dependencies_all.png:

Running tasks in different virtual environments

Coming soon...

Storing task parameters / configuration files alongside source code in src

Coming soon...

Command Line Usage

All pdpp commands (other than pdpp init) must be run from the project's base pdpp directory (i.e., the directory containing _import_, _export_, and other task subdirectories).

pdpp init

Initializes a new pdpp project in an empty directory.

pdpp new

Adds a new task to a pdpp project and launches an interactive rigging session (see pdpp rig below for more information) where you can specify the task's dependencies.

pdpp custom

Adds a new custom task to a pdpp project and launches an interactive rigging session for it (see pdpp rig below for more information). Custom tasks make no assumptions about the kind of source code or automation required by the task; it is up to the user to write a custom dodo.py file that will correctly automate the task in the context of the larger workflow. More information on writing custom dodo.py files can be found here.

pdpp sub

Adds a new sub-project task to a pdpp project and launches an interactive rigging session for it (see pdpp rig below for more information). Sub-project tasks are distinct pdpp projects nested inside the main project -- structurally, they function identically to all other pdpp projects. Their dependencies are defined as any local files contained inside their _import_ directory (which functions as if it were an input directory for a task) and their targets are defined as any local files contained inside their _export_ directory (which functions as if if were an output directory for a task).

pdpp rig

Launches an interactive rigging session for a selected task, which allows users to specify the task's dependencies (inputs), targets (outputs), and source code (src). In practice, explicit rigging of a task's source code is rarely necessary and it may be worth considering using a custom task (pdpp custom) instead of the default basic task. However, there are situations where you'll want to specify the task's main source code file, such as when you want to [run a task using a different virtual environment than other tasks](#running-tasks-in-different-virtual environments) or store a task's parameters or other configuration in files alongside the it's source code.

pdpp run or doit

Runs the project. The pdpp run command provides basic functionality; users may pass arguments to the doit command that provides a great deal of control and specificity. More information about the doit command can be found here.

pdpp graph

Produces four visualizations of the pdpp project:

• dependencies_sparse only includes task nodes.
• dependencies_file includes task nodes and data files.
• dependencies_source includes task nodes and source files.
• dependencies_all includes task nodes, source files, and data files.

pdpp extant

Incorporates an already-PDP compliant directory (containing input, output, and src directories) that was not created using pdpp into the pdpp project.

pdpp enable

Allows users to toggle tasks 'on' or 'off'; tasks that are 'off' will not be executed when pdpp run or doit is used.