About Me

• BSc Mathematics and Statistics - Plymouth University
• MSc Statistics - Sheffield University
• Worked as a statistician for C3 Resources and Plymouth University
• Currently work as a R / data science consultant for Mango Solutions

Agenda

• Reproducible workflows
• Process consistency
• Introducing remake

Single Script Analyses

|- ProjectName/
    |- data/
    |- output/
    |- run.R
    |- README.md

Multiple Script Analyses

|- ProjectName/
    |- data/
    |    |- raw/
    |    |- cleaned/
    |    |- prepared/
    |- output/ 
    |    |- fig/
    |    |- results/
    |- lib
    |    |- source.R
    |    |- explore.R
    |    |- prepare.R
    |    |- model.R
    |    |- evaluate.R
    |    |- utils.R
    |- run.R
    |- README.md

Problems

• Not scalable
• Undocumented dependencies
• Data can change
  • When it does - which parts of the analysis will need re-running?
• Manually re-running analyses is prone to error

Two Related Problems

• Making research reproducible
  • Recreate outputs from a paper
  • Record entire workflows
• Efficient R based workflows
  • When do we re-run code?

remake

• remake is “Make-like build management, reimagined for R”

• Written by Rich Fitz John

• remake allows you to:
  • Change parts of your workflow and only update what changed
  • Make your workflow reproducible since it incorporates the data importing, analysis and reporting in the same pipeline
• Note: remake is in active development

remake Components

• “Targets”: Objects of interest
• “Rules”: Functions returning targets
• “Dependencies”: Upstream targets

remake Components - The YAML File

• The yaml file tells remake everything that it needs to know.
• It is the “map” of your analysis

Example

• nycflights13 (install.packages("nycflights13"))
  • contains information about all flights that departed from NYC in 2013
  • 336,776 flights in total
• This example uses the datasets:
  • flights, airports, airlines
• I want to generate a report with
  • The number of flights from each airport each month
  • The mean departure delay per airline each month

The YAML File - Packages and Source Files

• At the start, write:
  • the packages you need to load
  • the functions you need to source

The YAML File - the Targets

• Next define the targets
  • The all target is the final output

The YAML File - the Targets

• Now we run the functions on the transformed data and produce the graphics

The YAML File - the Targets

• Finally, we create the output

Visualising Your Workflow

• remake uses the DiagrammeR package to visualize your workflows!
• You can diagram your whole pipeline using the diagram() function from remake.

Visualising Your Workflow

Running the Analysis

• Running remake::make() will produce the following output:

> remake::make()
[  LOAD ] 
[  READ ]                                              |  # loading sources
<  MAKE > all
[ BUILD ] transformedData                              |  transformedData <- mapNames()
[  READ ]                                              |  # loading packages
Joining, by = "carrier"
[ BUILD ] numFlights                                   |  numFlights <- numFlightsFromAirport(transformedData)
[ BUILD ] depDelay                                     |  depDelay <- meanDepartureDelayByCarrier(transformed...
[  PLOT ] figures/num_flights_per_airport.png          |  plotNumFlights(numFlights) # ==> figures/num_flights_per_airport...
[  PLOT ] figures/mean_departure_delay_per_carrier.png |  plotDepDelay(depDelay) # ==> figures/mean_departure_delay_per_ca...
[ BUILD ] report.html                                  |  render("report.Rmd")
...
[ ----- ] all

Re-running the Analysis

> remake::make()
<  MAKE > all
[    OK ] transformedData
[    OK ] numFlights
[    OK ] depDelay
[    OK ] figures/num_flights_per_airport.png
[    OK ] figures/mean_departure_delay_per_carrier.png
[    OK ] report.html
[ ----- ] all

Other Essential Tools

• git
• packrat
• testthat
• Roxygen2
• Continuous integration (travis-ci, Jenkins, etc.)
• Code coverage (covr)

Other Automation Tools

• make
• Flowr
• Ruigi
• Many more (https://github.com/pditommaso/awesome-pipeline)

Summary - Key Messages

• Consistency in your processes
• Consistency in analysis quality
• Reproducibility
• Reusability
• Single and multi-script analyses are not scalable
• They do not track dependencies well
• It can be difficult to know which functions to run and in what order
• remake (or similar) can handle these issues for you

Any Questions?

https://nathaneastwood.github.io

@nathaneastwood_

https://github.com/nathaneastwood