I have a machine learning model that takes some time to train. Data pre-processing and model fitting can take 15–20 minutes. That’s not so bad, but I also want to tune my model to make sure I’m using the best hyper-parameters. With 16 different combinations of hyperparameters and 5-fold cross-validation, my 20 minutes can become a day or more. Metaflow is an open-source tool from the folks at Netflix that can be used to make this process less painful.

AWS has announced support for container images for their serverless computing platform Lambda. AWS doesn’t provide an R runtime for Lambda, and this was the excuse I needed to finally try to make one. An R runtime means that I can take advantage of AWS Lambda to put my R functions in the cloud. I don’t have to worry about provisioning servers or spinning up containers — the function itself is the star.

Locking down R package dependencies and versions is a solved problem, thanks to the easy-to-use renv package. System dependencies — those Linux packages that need to be installed to make certain R packages work — are a bit harder to manage. Option 1: Hard-coding The easiest option is to hard-code the system dependencies. I did this recently when I was creating a Dockerfile for a very simple Plumber API: RUN apt-get update -qq && apt-get -y --no-install-recommends install \ make \ libsodium-dev \ libicu-dev \ libcurl4-openssl-dev \ libssl-dev My Dockerfile used only three R packages and so its system dependencies were not complicated.

I’ve set myself an ambitious goal of building a Kubernetes cluster out of a couple of Raspberry Pis. This is pretty far out of my sphere of knowledge, so I have a lot to learn. I’ll be writing some posts to publish my notes and journal my experience publicly. In this post I’ll go through the basics of Kubernetes, and how I hosted a Plumber API in a Kubernetes cluster on Google Cloud Platform.

It’s no secret that I love R and begrudgingly use Python. But there’s a another option for data science, and it promises the speed of C with the ease of use of R/Python. That language is Julia, and it’s a delight to use. I took some time to learn the basics, and I’m sharing my impressions here. Julia is not the most popular language in the world Before I go on, there’s one thing I want to stress here: Julia is not as popular as Python or R for doing stuff with data.

drake is a package for orchestrating R workflows. Suppose I have some data in S3 that I want to pull into R through a drake plan. In this post I’ll use the S3 object’s ETag to make drake only re-download the data if it’s changed. This covers the scenario in which the object name in S3 stays the same. If I had, say, data being uploaded each day with an object name suffixed with the date, then I wouldn’t bother checking for any changes.

After I posted my efforts to use MLflow to serve a model with R, I was worried that people may think I don’t like MLflow. I want to declare this: MLflow is awesome. I’ll showcase its model tracking features, and how to integrate them into a tidymodels model. The Tracking component of MLflow can be used to record parameters, metrics and artifacts every time a model is trained. All of this information is presented in a very nice user interface.

There’s always a need for more tidymodels examples on the Internet. Here’s a simple machine learning model using the recent coffee Tidy Tuesday data set. The plot above gives the approach: I’ll define some preprocessing and a model, optimise some hyperparameters, and fit and evaluate the result. And I’ll piece all of the components together using targets, an experimental alternative to the drake package that I love so much. As usual, I don’t care too much about the model itself.

I’m obsessed with how to structure a data science project. The time I spend worrying about project structure would be better spent on actually writing code. Here’s my preferred R workflow, and a few notes on Python as well. The R package workflow In R, the package is “the fundamental unit of shareable code”. At rstudio::conf 2020, Hadley gave a rule of thumb for when to create a package, which I’ll paraphrase: “When you copy and paste a block of code three times, make a function.

Suppose I want a function that runs some setup code before it runs the first time. Maybe I’m using dplyr but I haven’t properly declared all of my dplyr calls in my function, so I want to run library(dplyr) before the actual function is run. Or maybe I want to install a package if it isn’t already installed, or restore a renv file, or any other setup process. I only want this special code to run the first time my function is called.