#Dead simple AB testing at Massive Scale with Scala and Spark
GIPHY
As a small engineering team, GIPHY needs creative, lightweight solutions to deliver features for a massive global user base. Leveraging the type-safety and expressiveness of Scala along with the performance of Spark allowed us to rapidly deliver a reliable AB testing solution for our search engine.
Demand for richer content in messaging and internet communication continues to grow rapidly. As a provider of such content, GIPHY has seen exceptional growth in traffic while maintaining a relatively small engineering team. Engineering problems must be approached with the following constraints in mind:
- solutions must be designed to scale
- features need to be delivered as quickly as possible
- unnecessary architectural complications and dependencies on other teams need to minimized
- code should be as generic and extensible as possible
Search is core to GIPHY and we needed to better incorporate data into our feature decision-making process. AB + Multivariate testing was an obvious capability. This talk will cover how we designed such a solution within the above-mentioned constraints and why Scala was a natural choice. Highlights include:
- parsing and modeling your data for use with functional paradigms in mind
- abstracting user experiments and statistical attribution in a generic and immutable manner
- maximizing Spark code reliability with compile-time safety and unit-testing
- adopting Bayesian inference over more classical frequentist approaches to empower more intuitive interpretations of experiments
- how a monorepo approach enables better code reuse and unification of services code with offline analytics
This talk is designed for a range of audiences and experience levels. Familiarities with functional-programming principles, big data architectures + Spark, and statistics relevant to AB testing are useful.
As a heavy user of Spark for the last couple years, I'm very familiar with how the APIs and engine have evolved and the powerful unification of type-safety and performance optimizations in the latest Dataset API. I'm also a passionate Scala enthusiast and love how it provides a very advanced type-system and functional abstractions with the practicality of the JVM platform, a rock-solid, high-performance, library rich ecosystem. I love sharing how these features of Scala can empower very elegant, yet pragmatic solutions.
Sean Quigley has always been fascinated with the abstract process of solving logical problems. This seems to be the one common theme on his journey from a fondness for math and puzzles as a child, to studying Physics and Economics at the University of Michigan, working as a hedge-fund quant, a data scientist in Ad-tech, and now a software engineer developing solutions for web-scale analytics and search at GIPHY. In his spare time, he enjoys martial arts, electronic music, and traveling.