Netflix Tech Blog

Sequential Testing Keeps the World Streaming Netflix Part 2: Counting Processes

thumbnail

Table of Contents

  1. Introduction
  2. Counting Processes in Sequential Testing
  3. Case Study 1: Drop in Successful Title Starts
  4. Case Study 2: Increase in Abnormal Shutdowns
  5. Case Study 3: Increase in Errors

Introduction

In this blog, we discuss the methodology of sequential testing for count metrics, as outlined in the NeurIPS paper "Anytime Valid Inference for Multinomial Count Data". This methodology allows for real-time detection of issues in count metrics, enabling prompt corrective actions.

Counting Processes in Sequential Testing

Sequential testing for count metrics involves incrementing a function by 1 whenever a new event occurs. This process is defined by an intensity function, allowing for the comparison of treatment and control groups in real-time. This sequential methodology is novel and efficient, requiring only two integers to test hypotheses.

Case Study 1: Drop in Successful Title Starts

A case study presented a decrease in successful title start events from treatment devices compared to control devices. By monitoring the count of title start events in real-time, the sequential testing methodology quickly identified the issue.

Case Study 2: Increase in Abnormal Shutdowns

An increase in abnormal shutdown events from treatment devices was detected using the sequential testing methodology. By comparing counts of abnormal shutdowns between treatment and control groups, the methodology rapidly pinpointed the problem for corrective action.

Case Study 3: Increase in Errors

Netflix also monitored the count of errors produced by treatment and control devices. An increase in 3.1.18 errors from treatment devices was swiftly identified using the sequential testing approach, showcasing the effectiveness of real-time count metric analysis.

In conclusion, the sequential testing methodology for count metrics offers a powerful tool for detecting issues promptly and facilitating data-driven decision-making in streaming services like Netflix.