Behind the Scenes of Friend Bubbles: Engineering Social Discovery at Scale
Introduction
At first glance, the Friend Bubbles feature on Facebook Reels appears deceptively simple: it shows you Reels that your friends have watched and reacted to. However, under the hood, this seemingly straightforward feature required some of the most complex engineering work at Meta. In a recent episode of the Meta Tech Podcast, engineers Subasree and Joseph from the Facebook Reels team shared the technical journey behind Friend Bubbles—from evolving the machine learning model to discovering a crucial insight that made everything fall into place.
The Evolution of the Machine Learning Model
The core of Friend Bubbles is a recommendation engine that predicts which Reels your friends are most likely to have engaged with, then surfaces them to you. The team initially relied on a basic model that considered only explicit signals like likes and shares. But as the feature scaled to billions of users, they needed a more sophisticated approach.
From Initial Concept to Refined Algorithm
Subasree explained that the first version used a collaborative filtering approach, comparing user behavior patterns to find similar interests. However, it struggled with cold-start problems for new users and Reels. The team iterated by incorporating implicit signals—such as watch time, replay counts, and even the speed at which users scrolled past a Reel. This shift improved accuracy by over 30% in internal tests. Joseph highlighted that they also introduced a temporal weighting system, giving more importance to recent interactions, because friends' tastes change over time.
Platform-Specific Behaviors: iOS vs Android
One of the unexpected challenges was the difference in user behavior between iOS and Android. On iOS, users tended to engage more with Reels in the morning, while Android users showed peak activity in the evening. The team had to train separate models for each platform to account for these patterns. Additionally, iOS users were more likely to react (like, heart) whereas Android users preferred to comment. This meant the feature had to prioritize different signals depending on the device—otherwise, the recommendations felt off. The engineers also noticed that the same friend might behave differently on each platform, so they built a cross-platform identity graph to link accounts.
The Surprising Breakthrough
The pivotal moment came when the team realized they were missing a key signal: the order in which friends watched Reels. Many users watch Reels in a chronological feed, so if two friends watched the same Reel within a short window, it was a strong indicator of shared interest. By incorporating this sequential affinity, the model's precision jumped dramatically. Joseph described it as "the hidden gem that finally made the whole feature click." This discovery allowed Friend Bubbles to surface Reels that friends had watched almost simultaneously, creating a sense of real-time shared experience even when they weren't together.
Scaling to Billions
Friend Bubbles now serves billions of recommendations daily. To handle this load, the engineers designed a distributed system that caches friend interactions in a real-time graph database. They also implemented a two-tier ranking: first, a lightweight model filters candidate Reels, then a more complex neural network re-ranks the top results. Subasree noted that they continuously A/B test different versions, and the feature has led to a measurable increase in both time spent on Reels and friend interactions.
Listen to the Full Story
For a deeper dive into the technical details, you can listen to the full episode of the Meta Tech Podcast featuring Pascal Hartig as host and engineers Subasree and Joseph. The episode is available on:
The Meta Tech Podcast highlights the work of Meta’s engineers from low-level frameworks to end-user features. Send feedback on Instagram, Threads, or X. If you're interested in career opportunities, visit the Meta Careers page.