The Engineering Behind Facebook Reels' Friend Bubbles: From Simple Idea to Global Scale

When a Simple Feature Demands Complex Engineering

At first glance, the Friend Bubbles feature on Facebook Reels appears deceptively straightforward: it highlights Reels that your friends have watched and reacted to. Yet, as often happens in software development, the most elegant user-facing features can require the deepest engineering work. In a recent episode of the Meta Tech Podcast, engineers Subasree and Joseph from the Facebook Reels team sat down with host Pascal Hartig to pull back the curtain on what it really took to bring Friend Bubbles to life—and to billions of users worldwide.

The Engineering Behind Facebook Reels' Friend Bubbles: From Simple Idea to Global Scale — Source: engineering.fb.com

The Challenge of Building Social Discovery

Friend Bubbles isn't just about surfacing popular content; it's about scaling social discovery—helping people find Reels their friends care about, even when those friends are spread across different devices, time zones, and network conditions. The team needed to build a system that could:

Aggregate friend activity in near real time.
Rank Reels based on social signals (watches, likes, shares).
Personalize the bubble display for each user.
Handle billions of interactions daily with low latency.

All of this had to feel instantaneous and effortless to the user. That required rethinking existing infrastructure and introducing new machine learning models.

Machine Learning Model Evolution

The Friend Bubbles feature went through several iterations of its machine learning model. Early versions relied heavily on collaborative filtering—essentially, if two friends watched similar Reels, the system would suggest those Reels to each other. But the team quickly discovered that friend affinity isn't always linear. A person might watch a Reel because a close friend liked it, but ignore the same suggestion from an acquaintance.

To capture these nuances, the engineers developed a graph-based neural network that models the strength and type of each friendship connection. The model learned to weigh signals differently: a “heart” reaction from a close friend counted more than a simple view from a distant friend. This evolution required months of training data collection, A/B testing, and iterative refinement.

Platform Differences: iOS vs Android Behaviors

One of the most surprising hurdles came from the different behaviors between iOS and Android users. The team expected the feature to work identically across platforms, but real-world usage revealed unexpected patterns. For example:

iOS users tended to watch more Reels in short bursts, often during commutes or waiting in line. Their friend activity data arrived in clusters, making it challenging to update bubbles without overwhelming the UI.
Android users exhibited more consistent engagement throughout the day, but their devices had varied network conditions and battery constraints, requiring optimizations to avoid draining resources.

The engineers had to adapt the model's update frequency and data compression techniques per platform. They also introduced client-side caching strategies that preserved bubble information even when network connectivity was intermittent.

The Surprising Discovery That Made It Click

After months of tuning, the feature still wasn’t performing as expected. Then came a breakthrough: the team realized that timing mattered far more than they had anticipated. Users didn’t just want to know what friends had watched—they wanted to know what friends were watching right now or very recently. The original model gave equal weight to activity from the past week, diluting the sense of shared real-time experience.

By introducing a temporal decay function that heavily weighted activity from the last few hours, engagement skyrocketed. The Friend Bubbles suddenly felt alive and connected. This simple change—prioritizing recency over volume—transformed the feature from a passive recommendation into an active social conversation starter.

Scaling Social Discovery to Billions

Once the model and platform quirks were resolved, the team turned to the monumental task of scaling. Friend Bubbles must handle billions of Reel views and friend interactions every day. The infrastructure relies on:

Distributed stream processing to capture friend activity in real time.
Graph databases to traverse friendship connections efficiently.
Caching layers that serve precomputed bubble candidates to reduce latency.

The engineers also built a feature flagging system to roll out updates gradually, monitoring metrics like daily active users, Reel watch time, and friend interaction rates. Each change was tested on a small percentage of users before global release. This careful approach allowed the team to iterate without disrupting the experience for the broader community.

Learn More: The Full Story on the Meta Tech Podcast

If you've ever underestimated a “simple” feature, this episode is for you. You can download or listen to the podcast below, or find it on your favorite platform:

The Meta Tech Podcast highlights the work of Meta’s engineers at every level – from low-level frameworks to end-user features. Send your feedback on Instagram, Threads, or X. And if you’re interested in building the next generation of social discovery, visit the Meta Careers page.

Originally published on the Engineering at Meta blog.