The Engineering Behind Friend Bubbles: 6 Surprising Insights from Meta's Reels Team

Friend Bubbles might look like a simple addition to Facebook Reels—just a ring of profile pictures showing which friends have watched or reacted to a Reel. But making that feature work smoothly for billions of users involved some of the deepest engineering challenges at Meta. In a recent episode of the Meta Tech Podcast, engineers Pascal Hartig, Subasree, and Joseph lifted the hood on the whole process, sharing how they evolved machine learning models, tackled platform-specific quirks, and ultimately unlocked a breakthrough that made everything click. Here are six surprising insights from that conversation.

1. The Illusion of Simplicity

On the surface, Friend Bubbles appears trivial: show a friend’s avatar if they’ve engaged with the same Reel. In reality, that requires real-time data pipelines that aggregate billions of interactions per second. The team had to ensure that latency stayed under a few hundred milliseconds even when a user scrolls through hundreds of Reels. Metadata like which reactions count, how to deduplicate identical reactions, and whether to prioritize close friends all demanded custom logic. The engineering team spent months just building the infrastructure to manage the sheer scale before they could even think about the user interface. This case proves that the most intuitive features often rely on the most complex backend systems.

The Engineering Behind Friend Bubbles: 6 Surprising Insights from Meta's Reels Team — Source: engineering.fb.com

2. Evolving the Machine Learning Model

The machine learning model behind Friend Bubbles didn’t start as a single algorithm. It went through several iterations. Early versions simply ranked friends by recency of interaction, but that led to stale recommendations. The team then switched to a collaborative filtering approach that considered both explicit signals (likes, shares) and implicit signals (watch time, scroll velocity). Even so, the model struggled with cold-start problems for new Reels. To fix this, they introduced a content embedding layer that learned the visual and audio patterns of Reels, allowing the system to predict friend engagement even without prior data. The final model now updates every few minutes, balancing freshness with computational efficiency.

3. iOS vs. Android: A Tale of Two Platforms

One of the biggest surprises was how differently the feature behaved on iOS compared to Android. On iOS, the animation of bubbles expanding and collapsing consumed excessive CPU, causing frame drops. The team had to reimplement the rendering using Metal, Apple’s low-level graphics API. On Android, the problem was battery drain due to constant polling to check for new friend activity. The solution there was to switch to a push-based system using Firebase Cloud Messaging. These platform-specific optimizations were critical: what works perfectly on one OS can completely break the user experience on another. The engineers documented every tweak so future feature teams can avoid the same trial-and-error.

4. The Breakthrough: Contextual Pinning

After months of tuning the model and optimizing performance, the feature still didn’t “feel right” to users. The team discovered the issue was contextual: Friend Bubbles were showing friends who had watched the Reel hours ago, making the experience feel disconnected. The breakthrough came when they introduced contextual pinning. Instead of showing any friend activity, the system now prioritizes bubbles from friends who watched the same Reel within the same session or who are currently online. This small change made the feature feel synchronous and social, dramatically increasing engagement. It was a classic case of a human insight—people want to feel they’re sharing a moment—driving a technical solution.

5. Scaling Social Discovery to Billions

Friend Bubbles is part of a larger effort to make Reels a social discovery platform, not just a feed of trending content. To scale, the team built a distributed graph system that maps friend connections to video interaction patterns. This system handles billions of edges and must serve read requests in under 50 milliseconds. They also introduced a fallback mechanism: if a user has very few friends who engage with Reels, the feature defaults to showing “trending” bubbles from the broader community. This ensures the experience never feels empty. The architecture was designed to be sharded across regions and to survive single-datacenter failures without data loss.

6. What Engineers Learned from the Podcast

The Meta Tech Podcast episode not only documented the technical journey but also revealed softer insights. For instance, Subasree emphasized the importance of cross-team collaboration—the ML engineers, infrastructure team, and UI designers had to align on weekly syncs. Joseph highlighted the value of dogfooding: team members used a beta version of Friend Bubbles for a month before launch, catching usability issues like bubble overlap and stale avatars. The episode also touched on career growth: tackling a “simple” feature turned out to be a high-impact project that stretched everyone’s skills. For any engineer, these lessons apply whether building a new feature from scratch or refining an existing one.

Friend Bubbles might look like a small social gimmick, but behind it lies a world of thoughtful engineering. From machine learning evolution to platform-specific fixes and a context-driven breakthrough, the story shows that the best features are often the hardest to build. To hear the full discussion—including the team’s future plans for social discovery in Reels—check out the Meta Tech Podcast episode wherever you get your podcasts, including Spotify and Apple Podcasts. And if you’re a developer interested in tackling challenges at this scale, visit the Meta Careers page.

Tags: