Engineering Social Discovery: Inside Meta's Friend Bubbles Feature

In a recent episode of the Meta Tech Podcast, software engineers Subasree and Joseph from the Facebook Reels team peeled back the layers on Friend Bubbles, a feature that at first glance seems simple: it highlights Reels your friends have watched and reacted to. But beneath the surface lies a complex web of engineering challenges, from evolving machine learning models to handling behavioral differences between iOS and Android users. The conversation reveals the surprising discovery that finally made the feature click and how it scales to billions of users. For anyone who has ever underestimated a “simple” feature, this deep dive is a must-read.

What is the Friend Bubbles feature and why is it significant?

Friend Bubbles is a social discovery feature that highlights Reels your friends have engaged with—either watched or reacted to. On the surface, it appears straightforward: you see a bubble indicating that a friend has interacted with a particular Reel. However, this simplicity belies the profound engineering effort required. The significance lies in its ability to personalize the social experience at massive scale. By surfacing content your friends find interesting, it creates a more engaging, community-driven feed. This feature not only enhances user interaction but also strengthens the social fabric of the platform. It leverages existing relationships to drive content discovery, making the Reels experience feel more relevant and connected. The engineering team had to overcome hurdles related to real-time data processing, privacy constraints, and recommendation relevance to bring this feature to life.

Engineering Social Discovery: Inside Meta's Friend Bubbles Feature — Source: engineering.fb.com

What were the main engineering challenges behind Friend Bubbles?

The primary challenges revolved around scalability, latency, and personalization. With billions of users, tracking each friend interaction in real time while maintaining low latency for the end user required a robust infrastructure. Another major hurdle was ensuring the feature respected privacy settings and did not inadvertently expose sensitive data. Additionally, the team had to design a system that could handle the unpredictable nature of social graphs—friends lists can change rapidly, and signals like “watched” or “reacted” needed to be processed almost instantaneously. The engineers also faced the classic recommendation dilemma: balancing novelty with relevance. They needed to surface content that friends actually cared about without overwhelming users with too many bubbles. Finally, platform differences between iOS and Android introduced separate sets of constraints, from memory management to API availability, which required tailored solutions.

How did the machine learning model evolve for this feature?

The machine learning model underpinning Friend Bubbles underwent significant evolution. Initially, the team experimented with a simple collaborative filtering approach that recommended Reels based on aggregate friend interactions. However, this proved too generic and failed to capture the nuance of individual friend relationships. They then shifted to a graph neural network (GNN) architecture, which could model the complex social graph more effectively. The GNN learned embeddings for users and Reels, capturing not just direct friend connections but also second-degree relationships. Another key evolution was incorporating temporal dynamics—the model became sensitive to recency, understanding that a friend’s interaction from yesterday might be more relevant than one from a month ago. Through iterative experiments, they discovered that a hybrid model combining content-based features (e.g., Reel topic) with social signals yielded the best results. The final model is continuously retrained to adapt to shifting user behaviors.

What surprising discovery made the feature click?

The breakthrough came when the engineers realized that timing and context were just as important as the content itself. Initially, Friend Bubbles showed interactions from any friend at any time, leading to mixed user engagement. Through A/B testing, they noticed a striking pattern: users were far more likely to engage with a bubble if the friend’s interaction happened within the last few hours and if the friend was someone they regularly interacted with. Another surprising finding was that reacted-to Reels (e.g., likes, comments) drove significantly more engagement than merely watched Reels. The team pivoted to emphasize recent reactions from close friends, and engagement metrics soared. This insight led to a redesigned ranking algorithm that prioritized timeliness and relationship strength over raw quantity. The discovery underscored the importance of behavioral data and challenged the team’s assumptions about social discovery.

How did user behavior differ between iOS and Android?

The team observed notable differences in how iOS and Android users interacted with Friend Bubbles. iOS users tended to engage more with the social context, often tapping on bubbles to see which specific friend reacted. In contrast, Android users were more likely to simply watch the Reel without exploring the social layer. This divergence influenced the user interface design: on iOS, the team added richer tooltips and friend profile previews, while on Android they simplified the interaction to reduce friction. Another difference was in notification preferences—iOS users were more receptive to push notifications about friend activity, whereas Android users preferred in-app badges. These variations required separate optimization strategies, from different API endpoints to distinct caching policies on each platform. The engineers had to ensure that while the core logic remained consistent, the user experience felt native to each ecosystem.

How does Friend Bubbles scale to billions of users?

Scaling Friend Bubbles required a multi-pronged approach. First, the team built a distributed graph processing pipeline that can compute friend interactions in near real-time using clusters of servers. They employed data sharding to distribute the social graph across many nodes, ensuring that no single server becomes a bottleneck. For prediction, they used a cascade model: the top-N most relevant Friend Bubbles are precomputed for each user and served via a lightweight recommender system. This reduces the load on the ML model during peak times. Additionally, they implemented adaptive serving—the system adjusts the number of bubbles displayed based on device performance and network conditions. On the backend, they rely on Zstandard compression for data transfer and advanced caching layers to minimize latency. Continuous monitoring and auto-scaling ensure the system handles spikes, such as during viral Reel events, without degradation.

What role did the Meta Tech Podcast play in sharing this story?

The Meta Tech Podcast served as the medium for the team to share the engineering journey behind Friend Bubbles publicly. Hosted by Pascal Hartig, the episode featured engineers Subasree and Joseph, who walked listeners through the feature’s development from concept to deployment. The podcast format allowed for in-depth technical discussions that written documentation might not capture—e.g., the iterative design process, the “eureka” moments, and the personal anecdotes from the team. By making this episode available on platforms like Spotify, Apple Podcasts, and Pocket Casts, Meta provided transparency into its engineering culture and highlighted the challenges of building seemingly simple features at scale. The podcast also serves as a recruitment tool, showcasing the kind of impactful work engineers can do at Meta. It invites feedback from the community and reinforces Meta’s commitment to sharing knowledge across the industry.