Behind the Scenes of Meta's Friend Bubbles: Engineering Social Discovery at Scale

The Feature That Looks Simple, Yet Required Deep Engineering

At first glance, Meta’s Friend Bubbles feature appears deceptively straightforward. It simply highlights Reels that your friends have watched and reacted to. But as any seasoned engineer knows, the most user-friendly features often demand the most complex underlying work. In a recent episode of the Meta Tech Podcast, software engineers Subasree and Joseph from the Facebook Reels team peeled back the curtain on what it really took to bring Friend Bubbles to life. They discussed the evolution of the machine learning model powering the feature, the surprising behavioral differences between iOS and Android users, and the unexpected breakthrough that finally made the entire concept click.

Why a “Simple” Feature Requires Heavy Lifting

Friend Bubbles had to solve a fundamental problem: social discovery at scale. With billions of Reels being watched every day, how do you surface the ones that are truly relevant to each user based on their friends’ activity? The engineering team needed to build a system that could:

• Identify which Reels a user’s friends have watched and reacted to (like, comment, share).
• Filter and rank those Reels to show the most engaging ones without overwhelming the user.
• Deliver this personalized feed quickly, across billions of devices, while respecting privacy and data constraints.

The technical challenges multiplied as the team scaled from small test groups to the entire global user base. Every millisecond of latency and every degree of personalization mattered. The solution required a deeply integrated machine learning pipeline that could learn from user behavior in real time.

Evolving the Machine Learning Model for Friend Bubbles

The ML model behind Friend Bubbles didn’t start as a single algorithm. It evolved through multiple iterations. Early versions tried to simply aggregate friend activity—show Reels that had the most friend reactions. But that approach lacked personalization: two users with the same set of friends would see identical bubbles. The team needed a model that could weight friend interactions differently based on how close a user is to each friend, how recently the interaction happened, and the type of reaction (e.g., a comment might indicate deeper interest than a simple like).

Joseph explained that they eventually built a multi‑task learning framework that jointly predicted both the likelihood of a user engaging with a Reel and the likelihood that their friends would also engage. This shared representation allowed the model to capture subtle social signals—like a friend who always watches full Reels versus one who quickly skips—and adjust ranking accordingly.

iOS vs Android: Uncovering Behavioral Asymmetries

One of the most surprising findings during development was how differently iOS and Android users interact with Reels and their friends’ activity. Subasree noted that on iOS, users tend to engage more with visual notifications and previews, while Android users are more likely to scroll through a list of text‑driven suggestions. This meant the Friend Bubble UI had to adapt: iOS received richer, animated bubbles that hinted at the content, while Android got a cleaner, more informational layout.

But the differences went beyond UI. The underlying friend graph dynamics varied by platform. iOS users generally have smaller, more tightly knit groups of friends, making it easier to surface highly relevant Reels. Android users, on the other hand, often have larger networks with weaker ties, requiring the model to prioritize stronger connections even if that meant showing fewer bubbles. The team had to train separate platform‑specific models to handle these nuances, increasing engineering complexity but improving user satisfaction.

The Breakthrough That Made Friend Bubbles Click

Despite significant progress, the feature still felt “off” during early internal testing. Users reported that the bubbles seemed random or uninteresting. The turning point came when Subasree and Joseph discovered that timing was everything. A friend could have watched a Reel hours ago, but if that Reel was only just going viral, showing it now felt stale. Conversely, showing a Reel that a friend watched just seconds ago—even if it wasn’t popular—felt exciting and fresh.

They implemented a time‑decay weighting system that gave more prominence to recently watched Reels, but also introduced a “trending among friends” signal that boosted Reels with a sudden spike in friend activity. This hybrid approach allowed the model to balance recency and relevance. The effect was immediate: user engagement with Friend Bubbles doubled in the first week after rollout.

Lessons for Engineers Building Social Features

The story of Friend Bubbles offers valuable takeaways for anyone building social discovery features at scale:

1. Don’t underestimate “simple” features. They often require the deepest engineering work.
2. Personalization must account for platform differences. What works for iOS may fail for Android.
3. Model evolution is a journey. Start simple, iterate based on data, and don’t be afraid to abandon early approaches.
4. Human‑like intuition (like “timing matters”) can be encoded mathematically. The surprising discovery about recency was key.

For more insights from the Meta Tech Podcast, including the full conversation with Subasree and Joseph, check out the episode on Spotify, Apple Podcasts, or Pocket Casts. Follow Meta Engineering on Instagram, Threads, or X for updates. And if you’re inspired to build the next generation of social discovery, explore career opportunities at Meta Careers.