Meta’s ‘Friend Bubbles’ for Reels: The Billion-User Scale Social Discovery Feature That Almost Didn’t Work
Meta engineers detail the complex engineering behind Friend Bubbles on Reels, a social discovery feature that highlights friends' watched and reacted Reels, in a new podcast episode.
Meta Engineers Reveal the Hidden Complexity Behind Friend Bubbles
Meta has pulled back the curtain on one of its most deceptively simple features: Friend Bubbles for Reels. The feature, which shows users which Reels their friends have watched and reacted to, required breakthrough engineering to scale to billions of users, engineers revealed in a new podcast.
"On the surface, Friend Bubbles looks trivial—just icons and counters," said Subasree, a software engineer on the Facebook Reels team. "But building a system that can serve billions of users in real-time with accurate friend activity was one of the hardest problems we've solved."
Background: The Feature That Hides a Billion-Dollar Problem
Friend Bubbles is a social discovery layer within Reels—Meta’s short-form video platform competing with TikTok. It displays circular avatars of friends who have interacted with a Reel, and when tapped, reveals their reaction (like, comment, share).
But behind that simple interface lies a massive real-time data pipeline. The feature must aggregate friend activity across millions of interactions per second, deduplicate, and personalize the display for each user—all while keeping latency under 200 milliseconds.
What They Found: iOS vs. Android & a Surprising Breakthrough
During development, the team discovered stark differences between iOS and Android user behavior. "iOS users tend to react in bursts, while Android users spread their interactions out over time," explained Joseph, another engineer on the team. "That forced us to build separate caching strategies for each platform."
But the real breakthrough came from a surprising source: the machine learning model. The team initially built a complex neural network to predict friend relevance, but it performed poorly at scale. "We realized the simplest signal—recency of interaction—was 90% as effective as the full model," Subasree noted. "We stripped the model down to a lightweight ensemble of decision trees, which cut latency by 60% and doubled feature adoption."
What This Means for Social Discovery at Scale
The Friend Bubbles case study offers lessons for any platform building social discovery features. The key takeaway: simplicity in user interface often demands complexity under the hood—but not in the way engineers expect. The optimal solution may come from minimizing, not maximizing, ML model size.
For Meta, the feature has already boosted Reels engagement by 12% among users who see friend activity. The company is now exploring ways to apply the same lightweight model architecture to other discovery surfaces, such as Instagram Explore and Facebook Watch.
Internal Anchor Links
The full technical deep-dive is available on the Meta Tech Podcast. Listeners can stream the episode on Spotify, Apple Podcasts, or Pocket Casts.