Meta Engineers Unlock Social Discovery at Scale: Inside the Billion-User Reel Friend Bubbles Feature

Breaking: Meta's Friend Bubbles feature for Reels now reaches over a billion users, but the engineering journey was anything but simple.

The feature, which surfaces Reels that friends have watched and reacted to, required deep machine learning and platform-specific optimizations to scale.

According to an episode of the Meta Tech Podcast, software engineers Subasree and Joseph from the Facebook Reels team revealed that the biggest breakthrough came from a surprising discovery about user behavior patterns.

“On the surface, Friend Bubbles looks trivial—just showing what friends engaged with. But the infrastructure needed to surface those connections in real time across billions of users was a massive engineering challenge,” said Subasree, engineer on the Reels team.

Background

Meta originally launched Reels in 2020 as a competitor to TikTok. Friend Bubbles emerged from a need to boost social discovery without relying solely on algorithmic recommendations.

Earlier tests revealed different behavioral patterns on iOS versus Android, forcing the team to build separate model pipelines.

The feature uses a specialized machine learning model that factors in friend interactions, watch history, and reaction patterns to determine which Reels to surface.

What This Means

For users, Friend Bubbles transforms Reels from a purely algorithmic feed into a socially curated experience. Sharing reactions becomes passive—just by watching, you help friends discover content.

For Meta, this represents a strategic shift: leveraging the social graph to increase engagement and time spent on Reels, which now competes directly with TikTok and YouTube Shorts.

The engineering approach—particularly the machine learning model evolution—offers a blueprint for other social platforms aiming to blend algorithmic and social signals.

From Failure to Breakthrough: The Engineering Journey

Early prototypes of Friend Bubbles suffered from low-quality recommendations. The ML model initially struggled to distinguish between Reels friends merely scrolled past versus those they truly engaged with.

“We realized the model needed to incorporate implicit signals—like pause duration and rewatch behavior—not just explicit likes. That changed everything,” Joseph explained.

The team also optimized for mobile hardware constraints, reducing model size by 40% without sacrificing accuracy. This allowed Friend Bubbles to run on-device for faster delivery.

iOS vs. Android: Different Rules, Same Goal

An unexpected hurdle: user behavior varied significantly between platforms. iOS users tended to interact more with friend-suggested Reels, while Android users preferred algorithm-based discovery.

To solve this, the engineers built two separate ranking models—one per platform—trained on platform-specific data. The result was a 15% lift in engagement across both ecosystems.

Scaling to Billions: The Infrastructure Layer

Friend Bubbles relies on a real-time graph processing system that updates friend interaction states within milliseconds. Meta's custom database, TAO, handles the billions of relationship edges.

Caching strategies were redesigned to prevent “cold start” issues for new users, notably in regions where Reels adoption is still growing.

Listen to the full episode of the Meta Tech Podcast on Spotify, Apple Podcasts, or Pocket Casts.