Abstract
Meta researchers have developed a centralized auction-based system to manage the "notification storm" in large-scale applications like Instagram. By treating notification opportunities as items in a fair-division market based on Competitive Equilibrium from Equal Incomes (CEEI), the system successfully balances competing product goals, increases user engagement, and reduces the total number of notifications sent. Production results show impact on 2.9 billion daily notifications with improved stability through first-price auction mechanisms.
Introduction
In a massive ecosystem like Instagram, users are subject to over 550 different types of notifications, ranging from social interactions like likes and comments to content updates such as story additions. Historically, these notifications were managed by independent product teams operating in silos. This decentralized approach often led to a "tragedy of the commons," where excessive notifications from multiple sources resulted in user fatigue and the eventual disabling of notifications altogether.
To address this, researchers proposed a centralized system to fairly allocate user attention, align team goals with platform health, and optimize notification volume and timing.
Methods: The Auction Framework
The core of the proposed solution is an auction approach grounded in the theory of Competitive Equilibrium from Equal Incomes (CEEI).
The Market Mechanism
In this market, different notification types (e.g., the "Stories" team) act as buyers. Each team is allocated a budget of "faux currency" to bid on the opportunity to send a notification to a specific user. A unique "Platform Bidder" acts as an auxiliary buyer, subsidizing bids for notifications that provide high global value to the platform, even if the originating team has a limited budget.
Optimization and Pacing
The system utilizes the Eisenberg-Gale Convex Program to maximize the geometric mean of all participants' utilities, ensuring "proportionality"—a fairness guarantee where every team receives a fair share of value relative to its budget. To handle the real-time, sequential arrival of notification opportunities, the researchers developed the PACE (Pace According to Current Estimated utilities) algorithm. This online algorithm uses pacing multipliers to ensure teams spend their budgets steadily throughout the day.
The study also compared First-Price (FP) and Second-Price (SP) auction formats. While SP auctions are common in digital advertising, the researchers found that FP auctions provided significantly more stable pacing multipliers in this context.
Evaluation and Results
The system was evaluated through production A/B tests on Instagram involving two groups of 77.4 million users each. The tests focused on four major notification categories: Comment Subscribed, Feed Suite Organic, Like, and Story Daily Digest.
Key Findings
- Engagement and CTR: The auction system led to measurable increases in core platform engagement metrics and notification click-through rates (CTR).
- Spam Reduction: Despite higher engagement, the total number of notifications sent decreased by approximately 0.5%, demonstrating that the system successfully filtered out low-value content.
- Operational Scale: Since its launch in July 2022, the system manages approximately 2.9 billion daily notifications.
- Auction Stability: First-price auctions reduced the variance of pacing multipliers by 10% to 50% compared to second-price auctions, making the system easier to tune and more predictable.
Conclusion
By shifting from a first-come-first-served model to a centralized, fair-division auction, Meta researchers have demonstrated a scalable way to manage user attention. The system prioritizes quality over quantity, ensuring that even smaller product features can reach users without contributing to a "notification storm." Future work will focus on automating budget discovery and extending the model to handle batched notifications.
Summary updated