Facebook’s Distributed Recommendation System: Architecture, Algorithms, and Performance

To maintain user experience at massive scale, recommendation systems traditionally rely on machine‑learning algorithms that process complete datasets, but the rapid growth of input data makes centralized approaches infeasible, prompting the development of distributed machine‑learning methods.

Facebook’s recommendation engine now handles roughly 100 billion ratings, over 1 billion users, and millions of items—far exceeding the Netflix Prize dataset—so it adopted Apache Giraph, a distributed iterative graph‑processing platform, as its core infrastructure.

The system uses collaborative filtering, formulating the problem as matrix factorization where the user‑item rating matrix R is approximated by the product of a user matrix and an item matrix, and the optimization seeks to minimize the distance between R and its reconstruction.

Because solving matrix factorization on such a scale is computationally intensive, Facebook employs iterative algorithms such as Stochastic Gradient Descent (SGD) and Alternating Least Squares (ALS), which converge to a good solution without needing to scan every sample in each iteration.

The standard Giraph approach treats users and items as vertices and ratings as edges, causing massive network traffic (e.g., 80 TB per iteration) and uneven vertex degree, which leads to memory bottlenecks and prevents true SGD updates from being globally synchronized.

To overcome these issues, Facebook introduced a work‑to‑work message‑passing scheme: the graph is partitioned into “workers” each containing a subset of items and users; workers circulate item‑update information clockwise, ensuring that each iteration only processes internal ratings, making communication volume independent of the number of ratings and eliminating degree‑distribution problems. The system also blends SGD and ALS into a hybrid “rotating‑mix” solver.

Performance is evaluated through A/B testing, measuring metrics such as average item rating, precision for the top‑1/10/100 items, overall precision, and Root Mean Squared Error (RMSE) on a held‑out test set.

For fast top‑K recommendation, Facebook explores using a ball‑tree index to accelerate item‑vector search by 10–100×, or clustering items by feature vectors to first select promising groups and then rank within them, trading some accuracy for speed.

Experimental results released in July 2014 show that Facebook’s hybrid method on Spark MLlib’s ALS benchmark is roughly ten times faster than the standard ALS implementation and can comfortably process over 100 billion ratings.

The approach is already deployed across multiple Facebook applications (e.g., page and group recommendations), with candidate selection limited to high‑degree entities, and future work includes leveraging social graphs, automated parameter tuning, and improved partitioning strategies.