Meta gets advertising model features into production 3.6 times faster

A faulty feature used to block the entire release

Meta’s Facebook Feature Framework, or F3, is used to create, compile and serve the machine learning features behind advertising ranking models. Processing a single request can require more than 20,000 features.

F3 reduced computing costs by compiling all features into a single combined graph. However, the graph also served as the unit of deployment. If a single feature contained an invalid definition, the entire release process stopped, preventing even the valid features from reaching production.

In Meta’s example, five of seven independently created features were valid, while two were faulty. The first compilation attempt failed on one error and the second on the other, so the complete graph could only enter the production environment on the third attempt.

F3 Lightning separates deployment from execution

F3 Lightning is designed to give every feature its own deployment unit while preserving the efficiency of F3’s shared computation. The compiler treats each feature as an independent configuration. The configuration contains all the nodes required to extract the feature, in the correct order, and does not depend on another feature’s configuration.

This means a faulty feature does not prevent the others from being deployed. Features can be selected, validated, deployed and rolled back on separate schedules. Instead of operating in batches, the release process uses an ordered queue in which the priority of changes is determined by the opportunities they unlock, as well as their impact and risk.

A new feature that is not yet used by any model can be released quickly through the standard validation process, according to Meta. If a change affects a feature already used by a production model, additional model-specific tests are required beforehand.

Four approaches restore the benefits of shared computation

In the first prototypes supporting independent releases, features were also executed separately. As a result, the same reading, filtering and aggregation operations had to be performed multiple times, significantly increasing processor usage.

F3 Lightning therefore separates deployment while bringing execution back together at startup. The system builds a single execution plan from the active feature configurations, with identical computations appearing only once.

Meta highlights four technical steps. The compiler creates stable identifiers based on the operation, its parameters and its inputs. The system then creates a single execution plan for all active features. The other optimizations described include using a shared memory buffer for all features and dense indexing for direct array access.

Faster release cycles with unchanged computing costs

According to a report published by Meta on November 3, 2026, comparable features reach the production environment about 3.6 times faster with F3 Lightning than with the previous F3. The company says this is achieved without increasing the infrastructure’s computing costs.

The shorter release cycle enables more model iterations in the same amount of time. In Meta’s own advertising machine learning infrastructure, the change reduces the risk that a single faulty feature will delay the deployment of other features that have already been validated.