F3 Lightning: How Meta Accelerates Ads ML Features to Production at Scale

Why it matters

Meta has introduced F3 Lightning, an infrastructure layer built on its existing Facebook Feature Framework, known as F3, which handles authoring, compiling, and serving the machine learning features behind ads ranking models. F3 originally merged every feature into a single graph to share computation and cut costs, but that merged graph also served as the release unit, meaning one faulty feature definition could block all others from shipping. As the framework grew from a few hundred features to tens of thousands, such failures became more frequent, and expensive model-related tests delayed entire releases. F3 Lightning addresses this by making each feature an independently compiled, validated, deployed, and rolled-back configuration, so the release path becomes an ordered queue rather than an all-or-nothing batch. The main engineering challenge was preserving F3's efficiency once features stopped sharing a merged graph, since early prototypes raised CPU usage. Meta solved this with four optimizations: stable computation identifiers, a single shared execution plan, one shared memory buffer, and dense index mapping. Comparable extraction features now reach production faster, with no added infrastructure compute cost.