Meta's F3 Lightning ships ads ML features ~3.6x faster

Meta's original F3 framework, first presented at AI @ Scale 2020, compiled all ads ranking features into a single merged graph to share computation. That design cut execution costs but made the whole graph the release unit, so one invalid definition held back every other feature. As the feature count grew, release speed eroded despite repeated improvements.

F3 Lightning separates the release boundary from the execution boundary. Each feature is compiled into its own configuration and released through an ordered queue, while startup-time optimizations—stable computation identifiers, a single execution plan, one shared memory buffer, and dense index mapping—restore the shared-computation efficiency. The result is per-feature rollback without the CPU penalty that early prototypes incurred.

Meta notes three lessons from the project: deployment and execution boundaries should be separate; work that never changes between requests belongs at startup; and two designs can run the same operations in the same order yet differ substantially in CPU cost depending on where results live and how many copies exist. The ~3.6x faster release cycle, with no increase in infrastructure compute cost, is the outcome Meta reports.