Meta's F3 Lightning Decouples Ads ML Feature Releases to Cut Deployment Time by 3.6x

Details

F3 Lightning decouples feature releases from shared execution, enabling independent compilation, validation, and rollback for each of the 20,000+ ads ML features.
Comparable extraction features reach production ~3.6x faster than the prior F3 framework, with no increase in infrastructure compute cost.
Four optimizations restore F3's efficiency: stable computation identifiers, a single shared execution plan, one shared memory buffer, and dense index mapping.
Stable identifiers derived from operation parameters allow the runtime to recognize and deduplicate shared work across independently compiled features.
A single execution plan built at startup via topological sort ensures shared computations run once, while a shared memory buffer eliminates cache misses from scattered allocations.
Dense index mapping replaces runtime hashing with direct array access, removing lookup overhead from the request path.
The design separates deployment boundaries from execution boundaries, preventing a single feature failure from blocking the entire release pipeline.

Key point