SilverTorch: Index as Model — A New Retrieval Paradigm for Recommendation Systems
Meta unveiled SilverTorch, a unified, model‑based retrieval system that replaces the traditional microservice mesh in recommendation pipelines. By embedding indexing, filtering, ANN search and neural reranking into a single PyTorch model, SilverTorch achieves 23.7× higher request throughput and 20.9× better compute‑cost efficiency versus CPU‑based baselines while improving recommendation accuracy. The architecture scales across dozens of apps, supports sub‑100 ms latency, and enables rapid iteration on retrieval features. Its research paper has been accepted to SIGIR 2026.
Migrating Data Ingestion Systems at Meta Scale
Meta has completely overhauled its massive data ingestion pipeline that extracts petabytes of social‑graph data from one of the world’s largest MySQL deployments. The new self‑managed warehouse architecture replaced customer‑owned pipelines and was migrated in three disciplined phases—shadow, reverse‑shadow, and...
Labyrinth 1.1: Making End-to-End Encrypted Backups Even More Reliable
Meta has launched Labyrinth 1.1, an upgrade to its encrypted storage protocol for Messenger. The new sub‑protocol pushes each message directly into the recipient’s encrypted backup, eliminating reliance on device‑online status. This change improves backup reliability when users lose, replace,...
How Meta Is Strengthening End-to-End Encrypted Backups
Meta has upgraded its end‑to‑end encrypted backup infrastructure for WhatsApp and Messenger with a hardware security module (HSM) based Backup Key Vault. The vault stores recovery codes in tamper‑resistant HSMs, keeping them inaccessible to Meta or cloud providers. New features...
Modernizing the Facebook Groups Search to Unlock the Power of Community Knowledge
Meta has overhauled Facebook Groups Search by deploying a hybrid retrieval architecture that combines traditional inverted‑index lookup with dense vector embeddings. The new system uses a 12‑layer, 200‑million‑parameter semantic model alongside the Unicorn lexical index, and ranks results with a...
Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale
Meta’s Capacity Efficiency Program has deployed a unified AI agent platform that automates both detection and remediation of performance regressions and optimization opportunities across its hyperscale infrastructure. By encoding senior engineers’ domain expertise into reusable skills and standardized tool interfaces,...
Post-Quantum Cryptography Migration at Meta: Framework, Lessons, and Takeaways
Meta released a detailed guide on its post‑quantum cryptography (PQC) migration, outlining a multi‑year rollout of PQ‑enabled TLS across its internal infrastructure. The company introduced a five‑tier PQC Migration Level framework—PQ‑Unaware to PQ‑Enabled—to help organizations assess and prioritize quantum‑risk mitigation....
Trust But Canary: Configuration Safety at Scale
Meta’s Configurations team explained how the company safeguards massive configuration rollouts using canary and progressive deployment techniques. The discussion highlighted health‑check metrics and monitoring signals that detect regressions early, and an incident‑review culture that focuses on system improvement rather than...
How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines
Meta built a pre‑compute engine of 50+ specialized AI agents that scanned its 4,100‑plus file, three‑repo data pipeline and produced 59 concise context files capturing tribal knowledge. This "compass" layer lifted AI coverage from roughly 5% to 100% of the...
KernelEvolve: How Meta’s Ranking Engineer Agent Optimizes AI Infrastructure
Meta unveiled KernelEvolve, an autonomous agent that automates low‑level kernel creation and tuning for its diverse AI accelerator fleet—including NVIDIA GPUs, AMD GPUs, custom MTIA silicon, and CPUs. By treating kernel optimization as a search problem, the system compresses weeks...
Meta Adaptive Ranking Model: Bending the Inference Scaling Curve to Serve LLM-Scale Models for Ads
Meta unveiled its Adaptive Ranking Model, a request‑centric inference stack that lets the company run LLM‑scale recommendation models for ads with sub‑second latency. The system combines inference‑efficient scaling, deep model‑hardware co‑design, and a multi‑card GPU serving layer to handle up...
Ranking Engineer Agent (REA): The Autonomous AI Agent Accelerating Meta’s Ads Ranking Innovation
Meta introduced the Ranking Engineer Agent (REA), an autonomous AI system that runs end‑to‑end machine‑learning experiments for ads ranking. REA generates hypotheses, launches training jobs, debugs failures, and iterates without continuous human oversight, using a hibernate‑and‑wake cycle for multi‑day workflows....
Patch Me If You Can: AI Codemods for Secure-by-Default Android Apps
Meta’s Product Security team unveiled a two‑pronged solution to harden Android apps at scale: secure‑by‑default frameworks that wrap risky OS APIs, and generative‑AI‑driven codemods that automatically migrate existing code to those frameworks. The AI system can propose, validate, and submit...
Investing in Infrastructure: Meta’s Renewed Commitment to Jemalloc
Meta announced a renewed focus on jemalloc, the high‑performance memory allocator that underpins its infrastructure. The company has unarchived the open‑source repository and outlined a roadmap to cut technical debt, modernize the codebase, and add features such as a stronger...