29 August 2026

12 new posts across 7 engineering blogs — 7 worth your time.

One week of Engineering Reading, 7 stories, as published.

How we saved 100 terabytes of memory by optimizing 1.1.1.1’s DNS cache

Cloudflare optimized its DNS cache platform, Big Pineapple, which stores over 250 billion entries. By implementing five successive memory storage changes, they reduced the per-entry memory footprint by over 50%, saving 100 terabytes of RAM across their fleet.

Why it matters — At massive scale, saving even a single byte per cache entry yields significant aggregate memory savings, demonstrating the value of micro-optimizations in data structures.

Cloudflare · 29 August 2026 · Read the original →

Infrastructure

MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet

Meta introduced MetaRoCE, a clean-sheet RDMA transport protocol designed specifically for AI workloads on commodity Ethernet. The release includes the protocol specification, a reference software implementation, and a compliance test suite to help scale GPU-to-GPU data transfers.

Why it matters — Custom RDMA protocols optimized for commodity Ethernet can prevent GPU idle time during large-scale AI training without requiring proprietary networking hardware.

Meta · 29 August 2026 · Read the original →

Infrastructure

MTIA 300: Meta’s First Training Chip with Built-in NICs and Communication-Offloading Engines

Meta details the MTIA 300, its custom training and inference accelerator optimized for recommendation models. The chip features built-in NIC chiplets and a co-designed communication library (HCCL) to offload communication overhead and outperform general-purpose GPUs.

Why it matters — Integrating network interface cards directly onto accelerator chips reduces communication bottlenecks in distributed AI training workloads.

Meta · 29 August 2026 · Read the original →

Data & ML

Scaling Conditional Learned Retrieval for Pinterest Home Feed

Pinterest discusses scaling candidate generation for its home feed using a Conditional Learned Retrieval approach. The system improves on traditional two-tower retrieval models to handle large-scale User-to-Pin matching.

Why it matters — Moving beyond standard two-tower retrieval models to conditional learned retrieval allows recommendation systems to scale candidate generation more dynamically.

Pinterest · 29 August 2026 · Read the original →

Data & ML

MAPS: Netflix’s Multimodal Asset Personalization at Scale

Netflix describes MAPS (Multimodal Asset Personalization at Scale), a system that personalizes visual assets like artwork and video previews for users. The system automates the selection and delivery of these cues to optimize user discovery and engagement.

Why it matters — Personalizing user interface assets at scale requires a multimodal approach that balances content metadata, user preferences, and real-time delivery constraints.

Netflix · 29 August 2026 · Read the original →

Practice & Craft

Project Lighthouse — Part 3: Introducing project-lighthouse-anonymize

Airbnb has open-sourced its privacy-preserving anonymization code under Project Lighthouse. The release includes scalable algorithms and data quality frameworks designed to protect user privacy while maintaining utility for data analysis.

Why it matters — Combining scalable anonymization algorithms with data quality frameworks allows organizations to protect user privacy without destroying the utility of the underlying datasets.

Airbnb · 29 August 2026 · Read the original →

Infrastructure

How a global payment processor preserved AWS RAM shares and Lake Formation permissions during an AWS Organizations migration

This post outlines how to safely migrate AWS accounts between organizations while preserving AWS Resource Access Manager (RAM) shares and Lake Formation permissions. It addresses the challenges of maintaining resource trust boundaries when an account leaves its parent organization.

Why it matters — Migrating AWS accounts across organizational boundaries requires careful sequencing to prevent the automatic revocation of RAM shares and data lake permissions.

AWS Architecture · 29 August 2026 · Read the original →

Designing custom AI silicon is nice, but surviving an AWS Organization migration is the real test.

7 stories, every Saturday

Published here every week. Follow by RSS to get it as it lands.

Subscribe by RSS All issues

← Previous issue Next issue →