From 2,500 requests to 24: Rebuilding star history without the stargazers
GitHub re-engineered how star history trajectories are calculated, moving away from fetching complete lists of individual stargazer records page-by-page. The new approach reduces the required API calls for a repository timeline from 2,500 down to 24.
GitHub · 17 October 2026 · Read the original →
Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations
Meta built a system that harvests idle GPU and CPU inference capacity across its global fleet to run secondary ads recommendation models. The system relies on two elasticity levers that dynamically expand and shrink opportunistic workloads without impacting baseline latency or SLAs.
Meta · 17 October 2026 · Read the original →
TimeSeries Tiered Storage Journey: Kafka/Flink Streams to Native Cassandra Cold Reads
Netflix details the evolution of its petabyte-scale temporal event platform, transitioning from live Kafka and Flink streaming ingestion to a tiered storage backend leveraging native Cassandra cold reads. The writeup addresses the engineering required to handle queries across massive temporal datasets cost-effectively.
Netflix · 17 October 2026 · Read the original →
Inside Hyperdrive: How we solved a rare Hyperdrive MySQL caching bug
Cloudflare engineers share a detailed root-cause analysis of a subtle caching and connection pooling bug encountered while expanding Hyperdrive from Postgres to MySQL protocol support. The post highlights the challenges of building reliable stateful database connection poolers on edge networks.
Cloudflare · 17 October 2026 · Read the original →
CASTLE: training natural language search before the first real query
Airbnb addressed the cold-start problem for its natural language search engine by generating synthetic training queries and relevance labels from de-identified booking and search history before launch. The synthetic pipeline allowed the search model to perform effectively before receiving live user queries.
Airbnb · 17 October 2026 · Read the original →
Shipping Media Over QUIC at Meta Scale for Instagram Live Ingest
Meta replaced legacy video ingest protocols on Instagram Live with Media over QUIC (MoQ), unifying their live streaming network stack. The writeup details the deployment across Instagram Live ingest and how MoQ improves stream stability under variable network conditions.
Meta · 17 October 2026 · Read the original →
Bulk Executor for Amazon DynamoDB, Part 2: Lessons we learned the hard way
AWS engineers share design tradeoffs, unexpected throttling behavior, and key-space distribution gotchas encountered while building an open-source DynamoDB bulk execution tool on top of AWS Glue. The post focuses on practical workarounds for maintaining performance under strict NoSQL constraints.
AWS Architecture · 17 October 2026 · Read the original →
Another week of building complex infrastructure to fix last year's complex infrastructure.
7 stories, every Saturday
Published here every week. Follow by RSS to get it as it lands.