22 August 2026

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

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

20× the CI traffic without getting slower: How we rebuilt Git serving at Datadog

Datadog built a tool called gitretriever to handle a 20-fold increase in CI Git traffic. The new system maintains low latency while reducing backend CPU usage. This allowed the platform to scale efficiently under heavy load.

Why it matters — A custom Git serving layer can handle a 20x increase in CI traffic while reducing backend CPU usage and maintaining low latency.

Datadog · 22 August 2026 · Read the original →

Infrastructure

A Tale of Two Flink Autoscalers

Netflix currently runs two Flink autoscalers in production, which is one more than they want. The first was built in-house years ago due to a lack of mature platform options, while the second is an Apache Flink community tool that scales workloads the homegrown system cannot support. The team is running both as they transition.

Why it matters — Running a homegrown autoscaler alongside a community-built alternative allows a team to scale workloads beyond original design limits during a gradual production migration.

Netflix · 22 August 2026 · Read the original →

Infrastructure

A revisit of remote Spectre attacks on Cloudflare Workers

Cloudflare re-evaluated remote Spectre attacks against its Workers platform to see if newer stabilization techniques pose a threat. The company previously deployed a defense called Dynamic Process Isolation (DyPrIs) to isolate suspicious scripts into separate processes. This post revisits the effectiveness of that defense against modern attack methods.

Why it matters — Dynamic Process Isolation (DyPrIs) can mitigate remote Spectre attacks by identifying malicious-looking scripts and isolating them into separate processes.

Cloudflare · 22 August 2026 · Read the original →

Data & ML

How we knew COVID was over (and what our models had to unlearn)

Airbnb's Forecasting Data Science team manages continuous forecasts for demand, bookings, and cancellations across thousands of markets. This post discusses the decision-making process behind when to retrain, rebuild, or leave forecasting models alone. It specifically details how their models had to unlearn pandemic-era trends as COVID-19 ended.

Why it matters — Forecasting models must be actively managed to unlearn anomalous historical trends, such as pandemic-era disruptions, to maintain accuracy across thousands of distinct markets.

Airbnb · 22 August 2026 · Read the original →

Architecture

Build a unified AI agent architecture with DynamoDB and Bedrock

Building AI agents often results in a fragmented data architecture where operational data is stored in DynamoDB and vector embeddings reside in a separate vector store. This duplication leads to higher infrastructure costs, synchronization complexity, and stale retrieval results. The post explores how native vector search in DynamoDB addresses these issues.

Why it matters — Consolidating operational data and vector embeddings into a single database with native vector search eliminates data synchronization lag and reduces infrastructure costs.

AWS Architecture · 22 August 2026 · Read the original →

Architecture

Consistency is the new latency: AI at the data layer

As AI applications transition from reactive bots to autonomous agents, their reliability becomes heavily dependent on the speed and accuracy of the underlying data layer. Many current AI architectures rely on a flawed assumption regarding data layer consistency. This post examines how data consistency issues impact the reliability of AI agents.

Why it matters — The reliability of autonomous AI agents is bound to the consistency and speed of the underlying data layer, making stale reads a critical failure point.

AWS Architecture · 22 August 2026 · Read the original →

Infrastructure

Improving infrastructure efficiency for growing demand in the age of AI

As demand for AI grows, engineering teams face physical constraints in energy, cooling, hardware availability, and physical space. Rather than simply building more data centers, teams must focus on improving the efficiency of existing infrastructure. This post addresses how to navigate these physical limitations.

Why it matters — Scaling AI infrastructure requires optimizing existing hardware and energy efficiency to work within physical constraints like cooling, power, and space.

Dropbox · 22 August 2026 · Read the original →

No matter how advanced the AI gets, we are still just debugging autoscalers.

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 →