Architecture AI Updates: November 19, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Booking.com Details How Its ML Platform Grew From MySQL and A/B Tests to Feature Stores, Interleaving and Agent Foundations
A Booking.com principal engineer presented about 20 years of the company’s data and ML platform evolution, across data management, ML engineering and domain intelligence. The speaker says model serving grew from a 2015 Scala service into a platform for 12 or more model formats, with 480 or more models and about 400 billion predictions a day at an average latency under 20 ms, and that the team settled on precomputed features behind an online store after three attempts at feature engineering. The talk also covers interleaving, which mixes results from several models to pre-select candidates before formal A/B tests, a Hadoop sunset in which 92% of workflows failed a usage and ownership check, and early work on agent memory, agent serving, access control and agent-to-agent communication. Source
2. Martin Fowler’s Site Publishes an Article Arguing Data Contracts Give AI Agents the Metadata They Need
Simon Harrer and Arif Wider published “Data Contracts” on martinfowler.com, describing a data contract as a machine-readable, verifiable and versioned description of a dataset’s schema, quality rules, service levels, ownership and terms of use. The article is built on the Open Data Contract Standard, a YAML format governed by the Bitol project at the Linux Foundation, and covers SemVer versioning, automated checks in CI/CD and scheduled production tests, and contracts as the interface of a data product. It frames agentic AI as a driver, noting that agents treat every value as truth and need explicit metadata, and it points to an “AI context” section in version 3.2 of the standard that carries instructions, constraints and verified query answers for agents. The authors also list costs, including upkeep and the risk of over-applying contracts, and advise starting where data problems hurt most. Source
3. ByteByteGo Describes an Analyze, Measure, Improve Loop for Evaluating AI Agents Using Three Gulfs of Failure
ByteByteGo published an excerpt on AI evals that sorts failures into three gulfs: comprehension (the developer versus the real data), specification (developer intent versus prompts or agent logic) and generalization (the data versus model behavior). Each gulf calls for a different fix. The recommended loop is to read traces of agent runs to find failure modes, measure how often each occurs with automated evaluators that are either code-based rules or an LLM acting as judge, and then improve through prompt or agent changes, more examples, better retrieval or fine-tuning. The excerpt works through an email-agent example and says evaluation rigor should scale with how costly errors are and how easily users can catch them. Source