Engineering JournalNew every weekday
Field notes from running Google Cloud in production at enterprise scale — Kubernetes, cloud security, DevSecOps, system architecture and AI infrastructure. Written for engineers who ship, not for slide decks. A new article every weekday, in English and German.
Start here
Eight focused engineering hubs. Each one gathers the deep-dive articles, real-world context and FAQs for a single domain — start where you want to go deepest.
Browse all topic hubs
FeaturedDeepSeek V4.1 Flash explained through its memory architecture, vendor benchmarks and worked agent costs, with practical API integration details.
Key takeawayThe architecture targets the cost of reading and retaining long agent histories.
Navicat 18 promises code insight, refactoring, data masking and AI that previews before it writes. I put it in front of a PostgreSQL table with 10 million rows and wrote down what worked, what hid, and what needs care.
A plain review of Gemini 4 Carbon, Google's new top model: the benchmark table and who ran each test, cost per task against Claude Opus 5.5 and GPT-6 Astra, why almost nobody can use it yet, how to get access, and who should care.
Ghidra MCP servers connect one decompiler to an AI agent. REA puts many analysis tools, Ghidra included, behind a single MCP server. How the two differ, what their setup changes on your machine, what can go wrong, and where the law draws lines.
Tailscale runs on WireGuard, so the real comparison is a bare tunnel against a tunnel plus a control plane. What that control plane adds, what it costs in speed and money, and what the October Tailscale SSH vulnerabilities actually change.
GPT-6.1 Terra vs Claude Sonnet 5.5 vs GPT-6.1 Sol compared on official specs, vendor benchmarks, independent Artificial Analysis results and real cost per agent turn: why Terra's half price shrinks to 44% per task, where Sonnet 5.5 still leads, what changes past 272K tokens, and which model to pick.
A JWT is a signed JSON object in three base64url parts: header, payload and signature. Anyone can decode it; only the key holder can sign it. How a JWT is built, how to verify one correctly, HS256 vs RS256 vs ES256, the attacks that still work (alg none, algorithm confusion, weak secrets, kid and jku injection) and how to decode a token without leaking it.
A PostgreSQL index is a separate structure that lets the database find rows without reading the whole table. Which type to use, from B-tree and hash to BRIN, GIN and partial or covering indexes, measured on 10 million rows with PostgreSQL 18.
What an AI gateway (LLM gateway) is, what sits inside one, how it differs from an API gateway, when you need one and what breaks in production: streaming, fallback, budgets, caching. With an open benchmark of nginx, Kong, Portkey and LiteLLM on the same mock model, a self-hosted vs managed comparison, and what the LiteLLM 1.107.0 compromise teaches about running one.
TypeScript 7 is the same language on a new native compiler written in Go. What changed between TypeScript 6 and 7, an own benchmark on type-fest, kysely and hono, every option that is now a hard error, the tooling that does not work yet, and a safe upgrade path through 6.0, with the 7.1 beta in view.
DuckDB is an in-process SQL database built for analytics: no server, columnar, vectorised, and able to query Parquet, CSV and S3 directly. What it is, a reproducible benchmark against SQLite, pandas and Polars on 11.2 million taxi trips, where it beats BigQuery and where it does not, plus Quack, DuckLake, MotherDuck and DuckDB 2.0.
Saving data and publishing an event are two writes to two systems, and no retry can make them atomic. The transactional outbox writes the event in the same database transaction and ships it afterwards. Schema, polling vs CDC with Debezium, idempotent consumers, ordering and operations.
Meta's Muse Spark Open against China's two strongest open-weight models, GLM 5.3 and Kimi K3: independent Artificial Analysis scores, Meta's own benchmark table and why its competitor numbers differ from the vendors', cost per task, the GPU memory each model needs, license fine print and safety results.
Part 1 of the NIS2 for Cloud Teams series. Article 21 lists ten minimum security measures and none of them mentions a cloud. This checklist maps each one to concrete controls in AWS, Azure and Google Cloud, separates what your provider covers from what you own, and gives a 30-day plan.
What OpenTelemetry is and is not, how traces, metrics and logs fit together, how the API, SDK, OTLP and the Collector move data, agent vs gateway deployments, zero-code instrumentation on Kubernetes, sampling, and the production mistakes that make observability expensive.