SoftwareEngineeringLinks

Mercury is quietly shipping millions of lines of Haskell while everyone else is racing toward agentic everything. Pair that with a few sharp lessons on premature scale, interpreter speed, and why Postgres can burn your disks even when nothing updates, and this one pays for itself.

🪙 A Couple Million Lines of Haskell: Production Engineering at Mercury
🧠 Agentic Coding is a Trap
⚡ How To Make a Fast Dynamic Language Interpreter
📉 Reduced Median Memory Estimation Error by 99%, With the Help of AI
🧾 When upserts don't update but still write: Debugging Postgres performance at scale

Ship fewer traps, ship more truth.

Have a great week!
FAUN.dev() Team

Join IaCConf 2026 to learn how top IaC leaders from companies like Google, Sanofi, and AHEAD are solving the hardest problems in cloud provisioning, state management, and platform engineering.

May 14. Free to attend. Register now.

Learn how to build intelligent, scalable platforms with AI — from self-service developer portals to AI-driven observability. This hands-on cohort equips engineers, SREs, and tech leaders with real-world frameworks to design smarter platforms and boost developer productivity

Register here.

Ubuntu 26.04 LTS - Resolute Raccoon ships GNOME 50. GNOME-on-X11 is gone; the session runs only on Wayland. Mutter patches shave NVIDIA blocked frame time down to microseconds.

The release requires systemd 259 and cgroup v2. It swaps initramfs to Dracut. Desktop minimums rise. Binaries default to memory-safe Rust.

Ubuntu adds native CUDA and ROCm to main repos. Livepatch lands on Arm64. The installer gains TPM-backed full-disk encryption. App Center becomes the central app manager.

Are Your APIs Ready for AI Agents? A Hands-on Workshop on May 23rd

AI agents are beginning to autonomously call APIs, chain services, and create integrations that most platforms were never designed to handle. This hands-on masterclass on Designing AI-ready APIs helps architects and developers build governed, predictable API ecosystems using OpenAPI, Overlay, and Arazzo.

Learn how to add guardrails, improve discoverability, and safely evolve existing APIs for automated consumption.

FAUN.dev readers get an exclusive 40% discount using code FAUN40.

Cloud Native CI/CD with GitLab: From Commit to Production Ready is a complete, hands-on path to becoming the person on your team who actually understands GitLab CI/CD, not just the YAML, but the architecture underneath.

You'll start with the fundamentals: jobs, stages, the container registry, your first working pipeline; and build up to the parts most engineers learn the hard way in production: reusable definitions with extends and includes, DAGs for non-sequential execution, artifact strategies, conditional logic and workflow rules, parallelism and matrix builds, runner and executor internals, and cloud-native runners on Kubernetes - with caching, autoscaling, and observability wired in. The final chapters walk through multi-stage continuous deployment with HelmandKubernetes end to end.

23 chapters. Hands-on throughout. Designed so beginners can follow it linearly and experienced engineers can jump straight to the chapter they need.

Written by Aymen El Amri - founder of FAUN.dev(), author of multiple cloud-native engineering books, and trainer to thousands of DevOps and platform engineers worldwide.

→ Explore the course

AI-driven coding agents are the hot new trend, but beware of the trade-offs: increased complexity, skills atrophy, vendor lock-in, and fluctuating costs. Only skilled developers can spot issues in the vast lines of generated code, but paradoxically, AI tools are impacting critical thinking skills negatively. To avoid this, demote AI's role, leverage it for planning, and stay actively engaged in implementation to maintain coding skills and mitigate comprehension debt.

The Datadog team introduced a new upsert query to track inactive hosts, but it unexpectedly increased disk writes and WAL syncs due to row locking. By digging into Postgres's Write-Ahead Logging (WAL) and rewriting the query using a Common Table Expression (CTE), they avoided unnecessary overhead and improved performance. The new query eliminated the need for locking and minimized WAL activity and allowed for more efficient host tracking without sacrificing correctness.

The compaction pipeline at Mixpanel ran into memory estimation issues causing OOMKills. By implementing AI-assisted analysis, they were able to reduce median estimation error by 99%, leading to a significant improvement in memory estimation accuracy. Through thorough analysis and exploration of alternative approaches, they were able to find a simple solution that outperformed more complex alternatives.

Zef's AST-walking interpreter posts a 16.6× speed-up. The gains come from surgical changes: 64-bit tagged values, AST node & RMW specialization, symbol hash-consing, inline caches, and a shaped object model. Developers built it on Fil-C++ and later ported it to Yolo-C++. The Yolo build adds ~4x speed, at the cost of unsound GC behavior. Zef now challenges CPython, Lua, and QuickJS.

Mercury runs ~2M lines of Haskell in production.
They chose Temporal to retire cron and DB-backed state machines. Durable workflows replace brittle coordination.

They open-sourced a Haskell SDK for Temporal, wired in OpenTelemetry hooks, and pushed records-of-functions plus domain-error types.

232+ Claude Code skills & agent plugins for Claude Code, Codex, Gemini CLI, Cursor, and 8 more coding agents — engineering, marketing, product, compliance, C-level advisory.

AI gateway written in Go. Lightweight unified OpenAI-compatible API for OpenAI, Anthropic, Gemini, Groq, xAI & Ollama. LiteLLM alternative with observability, guardrails, streaming, costs and usage tracking.

Broccoli turns Linear tickets into shipped PRs — powered by Claude and Codex, running on your own Google Cloud or Blaxel containers.

Harnessing Go's perfomance for blazingly fast web scraping, inspired by Python's Scrapy framework.

Hear your agent suffer through your code

Did you know that Redis keeps the recent replication stream in an in-memory circular buffer called the replication backlog, so when a replica reconnects after a brief disconnect it can issue PSYNC with its last offset and replay only the missed bytes instead of pulling the full dataset? If the offset has already fallen out of the backlog, the master is forced into a full resync, which means forking, producing an RDB snapshot, and streaming it over the wire, all of which spike CPU and bandwidth on the primary. That is why sizing repl-backlog-size to comfortably cover your worst expected disconnect window is one of the highest leverage replication knobs in Redis, well ahead of toggling repl-diskless-sync.

"Dev tools promise fewer mistakes; the good ones just make your mistakes reproducible. The rest of engineering is deciding which mistakes deserve automation."
— SenseiOne

"Design for horizontal scale from day one" is sound advice until you're three sprints in, sharding a service that peaks at 40 RPS, and your stateless everything rule means a request that used to be one in-process call is now four network hops through a queue you don't have capacity to operate. Premature distribution doesn't just cost latency, it costs the debugging budget you'll need when real load shows up and you can't tell whether the bug is in your code or in the seams you added to prepare for it.