Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
🔁 Airbnb replays real workloads to test changes
🐱 Mistral previews ML4 ahead of open weights
⚡ Next.js 16.4 adds lazy compilation
🕰️ Spanner Omni drops Google's atomic clocks
☸️ Kubernetes v1.35 fails nodes on cgroup v1
Plus: 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.
🔁 Airbnb replays real workloads to test changes LINK
- Airbnb built a system that captures real production MySQL traffic through ProxySQL and replays it offline, letting its database team load-test, plan capacity, and de-risk version upgrades across hundreds of clusters running millions of queries per second.
- Because logs are captured at the proxy with no application changes, the pipeline partitions queries per cluster into five-minute buckets, reassembles transactions in order, and rewrites each INSERT to pin last_insert_id so dependent reads behave consistently across engines.
- During the MySQL 5.7-to-8.0 upgrade, replay caught a query pattern whose latency jumped from 0.03 to 2.6 seconds reading 273 MB per join, plus unordered LIMIT queries returning different rows, letting teams fix issues before they hit production.
🐱 Mistral previews ML4 ahead of open weights LINK
- Mistral has launched a public preview of Mistral Large 4, a one-trillion-parameter multimodal model code-named "Le Chonk," pitched as the strongest open-weight model developed outside China with a focus on coding, cybersecurity, finance and visual grounding.
- The sparse model activates only 49 billion parameters during inference, accepts multimodal input but outputs text, and was trained from scratch over about two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.
- Mistral will publish the weights under a custom license on Oct. 27 after a three-week testing period; preview benchmarks look competitive-62% on DeepSWE coding, but ML4 doesn't lead every configuration, and the results stay provisional until outsiders can test the final checkpoint.
⚡ Next.js 16.4 adds lazy compilation LINK
- Next.js 16.4 makes Cache Components the recommended model for every app and the default for new projects, letting developers annotate parts of a component tree for client-side, server-side, or build-time caching, replacing App Router's implicit behaviors.
- A new route config can force a shell, prefetch, or full navigation to stay static, failing the build if dynamic content sneaks in, while new APIs defer loading cached content until an actual navigation instead of eager prefetching.
- All apps also gain Turbopack's disk cache using 20-25% less space via Zstandard compression, lazy server compilation that only rebuilds requested routes, smaller production bundles, and React 19.3 with stable View Transitions.
🕰️ Spanner Omni drops Google's atomic clocks LINK
- Google has made Spanner Omni generally available, letting its distributed SQL database run in your own data center, on rival clouds, or on a laptop by swapping out the Colossus file system and the atomic-clock-backed TrueTime service it depended on.
- In place of TrueTime, a software layer provides error-bounded time sync across heterogeneous hardware, working because Spanner already overlaps its uncertainty waits with other work, while a Colossus-like layer writes to local disks with automatic shard splitting and rebalancing.
- The catch is no availability SLA: patching, upgrades, backups and monitoring move in-house through Prometheus and Grafana, Google Cloud integrations like BigQuery and Gemini Enterprise are excluded, and the real tradeoff is managed Spanner versus tail latency and operational toil.
☸️ Kubernetes v1.35 fails nodes on cgroup v1 LINK
- Kubernetes has deprecated cgroup v1, and starting with v1.35 the kubelet's `failCgroupV1` setting defaults to true, so a node still running cgroup v1 will fail kubelet startup unless administrators apply a temporary `failCgroupV1: false` override.
- cgroup v2 has been stable since Kubernetes v1.25 and brings a unified hierarchy plus features v1 can't support, including container-aware OOM handling via `memory.oom.group`, PSI contention metrics, and tiered Memory QoS using `memory.high`, `memory.min`, and `memory.low`.
- Requirements include Linux kernel 5.8 or later and a runtime like containerd v1.4+ or CRI-O v1.20+, and since kubeadm's SystemVerification preflight now errors on cgroup v1 with v1.35 kubelets, migrate every node before the v1 fallback is removed in v1.38.
Other news you might like
- AI Changed How Spotify Builds. What We Learned (and Fixed) About Quality at Higher Velocity LINK
- How we made registry metadata 70% smaller LINK
- Shipping JPEG XL in Chrome LINK
- COSMIC shuts the door on AI code as GNOME debates letting bug reports in LINK
- Building Git infrastructure for agent-scale development LINK
- EmbeddingGemma 2: an open, lightweight multimodal embedding model LINK
- LibreOffice says ‘no AI’ is now a software feature LINK
- How fast is Python 3.15? LINK
🧰 Trending tools
Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a chatroom when your allowance runs out. LINK
Kilo Code for JetBrains: an open-source AI coding agent for JetBrains IDEs that lets you build and iterate using flexible model choices across projects. LINK
Mastra Factory: a TypeScript framework for building AI agents with staged workflows, memory, streaming, evals, tracing, and a Studio UI for testing. LINK
iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate and fix them. LINK
Tines 3B: builds and runs team automations, apps, and AI agents on one secure platform, keeping workflows unified and easy to manage LINK
Enter Pro: an AI-native platform for building and scaling applications, handling development and growth without stitching together separate tools and services. LINK
📚 Trending papers & reports
AI memory systems can wrongly resurrect outdated facts just because an old statement reappears, crashing accuracy from ~99% to ~11%, though a simple guard that refuses to reactivate retired values restores it to ~98%. LINK
Brain-guided code tuning uses signals from how humans read code to decide which small parts of a model to update, cutting the cost of fine-tuning coding assistants without losing accuracy. LINK
Spec-driven coding gives teams a 40-point checklist for writing clearer instructions to coding agents, and projects with the best specs spent half as many commits fixing bugs, ~12% versus ~20%. LINK
Coding agent rulebooks automatically extract each software project's unwritten contribution rules so AI coding assistants follow house style, boosting compliance by up to ~45% and matching the right norms with 86% precision. LINK
Automated bug-fixing scorecards show that judging AI code patches only by whether tests pass hides key trade-offs in security, maintainability, and cost, and that a sparse model matched larger ones using 3-6x fewer active parameters. LINK
See you tomorrow for a new dose of ☕️ Devshot!