Kubernetes kills IPVS, Cloudflare ships a browser without a browser, and Canonical lets AI write your patches
8 min read · 12 sources
- Kubernetes deprecated the kube-proxy IPVS backend in 1.35 and will remove it in 1.40, forcing migration to nftables or iptables.
- Cloudflare open-sourced Kitesurf, a stateless browser for AI agents that runs on Workers without a browser container.
- Salesforce saved more than $5 million by combining VPA and HPA across 1,400 Amazon EKS clusters.
- Canonical's new Edge archive gives Ubuntu Pro users opt-in access to AI-generated security patches via the Surge multi-agent system.
- GitHub's October 2026 report logged eight incidents, including a memory exhaustion failure in its runner-assignment service.
Kubernetes is finally pulling the plug on the kube-proxy IPVS backend, and the migration window is tight. If you’re still running IPVS in production, you have until Kubernetes 1.40 to move, and the nftables backend is where you’re headed.
Meanwhile, Cloudflare just dropped a browser engine that isn’t a browser at all. Kitesurf runs on Workers, parses HTML, executes JavaScript, and renders screenshots without a single browser container. It’s Apache 2.0, it speaks Chrome DevTools Protocol, and it might change how you think about AI agents that need to “see” the web.
Salesforce drove more than $5 million in cumulative infrastructure savings by pairing VPA with HPA across 1,400 EKS clusters.
kube-proxy IPVS is dead. Plan your migration now.
Kubernetes SIG Network has officially deprecated the kube-proxy IPVS backend as of Kubernetes 1.35, with removal scheduled for 1.40. That gives you roughly four minor releases to move off it.
IPVS was introduced to fix the latency problems in the iptables backend, both in the data plane and the control plane. The problem: the kernel IPVS API never fully mapped onto the Kubernetes Services API. The mismatch created edge cases that SIG Network had to carry forever, and the maintenance cost stopped being worth it.
Your options are the nftables backend, which has been GA since 1.33, or the iptables backend. If you’ve been avoiding nftables because it was new, that excuse expired a few releases ago. The deprecation notice is the push you need to test your Services against nftables now, not when 1.40 lands and your cluster refuses to start.
Cloudflare's Kitesurf: a browser engine with no browser
Cloudflare has open-sourced Kitesurf under Apache 2.0, and it’s a genuinely different take on headless browsing. It parses HTML, applies CSS, executes JavaScript, and renders screenshots entirely on Cloudflare Workers, with no browser container anywhere in the path.
The trick is compiling Rust to WebAssembly. Kitesurf leans on open-source Rust crates like Blitz, Stylo, Boa, and Linebender, all compiled to WASM and running on Workers’ compute. It supports the Chrome DevTools Protocol, so existing Puppeteer clients can talk to it without rewriting their tooling.
For anyone running AI agents that need to interact with web pages, this removes a whole class of infrastructure. No more browser containers to patch, scale, or watch for memory leaks. Deploy it to Workers or run it locally from the GitHub repo. The trade-off is performance and full browser fidelity, but for agent workloads that just need to load a page, click a button, and take a screenshot, this could be enough.
Technical debt is the price of learning, not a sin
Xin Yao’s talk at GOTO Copenhagen reframes technical debt as something other than a failure. His argument: debt is what you accumulate when you design software with incomplete knowledge. It’s the price of learning by doing, not a mark of poor engineering.
The problem starts when software doesn’t catch up with what you’ve learned. Debt becomes pathological when it’s permanent, when the codebase never absorbs the lessons that created the debt in the first place. Yao also notes debt can be a deliberate business strategy, a way to communicate with management about what’s being deferred and why.
The real tension is optionality. Every feature you ship removes some future options, and the challenge is balancing feature delivery against preserving room to maneuver. For engineers who’ve been told debt is always bad, this is a useful counterweight. For managers, it’s a reminder that the fix isn’t just more refactoring, it’s building shared understanding faster.
Salesforce's $5 million autoscaling playbook
Salesforce published a detailed breakdown of how it right-sizes Kubernetes workloads at scale on EKS, and the numbers are worth your time. The fleet: over 1,400 EKS clusters, individual clusters up to 2,500 nodes, roughly 1,000 workloads. The result: more than $5 million in cumulative infrastructure savings.
The approach is a multi-dimensional scaling strategy that pairs Vertical Pod Autoscaler with Horizontal Pod Autoscaler. The key insight is keeping the two controllers from reacting to the same signal. For Istio workloads, VPA adjusts CPU and memory while HPA scales replicas based on proxy connection count. Different signals, complementary actions.
The backstory matters. Salesforce’s static resource management caused production incidents, high operational toil, and scalability bottlenecks. The fix wasn’t a single tool, it was a coordinated strategy across the whole fleet. If you’re running EKS at any scale, this is a concrete reference architecture for how to make VPA and HPA work together instead of fighting each other.
Your OpenTelemetry pipeline is eating your CPU
A presentation at Cloud Native Days Austria 2026 makes the case that observability has a cost, and it’s higher than you think. OpenTelemetry pipelines consume significant CPU, memory, network, and storage as applications scale, and the overhead comes from the entire pipeline: span generation, collection, processing, exporting, and storage.
The uncomfortable finding is that a single overhead percentage is misleading. The cost is workload-dependent, so your neighbor’s 5% overhead tells you nothing about your own. The investigation used Telemetrygen to identify five common Collector configuration problems, including excessive attributes, unbounded queues, and expensive processing.
The practical takeaways: memory-limiter placement matters, sampling is a trade-off you have to make consciously, and tail sampling requires trace-aware routing when Collectors scale horizontally. If your observability stack is slowing down your applications, the fix is probably in your Collector configuration, not in buying more nodes.
Claude-BugHunter: 83 skills for red-team work
Claude-BugHunter is a self-contained Claude Code skill bundle for bug hunting and external red-team work. It ships with 83 skills, 15 slash commands, and 681 disclosed-report patterns across 24 vulnerability classes.
The bundle includes enterprise identity and infrastructure attack matrices, engagement-folder scaffolding, and Burp MCP integration. It installs as a plugin or via a copy install script, and it automatically loads relevant skills based on the topic you describe in plain English. For security teams that already use Claude Code, this is a way to standardize methodology across engagements without reinventing prompts every time.
PaperOrchestra: Google Research's automated paper writer
Google Research released PaperOrchestra, a multi-agent framework that turns raw research materials into submission-ready LaTeX manuscripts. It orchestrates specialized agents for outline generation, literature review, section writing, content refinement, and plotting.
Setup requires a Conda environment with Python 3.11 and API keys for models like Gemini or OpenAI. The dataset will be released separately later. If you’re evaluating whether multi-agent frameworks can produce usable research output, this is a concrete reference implementation from a team that publishes papers for a living.
Platform teams: obstacles vs. challenges
Octopus Deploy’s Steve Fenton uses Future of Platform Engineering survey data to draw a useful distinction. Obstacles are things that prevent platform success: lack of clear strategy, budget, and skills. Challenges are things like tooling sprawl, competing priorities, resistance to change, and technical complexity.
The finding that matters: high-performing platforms are more likely to face challenges than obstacles. Tooling sprawl is the top challenge for successful teams, not a sign of failure. The top obstacle is lack of clear strategy, which leads to frustration and teams building from scratch instead of using best-in-class tools. If your platform team is struggling, first ask whether you have a strategy problem or a challenge problem. The fixes are completely different.
Uber's Secure AI Framework for autonomous agents
Uber developed a Secure AI Framework to govern autonomous AI agents across their lifecycle, addressing prompt injection, tool poisoning, goal hijacking, and hallucination-driven actions. The framework has three layers: provisioning-time registration in the Uber AI Registry with automated analysis via Argus, runtime behavioral authorization through AI Registry, AI Guards, and Agent Guard runtime hooks, and post-deployment Continuous Assurance with unified execution traces and replay.
The notable design choice: per-agent intents worked better than generic thresholds. The framework extends traditional identity and API controls to evaluate whether an agent’s action fits the user’s request, purpose, permissions, and context. For anyone deploying autonomous agents, this is a reference architecture for what governance looks like when agents act on their own.
Canonical's Edge archive: AI-written security patches
Canonical is offering Ubuntu Pro users early access to AI-generated security patches through a new opt-in archive called Edge. The patches are built by a multi-agent system called Surge for backporting userspace fixes.
The Edge archive is not enabled by default. You have to manually enable it with sudo apt install pro-edge && sudo pro-enable-edge, with optional automatic updates. Surge is not used for kernel vulnerabilities, and stable releases remain strictly human-validated. If you’re comfortable running AI-generated patches on your systems, this is a way to get fixes earlier. If you’re not, the default is still human-reviewed patches.
GitHub's October availability report
GitHub’s October 2026 availability report details eight incidents causing degraded performance. The highlights: write outages, monitoring gaps during data migrations, and a regional network disruption that caused the runner-assignment service to wait on database calls until it exhausted memory and restarted.
As of November 9, 140 of 151 production services and 32 of 50 MySQL primaries run in Azure Central US, with compute capacity more than doubled across nine Kubernetes clusters. Targeted improvements include an 85% reduction in NAT-gateway usage and an 85% reduction in peak database reads for a high-traffic Actions endpoint. The report is a reminder that even GitHub has bad days, and the postmortem details are worth reading for the failure modes alone.
PlanetScale: IOPS control for Postgres restores
PlanetScale now lets users set storage IOPS and throughput when restoring Postgres backups. Large restores are often limited by disk throughput, and this gives you control over that bottleneck.
The settings are available in the Restore to new branch dialog for branches on AWS network-attached storage. CLI users get --storage-iops and --storage-throughput; API users get storage_iops and storage_throughput_mibs. The settings persist on the restored branch and can be changed later. If you’ve ever waited hours for a big Postgres restore, this is a quality-of-life improvement worth knowing about.
You May Also Like
Salesforce stumbles, Kubernetes gets pod-level resource control, and the cost of your coding agent's harness
Salesforce spent hours stumbling through a global outage that stalled requests on an internal login service and exhausted server resources across hundreds of …
Kagent 1.0 Is GA: Kubernetes Gets a Native Way to Run AI Agents
Solo.io shipped kagent 1.0, an enterprise distribution of the CNCF-hosted Agent Substrate that runs AI agents as schedulable actors on warm Kubernetes worker …
Gheist’s GitHub token heist stalls, Cloudflare maps Q3 outages, and autoscalers finally clash
Wiz’s research exposed “Gheist,” an automated campaign that slipped secret‑exfiltrating GitHub Actions into 850+ accounts before GitHub’s new token controls …




