The New Stack Makers

The etcd project — a distributed key-value store older than Kubernetes — recently faced significant challenges due to maintainer turnover and the resulting loss of unwritten institutional knowledge. Lead maintainer Marek Siarkowicz explained that as longtime contributors left, crucial expertise about testing procedures and correctness guarantees disappeared. This gap led to a problematic release that introduced critical reliability issues, including potential data inconsistencies after crashes.To rebuild confidence in etcd’s correctness, the new maintainer team introduced “robustness testing,” creating a framework inspired by Jepsen to validate both basic and distributed-system behavior. Their goal was to ensure linearizability, the “Holy Grail” of distributed systems, which required developing custom failure-injection tools and teaching the community how to debug complex scenarios.The team later partnered with Antithesis to apply deterministic simulation testing, enabling fully reproducible execution paths a

Helm — originally a hackathon project called Kate’s Place — turned 10 in 2025, marking the milestone with the release of Helm 4, its first major update in six years. Created by Matt Butcher and colleagues as a playful take on “K8s,” the early project won a small prize but quickly grew into a serious effort when Deus leadership recognized the need for a Kubernetes package manager. Renamed Helm, it rapidly expanded with community contributors and became one of the first CNCF graduating projects.Helm 4 reflects years of accumulated design debt and evolving use cases. After the rapid iterations of Helm 1, 2, and 3, the latest version modernizes logging, improves dependency management, and introduces WebAssembly-based plugins for cross-platform portability—addressing the growing diversity of operating systems and architectures. Beyond headline features, maintainers emphasize that mature projects increasingly deliver “boring” but essential improvements, such as better logging, which simplify workflows and integrate

Kubernetes has relied on role-based access control (RBAC) since 2017, but its simplicity limits what developers can express, said Micah Hausler, principal engineer at AWS, on The New Stack Makers. RBAC only allows actions; it can’t enforce conditions, denials, or attribute-based rules. Seeking a more expressive authorization model for Kubernetes, Hausler explored Cedar, an authorization engine and policy language created at AWS in 2022 and later open-sourced. Although not designed specifically for Kubernetes, Cedar proved capable of modeling its authorization needs in a concise, readable way. Hausler highlighted Cedar’s clarity—nontechnical users can often understand policies at a glance—as well as its schema validation, autocomplete support, and formal verification, which ensures policies are correct and produce only allow or deny outcomes.Now onboarding to the CNCF sandbox, Cedar is used by companies like Cloudflare and MongoDB and offers language-agnostic tooling, including a Go implementation donated by S

JupyterLite, a fully browser-based distribution of JupyterLab, is enabling new levels of global scalability in technical education. Developed by Sylvain Corlay’s QuantStack team, it allows math and programming lessons to run entirely in students’ browsers — kernel included — without relying on Docker or cloud-scale infrastructure. Its most prominent success is Capytale, a French national deployment that supports half a million high school students and over 200,000 weekly sessions from essentially a single server, which hosts only teaching content while computation happens locally in each browser.QuantStack, founded in 2016 as what Corlay calls an “accidental startup,” has since grown into a 30-person team contributing across Jupyter, Conda-Forge, and Apache Arrow. But JupyterLite embodies its most ambitious goal: making programming education accessible to countries with rapidly growing youth populations, such as Nigeria, where traditional cloud-hosted notebooks are impractical. Achieving a billion-user future

AWS’s approach to Elastic Kubernetes Service has evolved significantly since its 2018 launch. According to Mike Stefanik, Senior Manager of Product Management for EKS and ECR, today’s users increasingly represent the late majority—teams that want Kubernetes without managing every component themselves. In a conversation onThe New Stack Makers, Stefanik described how AI workloads are reshaping Kubernetes operations and why AWS open-sourced an MCP server for EKS. Early feedback showed that meaningful, task-oriented tool names—not simple API mirrors—made MCP servers more effective for LLMs, prompting AWS to design tools focused on troubleshooting, runbooks, and full application workflows. AWS also introduced a hosted knowledge base built from years of support cases to power more capable agents.While “agentic AI” gets plenty of buzz, most customers still rely on human-in-the-loop workflows. Stefanik expects that to shift, predicting 2026 as the year agentic workloads move into production. For experimentation, he r

At KubeCon + CloudNativeCon 2025 in Atlanta, the panel of experts - Kate Goldenring of Fermyon Technologies, Idit Levine of Solo.io, Shaun O'Meara of Mirantis, Sean O'Dell of Dynatrace and James Harmison of Red Hat - explored whether the cloud native era has evolved into an AI native era — and what that shift means for infrastructure, security and development practices. Jonathan Bryce of the CNCF argued that true AI-native systems depend on robust inference layers, which have been overshadowed by the hype around chatbots and agents. As organizations push AI to the edge and demand faster, more personalized experiences, Fermyon’s Kate Goldenring highlighted WebAssembly as a way to bundle and securely deploy models directly to GPU-equipped hardware, reducing latency while adding sandboxed security.Dynatrace’s Sean O’Dell noted that AI dramatically increases observability needs: integrating LLM-based intelligence adds value but also expands the challenge of filtering massive data streams to understand user behavi

AWS re:Invent has long featured CTO Werner Vogels’ closing keynote, but this year he signaled it may be his last, emphasizing it’s time for “younger voices” at Amazon. After 21 years with the company, Vogels reflected on arriving as an academic and being stunned by Amazon’s technical scale—an energy that still drives him today. He released his annual predictions ahead of re:Invent, with this year’s five themes focused heavily on AI and broader societal impacts.Vogels highlights technology’s growing role in addressing loneliness, noting how devices like Alexa can offer comfort to those who feel isolated. He foresees a “Renaissance developer,” where engineers must pair deep expertise with broad business and creative awareness. He warns quantum-safe encryption is becoming urgent as data harvested today may be decrypted within five years. Military innovations, he notes, continue to influence civilian tech, for better and worse. Finally, he argues personalized learning can preserve children’s curiosity and better

DevOps practitioners — whether developers, operators, SREs or business stakeholders — increasingly rely on telemetry to guide decisions, yet face growing complexity, siloed teams and rising observability costs. In a conversation at KubeCon + CloudNativeCon North America, IBM’s Jacob Yackenovich emphasized the importance of collecting high-granularity, full-capture data to avoid missing critical performance signals across hybrid application stacks that blend legacy and cloud-native components. He argued that observability must evolve to serve both technical and nontechnical users, enabling teams to focus on issues based on real business impact rather than subjective judgment.AI’s rapid integration into applications introduces new observability challenges. Yackenovich described two patterns: add-on AI services, such as chatbots, whose failures don’t disrupt core workflows, and blocking-style AI components embedded in essential processes like fraud detection, where errors directly affect application function.Ris

Major banks once built their own Linux kernels because no distributions existed, but today commercial distros — and Kubernetes — are universal. At KubeCon + CloudNativeCon North America, AWS’s Jesse Butler noted that Kubernetes has reached the same maturity Linux once did: organizations no longer build bespoke control planes but rely on shared standards. That shift influences how AWS contributes to open source, emphasizing community-wide solutions rather than AWS-specific products.Butler highlighted two AWS EKS projects donated to Kubernetes SIGs: KRO and Karpenter. KRO addresses the proliferation of custom controllers that emerged once CRDs made everything representable as Kubernetes resources. By generating CRDs and microcontrollers from simple YAML schemas, KRO transforms “glue code” into an automated service within Kubernetes itself. Karpenter tackles the limits of traditional autoscaling by delivering just-in-time, cost-optimized node provisioning with a flexible, intuitive API. Both projects embody AWS’

Clockwork began with a narrow goal—keeping clocks synchronized across servers—but soon realized that its precise latency measurements could reveal deeper data center networking issues. This insight led the company to build a hardware-agnostic monitoring and remediation platform capable of automatically routing around faults. Today, Clockwork’s technology is especially valuable for large GPU clusters used in training LLMs, where communication efficiency and reliability are critical. CEO Suresh Vasudevan explains that AI workloads are among the most demanding distributed applications ever, and Clockwork provides building blocks that improve visibility, performance and fault tolerance. Its flagship feature, FleetIQ, can reroute traffic around failing switches, preventing costly interruptions that might otherwise force teams to restart training from hours-old checkpoints. Although the company originated from Stanford research focused on clock synchronization for financial institutions, the team eventually recogni

At JupyterCon 2025, Jupyter Deploy was introduced as an open source command-line tool designed to make cloud-based Jupyter deployments quick and accessible for small teams, educators, and researchers who lack cloud engineering expertise. As described by AWS engineer Jonathan Guinegagne, these users often struggle in an “in-between” space—needing more computing power and collaboration features than a laptop offers, but without the resources for complex cloud setups. Jupyter Deploy simplifies this by orchestrating an entire encrypted stack—using Docker, Terraform, OAuth2, and Let’s Encrypt—with minimal setup, removing the need to manually manage 15–20 cloud components. While it offers an easy on-ramp, Guinegagne notes that long-term use still requires some cloud understanding. Built by AWS’s AI Open Source team but deliberately vendor-neutral, it uses a template-based approach, enabling community-contributed deployment recipes for any cloud. Led by Brian Granger, the project aims to join the official Jupyter ec

In an interview at JupyterCon, Brian Granger — co-creator of Project Jupyter and senior principal technologist at AWS — reflected on Jupyter’s evolution and how AI is redefining open source sustainability. Originally inspired by physics’ modular principles, Granger and co-founder Fernando Pérez designed Jupyter with flexible, extensible components like the notebook format and kernel message protocol. This architecture has endured as the ecosystem expanded from data science into AI and machine learning. Now, AI is accelerating development itself: Granger described rewriting Jupyter Server in Go, complete with tests, in just 30 minutes using an AI coding agent — a task once considered impossible. This shift challenges traditional notions of technical debt and could reshape how large open source projects evolve. Jupyter’s 2017 ACM Software System Award placed it among computing’s greats, but also underscored its global responsibility. Granger emphasized that sustaining Jupyter’s mission — empowering human reason

Jupyter AI v3 marks a major step forward in integrating intelligent coding assistance directly into JupyterLab. Discussed by AWS engineers David Qiu and Piyush Jain at JupyterCon, the new release introduces AI personas— customizable, specialized assistants that users can configure to perform tasks such as coding help, debugging, or analysis. Unlike other AI tools, Jupyter AI allows multiple named agents, such as “Claude Code” or “OpenAI Codex,” to coexist in one chat. Developers can even build and share their own personas as local or pip-installable packages. This flexibility was enabled by splitting Jupyter AI’s previously large, complex codebase into smaller, modular packages, allowing users to install or replace components as needed. Looking ahead, Qiu envisions Jupyter AI as an “ecosystem of AI personas,” enabling multi-agent collaboration where different personas handle roles like data science, engineering, and testing. With contributors from AWS, Apple, Quansight, and others, the project is poised to ex

In this episode of The New Stack Podcast, hosts Alex Williams and Frederic Lardinois spoke with Keith Ballinger, Vice President and General Manager of Google Cloud Platform Developer Experience (GPC), about the evolution of agentic coding tools and the future of programming. Ballinger, a hands-on executive who still codes, discussed Gemini CLI, Google’s response to tools like Claude Code, and his broader philosophy on how developers should work with AI. He emphasized that these tools are in their “first inning” and that developers must “slow down to speed up” by writing clear guides, focusing on architecture, and documenting intent—treating AI as a collaborative coworker rather than a one-shot solution. Ballinger reflected on his early AI experiences, from Copilot at GitHub to modern agentic systems that automate tool use. He also explored the resurgence of the command line as an AI interface and predicted that programming will increasingly shift from writing code to expressing intent. Ultimately, he envision

At the PyTorch Conference 2025 in San Francisco, Luca Antiga — CTO of Lightning AI and head of the PyTorch Foundation’s Technical Advisory Council — discussed the evolution and influence of PyTorch. Originally designed to be “Pythonic” and researcher-friendlyAntiga emphasized that PyTorch has remained central across major AI shifts — from early neural networks to today’s generative AI boom — powering not just model training but also inference systems such as vLLM and SGLang used in production chatbots. Its flexibility also makes it ideal for reinforcement learning, now commonly used to fine-tune large language models (LLMs).On the PyTorch Foundation, Antiga noted that while it recently expanded to include projects likev LLM ,DeepSpeed, and Ray, the goal isn’t to become a vast umbrella organization. Instead, the focus is on user experience and success within the PyTorch ecosystem.Learn more from The New Stack about the latest in PyTorch:Why PyTorch Gets All the LoveLightning AI Brings a PyTorch Copilot to Its

Harness co-founder Jyoti Bansal highlights a growing issue in software development: while AI tools help generate more code, they often create bottlenecks further along the pipeline, especially in testing, deployment, and compliance. Since its 2017 launch, Harness has aimed to streamline these stages using AI and machine learning. With the rise of large language models (LLMs), the company shifted toward agentic AI, introducing a library of specialized agents—like DevOps, SRE, AppSec, and FinOps agents—that operate behind a unified interface called Harness AI. These agents assist in building production pipelines, not deploying code directly, ensuring human oversight remains critical for compliance and security.Bansal emphasizes that AI in development isn't replacing people but accelerating workflows to meet tighter timelines. He also notes strong enterprise adoption, with even large, traditionally slower-moving organizations embracing AI integration. On the topic of an AI bubble, Bansal sees it as a natural par

The agentic AI space faces challenges around secure, governed connectivity between agents, tools, large language models, and microservices. To address this, Solo.io developed two open-source projects: Kagent and Agentgateway. While Kagent, donated to the Cloud Native Computing Foundation, helps scale AI agents, it lacks a secure way to mediate communication between agents and tools. Enter Agentgateway, donated to the Linux Foundation, which provides governance, observability, and security for agent-to-agent and agent-to-tool traffic. Written in Rust, it supports protocols like MCP and A2A and integrates with Kubernetes Gateway API and inference gateways.Lin Sun, Solo.io’s head of open source, explained that Agentgateway allows developers to control which tools agents can access—offering flexibility to expose only tested or approved tools. This enables fine-grained policy enforcement and resilience in agent communication, similar to how service meshes manage microservice traffic. Agentgateway ensures secure an

David Cramer, founder and chief product officer of Sentry, remains skeptical about generative AI's current ability to replace human engineers, particularly in software production. While he acknowledges AI tools aren't yet reliable enough for full autonomy—especially in tasks like patch generation—he sees value in using large language models (LLMs) to enhance productivity. Sentry's AI-powered tool, Seer, uses GenAI to help developers debug more efficiently by identifying root causes and summarizing complex system data, mimicking some functions of senior engineers. However, Cramer emphasizes that human oversight remains essential, describing the current stage as "human in the loop" AI, useful for speeding up code reviews and identifying overlooked bugs.Cramer also addressed Sentry's shift from open source to fair source licensing due to frustration over third parties commercializing their software without contributing back. Sentry now uses Functional Source Licensing, which becomes Apache 2.0 after two years. T

Cursor, the AI code editor, recently integrated with Linear, a project management tool, enabling developers to assign tasks directly to Cursor's background coding agent within Linear. The collaboration felt natural, as Cursor already used Linear internally. Linear's new agent-specific API played a key role in enabling this integration, providing agents like Cursor with context-aware sessions to interact efficiently with the platform.Developers can now offload tasks such as fixing issues, updating documentation, or managing dependencies to the Cursor agent. However, both Linear’s Tom Moor and Cursor’s Andrew Milich emphasized the importance of giving agents clear, thoughtful input. Simply assigning vague tasks like “@cursor, fix this” isn’t effective—developers still need to guide the agent with relevant context, such as links to similar pull requests.Milich and Moor also discussed the growing value and adoption of autonomous agents, and hinted at a future where more companies build agent-specific APIs to supp

In this episode of The New Stack Agents, ServiceNow CTO and co-founder Pat Casey discusses why the company runs 90% of its workloads—including AI infrastructure—on its own physical servers rather than the public cloud. ServiceNow maintains GPU hubs across global data centers, enabling efficient, low-latency AI operations. Casey downplays the complexity of running AI models on-prem, noting their team’s strong Kubernetes and Triton expertise. The company recently switched from GitHub Copilot to its own AI coding assistant, Windsurf, yielding a 10% productivity boost among 7,000 engineers. However, use of such tools isn’t mandatory—performance remains the main metric. Casey also addresses the impact of AI on junior developers, acknowledging that AI tools often handle tasks traditionally assigned to them. While ServiceNow still hires many interns, he sees the entry-level tech job market as increasingly vulnerable. Despite these concerns, Casey remains optimistic, viewing the AI revolution as transformative and ul