Netflix swaps homegrown batch scheduler for Kueue, Kubernetes gets native GPU scheduling

Codeplain’s open‑source Plain language treats specs as the single source of truth, letting AI regenerate implementation whenever a change is needed. By shifting review from code to specs, developers reduce cognitive load and cut maintenance overhead, a model gaining traction alongside tools like Amazon’s Kiro and GitHub’s Spec Kit.

Netflix migrated millions of batch jobs from its proprietary Compute Managed Batch system to Kueue, a Kubernetes‑native job queue. The switch cut queuing complexity, raised resource utilization and kept the change invisible to end‑users. Kueue now handles preemption, topology‑aware scheduling and fair‑share across tenants on the Titus platform.

NASA is trialing a medical AI assistant for deep‑space crews that runs entirely offline, using Red Hat’s open‑source RamaLama to containerize large‑language and vision models. This lets astronauts get instant diagnostic help without a Earth link, a game‑changer for Moon or Mars missions where communication delays are critical.

Kubernetes' Device Management Working Group has pushed Dynamic Resource Allocation (DRA) to GA, giving clusters native support for GPUs, TPUs, NICs and other accelerators. The new API lets pods request and share specialized hardware after start, cutting the need for custom hacks and opening AI, edge and telecom workloads to reliable, large‑scale scheduling.

Cloudflare’s Workers Images binding was silently dropping large image data because of a rare race condition in the Rust hyper HTTP library. After six weeks of debugging, engineers patched the library with just four lines, restoring reliable image processing at edge scale.

OpenAI announced Jalapeño, its first custom Intelligence Processor built with Broadcom, to run large language models more efficiently. Early tests show higher performance per watt than current GPUs, letting OpenAI cut reliance on Nvidia and tighten control over the AI inference stack.

AWS Lambda now offers MicroVMs, giving each user or AI‑generated task its own lightweight VM with Firecracker isolation, rapid startup and persistent state. This bridges the gap between containers and full VMs, letting developers build interactive, multi‑tenant services without custom virtualization overhead.

Cloudflare launched the open‑source One stack, a library of agent skills that can plan, deploy and migrate Zero Trust environments. It encodes thousands of hours of migration expertise, including ready‑made logic for moving from Zscaler or Palo Alto to Cloudflare One, letting teams automate what used to take weeks.

Chainguard has made its Java Libraries generally available, delivering drop‑in, back‑ported fixes for dozens of critical and high‑severity CVEs across the Spring Boot ecosystem. By swapping a single Maven coordinate, teams can patch legacy Spring Boot versions without code changes, preserving auditability and accelerating delivery.