TLDR DevOps 2026-02-25

Most AI SRE evaluations are set up to fail before they start:

• 80% of AI projects fail, often due to misaligned expectations and undefined metrics
• PoCs scoped under 3 months are 3X more likely to reach production
• Most PoCs run in sandbox environments, so anything the tool learns won't transfer to production

The RunLLM AI SRE Buyer's Guide covers what to measure, what to avoid, and six criteria that separate tools that work in production from tools that only look good in demos. Free to download.

Kubernetes v1.35 ("Timbernetes") shipped with stable in-place Pod resource resizing, gang scheduling for distributed AI workloads, and a new default to KYAML output format designed to reduce configuration errors. The release addresses operational pain points for teams running mixed production workloads—particularly AI/ML jobs—by enabling coordinated Pod placement, CPU/memory adjustments without container restarts, and stricter YAML formatting guardrails.

Pulumi released a new onError hook feature that lets developers automatically retry failed cloud infrastructure operations or handle errors with custom code, useful for transient failures like DNS propagation delays or resource readiness issues. The feature is available in Node, Python, Go, and .NET SDKs as of version 3.219.0.

Amazon Elastic Container Service now offers native blue-green deployments, enabling zero-downtime releases with built-in lifecycle hooks, bake time, and rollback without requiring AWS CodeDeploy. While CodeDeploy still supports canary and linear traffic shifting and CodePipeline integration, ECS native is the recommended default for simpler, service-centric deployments in CDK.

AI has revived lines of code as a success metric. When code is nearly free to generate, volume measures nothing and correlates with more duplication, churn, and false productivity gains. Instead, engineering leaders should track outcomes like time-to-value, code durability, defect rates, and team comprehension, which reflect real impact rather than output.

Agentic DevOps tools like GitHub Copilot and Microsoft's Azure SRE Agent shift CI CD from deterministic pipelines to AI-driven judgment that can autonomously debug tests, remediate incidents, and reduce toil. While increasing velocity, they introduce opaque decision-making, novel failure modes, and new operational risks that require guardrails, shadow deployments, and strong oversight.

A Ruby developer spent 300 days introducing RuboCop into a 1M+ line legacy Rails monolith, enabling cops incrementally and fixing offenses through hundreds of small, reviewable PRs instead of a big-bang migration. The effort reduced PR review noise, uncovered real bugs and risky patterns, led to upstream contributions to RuboCop, and concluded once the remaining cops offered diminishing returns relative to ROI.

Azure Quick Review azqr was completely refactored using GitHub Copilot plan and agent modes under architectural supervision, eliminating major technical debt and 72 scanner packages. The overhaul removed 16,506 net lines across 560 files and improved scan performance by up to 90 percent.

Office comedy meets myth-busting tech. See why many of the world's largest companies still trust Big Iron in Broadcom's hilarious web series. Start watching episodes now.

Ubuntu 26.04 LTS 'Resolute Raccoon' will default to OpenJDK 25 when it launches on April 23, matching the long-term support cycles of both the OS and Java version.

Datadog Code Coverage provides unified, line-level visibility across repositories, enforces consistent testing standards, and uses automated test generation to identify and fill coverage gaps, helping teams maintain code quality amid faster, AI-assisted development.

The Rust compiler team has launched an anonymous 5-minute Debugging Survey (open until March 13) to better understand developers' struggles with debugging Rust.