TLDR DevOps 2026-05-04

Cursor, Pierre, Discord, Reddit, Bazel, Meta, Intercom. All running CI on Buildkite. Silently.

That's kind of the point. We've spent thirteen years making CI that is architected once and then mostly forgotten.

When you do have a fail or flake, we provide immediate insight.

Boring outcomes.

The opposite of boring CI is "interesting" CI, and... nobody wants interesting CI.

Our 30-day trial unlocks everything, no credit card, no auto-upgrade. There's a real engineer (his name's Ola) on standby if you get stuck.

Start building here →

Kubernetes v1.36 rolled out Pod-Level Resource Managers as an alpha feature, allowing performance-critical pods to allocate exclusive, NUMA-aligned resources to main application containers while sidecars share a separate pod-level resource pool. The enhancement solves a longstanding trade-off where users previously had to either waste resources by giving every container (including lightweight sidecars) exclusive CPU allocations or forfeit the pod's Guaranteed QoS class entirely, with practical applications for ML training, high-frequency trading, and low-latency databases.

Amazon CloudWatch introduces a visual editor in the EC2 console to configure and manage the CloudWatch agent without JSON, enabling graphical setup, one-click deployment, automated policies, and fleet-wide observability across instances at no additional cost beyond standard usage pricing.

OpenTofu migration to Terraform is typically a near drop-in replacement requiring registry updates, CI/CD changes, version alignment, and careful state backups, with orchestration platforms like Spacelift supporting a smooth transition.

Rapid AI-driven growth is increasing system load and tempo for platforms like GitHub and Anthropic, causing saturation risks, outages, and architectural strain. LLMs boost productivity but may exacerbate reliability challenges rather than mitigate them.

Fully agentic coding creates “cognitive debt” by distancing developers from the code, weakening the judgment and debugging skills needed to supervise AI-generated work. The better approach is to use AI as a secondary tool for planning, research, and small delegations while staying actively involved in implementation and only generating code you can fully review.

Broadcom emphasizes private cloud resurgence driven by AI and data sovereignty concerns, integrating Kubernetes as VCF's core while advancing open source adoption and unified platform engineering for cloud native workloads at scale.

A better GitHub replacement should focus less on editor-like features and more on open-source maintainer coordination across projects, especially dependencies, downstream users, active forks, release impact, and migration signals. The core idea is that modern software reuse happens through package manifests and dependency graphs, so forges should treat dependencies as first-class relationships with downstream testing, dependent feeds, safer CI defaults, package caching, and better project-status visibility.

Meta introduced two major updates to its end-to-end encrypted backup system for WhatsApp and Messenger, including over-the-air fleet key distribution that allows Messenger to update HSM security keys without app updates (using validation bundles signed by both Cloudflare and Meta), and a commitment to publicly publish evidence of secure HSM fleet deployments on their blog. The HSM-based Backup Key Vault stores users' recovery codes in tamper-resistant hardware that Meta claims neither the company nor any third party can access, with the system deployed across multiple datacenters using majority-consensus replication.

Meta built a unified AI agent platform that uses standardized tools and encoded senior-engineer expertise to find, diagnose, and fix infrastructure performance issues at hyperscale.

At NAB Show 2026, widespread AI adoption highlighted a gap in operating and observing AI systems.