Goodput, Not Throughput, Is the Right Metric for LLM Serving
A CNCF benchmarking study shows one configuration hit 50% more throughput while degrading user experience 10x. The metric you're optimizing for is probably wrong.
How 121 Million gRPC Connections Exposed a Fatal DNS Assumption
Bitdrift's T20 World Cup incident: a single DNS routing policy caused 80% errors when 121M devices connected at once. The fix required zero code changes.
AWS DevOps Agent + Kiro CLI: The Incident Remediation Loop That Fixes Itself
AWS has published a production blueprint where DevOps Agent investigates incidents and Kiro CLI applies the fix — reducing MTTR by 75% with a single human touchpoint: PR approval.
Amazon SQS at 20: How a Simple Queue Became the Backbone of Cloud Infrastructure
Launched July 13, 2006 alongside EC2 and S3, Amazon SQS turns 20 with 70,000 TPS throughput and 1 MiB payloads. What two decades of iteration reveals about distributed systems design.
Is a Kubernetes Pod the Right Deployment Unit for an AI Agent?
CNCF's kagent team asks whether Pods — ideal for long-lived microservices — are the right abstraction for AI agents that wake briefly, execute, and go idle. Their answer: probably not.
MAPFRE's Fraud Detection Jumped 135% Accuracy — Using EMR Serverless and Graph Analytics
How MAPFRE USA replaced rules-based insurance fraud detection with a graph-enriched ML pipeline on AWS, producing $6.8M in 2025 savings and 50–135% accuracy improvements.
March Networks Moved 5,600 TB of Video to AWS — and Cut Storage Costs by 79%
How enterprise video surveillance found its cloud-native architecture: March Networks' S3 and Glacier stack delivers $1.35M in annual savings for a single retail deployment.
What Running Kubernetes at Million-Cluster Scale Taught AWS About Zonal Failures
AWS operates hundreds of thousands of EKS clusters across 30+ regions. The core lesson from that scale: most availability problems don't come from components failing — they come from components reacting in ways that make failures worse. Here's what static stability means in practice.
Why Retrieval Quality Is the Defining Challenge in AI Agent Architecture
Agentic systems have two jobs: build context, then use it. Most failures that look like LLM problems start in the context-building step. Hybrid search with RRF is the current production standard — here's what that means for your stack.
Specification-Driven Data Workflows on AWS: Why Regulated Industries Are Adopting This Pattern
AWS architects outline a three-layer pattern — intent, composition, processing — that separates what a data workflow should do from how it executes. Clinical trials report onboarding datasets in days, not weeks.
How S&P Global Hit 15-Minute RTO on Capital IQ Using Amazon FSx and NetApp ONTAP
S&P Global's Capital IQ platform now achieves sub-15-minute read-only failover using Amazon FSx SnapMirror and FlexClone pre-provisioning. The design turns disaster recovery into a traffic redirect.
How AWS Engineers Trace Agentic AI Failures with OpenTelemetry and OpenSearch
AWS has built a portable observability pattern for agentic AI using OpenTelemetry traces and OpenSearch Agent Health — making wrong tool calls, latency spirals, and context cache failures inspectable before they hit production.
Coinbase Runs 1,200 AI Agents — and Cut Its AI Bill in Half
Coinbase deployed 1,200 full-time AI agents, shrank engineering teams from 10 to 2, and halved its AI spend while token usage climbed. The strategy: route every task to the cheapest model that can handle it.
Sandboxing Your AI Agents Is Necessary. It's Not Enough.
CNCF's July 2026 analysis of production agent fleets identifies the gap that sandboxing alone leaves open: resource efficiency at scale. Isolation solves security. Serverless-style runtimes solve the fleet management problem. You need both.
AWS Brings Claude Sonnet 5 to Bedrock and Launches WorkSpaces for AI Agents
AWS's July 6 roundup covers two additions for production AI builders: Claude Sonnet 5 on Amazon Bedrock and WorkSpaces for AI Agents reaching general availability for desktop-app automation.
Platform Engineering 2.0: Five Pillars for AI-Native Workloads
Platform Engineering 1.0 solved developer delivery. Platform Engineering 2.0 must solve GPU allocation, model serving, agentic governance, and FinOps for AI consumption. Here's the framework.
The 4-Body Problem of SRE: Why Autonomous Operations Require Unified Context
Autonomous SRE operations don't fail because models aren't good enough. They fail because the four bodies of operational knowledge — code, infrastructure, signals, and tribal wisdom — stay fragmented.
Data Sovereignty Is No Longer a Geography Problem — It's an Architecture Problem
The EU's Cloud and AI Development Act reframes data sovereignty as jurisdictional control, not physical location. Here's how platform engineers are enforcing it through Kubernetes admission controllers, OpenStack, and policy-as-code.
kpt Joins CNCF: The Kubernetes Config Toolchain That Shows You What You're Deploying Before You Deploy It
kpt's CNCF sandbox debut reintroduces a Kubernetes configuration toolchain built around WYSIWYG deployment — you inspect the final manifest before it hits the cluster, not a template that generates it.
AWS EKS Kubernetes Version Rollbacks: Upgrading Is No Longer a One-Way Door
Amazon EKS now lets teams reverse a Kubernetes version upgrade within 7 days. Here's how rollback works, what the limits are, and what it changes for platform engineers.
Why Traditional CI/CD Fails for LLMs—and the Release Gates That Fix It
CI/CD was built for deterministic code. LLMs aren't deterministic. Here's the core mismatch and the threshold-based release gates teams are building to address it.
AWS Certificate Manager Adds ACME Support for Automated TLS
AWS Certificate Manager now runs a fully managed ACME server endpoint — automating TLS certificate issuance and renewal via Certbot, cert-manager, or any ACMEv2 client.
AWS CloudFormation Express Mode: 4x Faster, Zero Extra Cost
AWS CloudFormation's new Express mode cuts deployment times by up to 4x by skipping stabilization wait times — enabling sub-minute feedback loops for AI-assisted infrastructure at no added charge.
Lessons from Running 1 Million Lambda Functions: What Scale-to-Zero Actually Costs
Building true-serverless multi-tenant SaaS at one million Lambda functions exposes problems documentation doesn't mention — synchronized scheduling, observability cost explosions, and the SQS idle trap.