Most saved links die in the bookmarks bar. These didn't. I've read each of them enough times to quote them in arguments, and they still hand me something new on every pass. Short list by design; if it's here, it earned it.
Every ECR question I've ever fielded, answered in one place. IRSA instead of token-refresh cron hacks, lifecycle policies before the storage bill grows teeth, scan-on-push wired into admission control. I reopen it before touching any registry work.
Seniority as blast radius, not years. Opens with a candidate who nails the system design in eight minutes and still doesn't get the offer, then explains exactly why. I reread it before every interview loop I run.
$30k/month down to $16.5k with receipts. The money hides three levels deep in Cost Explorer usage types: cross-AZ NAT traffic, orphaned Secrets Manager entries, idle IPv4. Cost as a platform default, not a cleanup sprint.
Reads OpenAI's decade of scaling posts closer than most of us read our own runbooks, then lands the right conclusion: copy the questions, not the 7,500-node architecture. “Can a node be Ready but still unsafe for GPU work?” lives in my head now.
Your dashboard says traffic is flat while users watch the model think forever. The framing that fixes it: the HTTP request is only the envelope, the GPU sees token work. I quote this whenever someone proposes autoscaling inference on request count.
Why does a pod resolving api.github.com fire five DNS queries? ndots:5 plus search domains, demonstrated in live CoreDNS logs instead of asserted. The trailing-dot trick at the end is the bit I keep re-explaining to people.
