Vercel, AWS, GCP, observability, IaC, and reliability engineering.
Edge middleware promised sub-50ms execution. Our p95 said otherwise. Here's what we found when we instrumented it properly, and the three changes that brought latency back under control.
A DynamoDB throttle event left our Terraform state half-written and locked. Here's the postmortem, the recovery steps, and the guardrails we added so it doesn't happen again.
A single NAT Gateway line item quietly ate our cloud budget. Here's the traffic audit, the VPC endpoint rollout, and the gotchas nobody mentions in the AWS docs.
We ran the same Next.js API workload on Cloud Run and Lambda for three months. Cold starts, cost, observability, and one nasty timeout bug shaped the decision.
A scheduled job that hadn't fired in six hours, no alert, no error in Sentry, and a billing email that didn't get sent. Here's exactly what broke, how we caught it, and the cron monitoring pattern we run now.
A war story about Sentry transactions, span ingestion, and a 6x bill spike — plus the dynamic sampling, SDK config, and quota guardrails we now ship by default.
We moved part of a production AWS estate from Terraform to Pulumi over six months. Here's what actually changed, what broke, and where we'd quietly stay on HCL.
We rolled out OpenTelemetry across a Node and Go fleet, picked tail-based sampling because everyone said to, and learned why head-based wins for most teams. Here's the tradeoff we wish someone had drawn for us.
Vercel's bandwidth and function invocation costs got loud at scale. We moved the hot read path to S3 + CloudFront while keeping the DX. Here's the architecture, the numbers, and what broke.
We moved a Next.js app from CloudFront + Lambda@Edge to Vercel and learned the hard way that signed cookies, edge regions, and middleware ordering don't translate cleanly. Here's what bit us.
We moved a 40-module Terraform monorepo to HCP Terraform Stacks. Here's what broke, what we gained, and the four decisions we'd reverse if we started over.
A real story of trimming RDS failover from a customer-visible 90 seconds down to roughly 12. The fix wasn't a bigger instance — it was DNS, connection pools, and Proxy.
A Cloud Run service that ran fine for eighteen months started timing out checkout on a Friday afternoon. The fix wasn't more CPU — it was a misread of how min-instances, concurrency, and startup CPU boost actually interact.
A quiet $40/day NAT Gateway line item turned into the second-largest cost on our AWS account. Here's how we found it, what was actually driving it, and the VPC endpoint plumbing that fixed it.
A single deploy turned a calm Sentry account into a $4k surprise. Here's what happened, what we changed, and how to stop event floods before finance notices.
We spent six months partially migrating a production AWS estate from Terraform to Pulumi. Here's what we kept, what we rolled back, and the boring reasons IaC choices rarely come down to language.
A war story about how head-based sampling in OpenTelemetry quietly hid a real latency regression for three weeks, and the tail-sampling setup we landed on after the incident.
A client's Vercel invoice tripled in a quarter. We spent two weeks tearing apart their Next.js app and shaved roughly 55% off the bill. Here's what mattered, what didn't, and what we'd do differently.
