Mikael Lirbank

Built a new agent skill for @neon.com's agent-skills repo. It helps AI coding agents find the query patterns that quietly run up your Postgres egress bill.

1/ Some constraints are real. Some are assumptions that sound plausible and never get tested. It's worth knowing which is which before they shape your decisions. I had one recently. 🧵

6/ The cross-provider penalty is effectively zero when both providers have data centers in the same metro. Provider boundaries and geographic proximity are not the same constraint.

1/ Some constraints are real. Some are assumptions that sound plausible and never get tested. It's worth knowing which is which before they shape your decisions. I had one recently. 🧵

4/ The thing that actually matters: geographic distance. Same Cloudflare infrastructure, same database, different regions: Far (San Jose → Virginia): 742ms Near (Ashburn → Virginia): 31ms That's 23x. Distance is real.

1/ Some constraints are real. Some are assumptions that sound plausible and never get tested. It's worth knowing which is which before they shape your decisions. I had one recently. 🧵

2/ Conventional wisdom: your database and application need to be on the same cloud provider. Cross-provider means crossing the internet, and crossing the internet means slow. It's the kind of thing that drives platform choices.

Good test coverage gives us confidence and prevents regressions. But testing for race conditions is hard. Even when tests run concurrently, the timing almost never lines up to trigger the bug. The test passes whether our code handles concurrency correctly or not.

5b Lock present - test passes. Lock removed - test fails. That's proof.

Good test coverage gives us confidence and prevents regressions. But testing for race conditions is hard. Even when tests run concurrently, the timing almost never lines up to trigger the bug. The test passes whether our code handles concurrency correctly or not.

4. What we actually need is a write lock. The second transaction blocks until the first one commits, then reads the updated value. Race condition gone. But now we have a new problem - how do we prove the lock is actually doing its job?

Good test coverage gives us confidence and prevents regressions. But testing for race conditions is hard. Even when tests run concurrently, the timing almost never lines up to trigger the bug. The test passes whether our code handles concurrency correctly or not.

2. Here's what that looks like. Two $50 credits hit an account with $100 at the same time. Both operations read the balance as 100. Both write 150. Final balance: $150 instead of $200. One credit vanished. No error. No rollback. Nothing in the logs.

🎉 Black Friday special: 100% off Neon Testing - grab your copy today! Just released v2.3.0 with support for Neon's expiring branches. Neon Testing tears down branches immediately after tests complete. But...

Hey @leerob.com is Bun 1.2 support something that needs to be done on the Vercel side?

So you know, these mean the same AWS data center: Vercel: Washington, D.C., USA (East) – iad1 Neon: AWS US East (N. Virginia) AWS Code: us-east-1 AWS Name: US East (N. Virginia) I find it especially helpful that @vercel.com name it "Washington, D.C" since AWS call it "N. Virginia".

I mean, so good! Classic, enduring, timeless.

Let's make it easier to pick a meta framework for your next project: - Next.js is for websites (great for e-commerce, CMS) - TanStack Start is for webapps (great for Figma, Excalidraw style apps) @t3.gg @tannerlinsley.com @leerob.com agree?