Upstream and Downstream Are Not Directions
The incident review started with a sentence that sounded ordinary. “The downstream service returned bad data.” Everyone knew what the speaker meant.
piljoong.dev
Senior staff engineer writing about backend systems, distributed IDs, and real-world architecture tradeoffs.
Latest posts
Turning Agent Work Into a System
In the last post, I argued that prompting is not the system. The missing layer was the system that keeps work in a known, controlled state while it changes. That is what I started calling Ordo.
From Prompting to Context. Still Not the System.
Early on, prompting felt like the core problem. If you wrote better prompts, you got better results. That framing made sense for a while.
AI Changed How We Build. It Did Not Change What Matters.
A view on engineering in the AI age AI is powerful. That is no longer a theoretical point. I spend far less time writing code from scratch than I used to. With a short prompt, I can generate something that would have taken hours or days before. I use that constantly now, and I don’t think there is any serious way to deny that the economics of building software have changed.
The 10-Join Query Nobody Wanted to Debug
The alert came in during business hours. A seller-facing marketplace listing page was showing incorrect items. Products that should never appear in standard listings were suddenly visible. This wasn’t a cosmetic issue. It was a correctness problem in a read path users relied on.
Distributed ID Formats Are Architectural Commitments, Not Just Data Types
Most systems start with auto-increment IDs because it’s the easiest possible thing that works. The database hands you numbers, you store them, life is good. There’s something comforting about watching IDs tick upward in perfect sequence—12345, 12346, 12347.