Content hashing for expensive stages
A late-stage failure in my pipeline meant rerunning from the top, which meant paying for image generation again....
Writing
Notes on systems, tooling, and the craft of shipping software.
A late-stage failure in my pipeline meant rerunning from the top, which meant paying for image generation again....
My pipeline called a language model to plan every scene in a video — shot type, camera angle,...
I built cost tracking, wired a rate table into it, and got a number I trusted. Then I...
I locked a server down, checked ufw status, and saw no rule for port 8000. Then I opened...
Terraform code rots faster than application code. Not because the language is bad, but because it’s easy to...
I needed a shared telemetry module reachable from every stage of a pipeline whose stages live in separate...
My scene planner divided total duration by ten and generated one image per slot. A 25-minute video produced...
A CLI tool can be fast, correct, and still feel wrong. The difference between a tool people tolerate...
My renderer printed this on every run: Hardware accelerated encoding disabled - "crf" option is not supported with...
The best deploy pipeline is the one you never think about. Code merges, tests run, artifacts build, production...
Your renderer works on the host. Inside a container it dies partway through, usually with something as unhelpful...
Everyone building on model APIs knows roughly what their bill is. Almost nobody knows what a single unit...
Your renderer works locally. You containerise it, run the same job, and every line of Persian is a...
I turned on concurrency in a generation pipeline and my cost telemetry started under-reporting. Not by a visible...
My pipeline generated 149 images in 26 minutes. The model itself returns in about 4 seconds. That accounting...