2026-03-31
So from the beginning of time (1843 Lovelace) humankind has been hacking around how to talk to computers. We’ve spent our time making languages to abstract primitives of machines ever since then.
A couple of years back some jerkface billionaire thought it would be funny “hell engrish is sorta like a language, so let us make some more or less neural net dojo squat on oss and the web, filtering out all the porn and rude assholes on social media, and get it to predict probable answers to some questions”.
The questions started innocent enough, with “what the hell is wrong with people?”, “why is my toe a yellowish black color?” and “what is the lyrics to the lumberjack song again?”. But then those engineering types thought: “Let’s make it spit out code, cause github was sucked up along with all the pecan pie recipes!”, and the end was nigh.
As the dojos made better and better text prediction factories, the engineering types made it do more and more of your job, to feed their own dopamine addiction and to earn brownie points from managemunt “Look what I just built! Isn’t it amazing? It’s almost there, just need to add a verification loop or six strapped on to prove it is not smoking doobie too much”.
So what’s left?
Your good old path to production is riddled with a priori sprinkles of assumptions, trust in meatbags to not check CNN every 10m while doing a 1000 line refactor, and even if they do, it’s ok, they will just take a little longer to get through all the context switching, and yeah, ball game is on this evening - got to watch!
I digress… where was I again? O yes. Your pipes and their flakiness.
You’d think you have it covered by adding a merge request, with a “not the same guy” rule here and a test environment here, ya know so people can test if they think it’s required. But ya don’t.
You need to move that warm and fuzzy trust you have in your team to something deterministically verifiable - your pipeline.
Most parts of your pipes are going to be replaced by cheap “Absolutely! Great question!” probability interns, so you had better fix that now.
Before developer Dave is too fwacking tired of reviewing this microslop from copilots he never liked in the first place, he just clicks the blue “Merge It” button. Before testing Tammy is in a rush to pick up her kids and forgets to test the “what could go wrong if you allow that <insert sql code here> feature into production feature flag which must never be turned on” edge case.
Pipelines are supposed to be feature funnels you plug into the mouth of your golden duck product, seeing how much you can expand its liver before it is ready to IPO and sell as a delicacy that will make you millions, so you can retire in that swimming pool of beer you always wanted… o, the bubbles on your skin, the smell…
I digress again.
Pipelines are about verifying that your product is still on track, that the system that runs your product still works, that it can handle the load, that it can pass under the glaze of the security team’s Sauron eye like a little hobbit, that it is good for the environment (tm). But mostly features baby, what the customer wants and we can sell, that’s what matters!
You don’t just one-shot gastown the hell out of your system (yet) and actually want proof that the whole town didn’t start fibbing and has actually started forming warring clans fighting each other via your system’s code.
You had better ensure that it is doing its pre-flight checks, in-flight operations, and post-flight maintenance well, else your duck will get hepatitis and your customers will jump on another duck they think they can trust more.
Proof of your system’s operational pudding.
If you want to prove your duck isn’t dying, read Claude’s thoughts on the matter.