Pascal (@paschals_robert) header
Pascal (@paschals_robert) profile picture

Full-stack engineer building https://teapaw.org/t.co/Nsd77USNDo Substack publication: https://teapaw.org/t.co/GTqCQdQBYw Drop a follow

/etc/init.d
Joined March 2022
heypascal.me is LIVE. My home on the internet. Months of building, rewriting, second-guessing every pixel. Full-stack engineer, founder, writer — one site, whole picture. If you've got something to build — an API, an AWS setup, a document workflow — the work-with-me section has real prices and a direct way to start. No pitch deck required. Here's what I built and what it cost me 🧵
2
3
20
1,348
The first person to use your backend will do the thing you never imagined: pass the id as a string, call the endpoints in the wrong order, send the request twice because the first one was slow. They are not using it wrong. They are showing you where your assumptions were.
1
5
I build for the second call, not the first: the duplicate request that should return the same answer, the missing field that should get a clear error instead of a crash, the slow response that should not make the client send it again.
1
7
Every confusing error message is a user telling you what they expected. Listen to it before you add a feature. Follow along if your first users keep surprising you. I write about building backends that survive contact with real people.
7
Every system has a job that runs at 2am and nobody remembers setting up. It sends the digest, cleans the old rows, renews the token. It works until the day it does not, and nobody notices until a user asks why the report never arrived.
1
4
I treat a scheduled job like a service with an owner: it logs every run, it tells someone when it fails, and it has a name that says what it does. A job that fails silently is not a background task, it is a bug with a delay on it. The fix is not a better job.
1
3
It is the alert that fires the first time the job misses, not the tenth. Follow along if your system has jobs nobody watches. I write about keeping small backends honest.
5
Your health check can take down your whole fleet. 10 app servers behind a load balancer. Each /health queries the shared database. Good advice: check something real. Then the database blips for 30 seconds. Every server fails the check at the same moment. The load balancer ejects all 10. Requests that never touched the database start failing. When the database recovers, the servers stay out until they pass several checks in a row. Auto scaling starts replacing healthy machines. No single server was broken. The check couldn't tell "I'm sick" from "the thing we all share is sick". The split I use now: /ready → the load balancer. Can THIS server do its own work? Shutting down? Event loop lagging? /health/deep → dashboards and alerts. Database, queue, cache, each probe with a short timeout. A database blip now pages a person instead of emptying the load balancer. Route on what one server owns. Report on everything else.
9
A 500 page that says "Something went wrong" is a dead end. The user does not know if their data saved, whether to retry, or who to tell. The error page is the moment the system talks to the user when it is at its worst, and most systems mumble.
1
7
A 500 that says the request was not processed and it is safe to try again, or that it was processed and not to send it again. The error is not the failure. The silence after it is.
1
5
If the user has to guess whether their payment went through, the bug is not the crash, it is the message. Follow along if your error messages leave users guessing. I write about the backend details that only show up when things break.
3
Clarity does not arrive before the first step. The first step is what makes it. I am in a phase I cannot see clearly, and I stopped waiting for the whole road to appear.
2
9
I wake up in the same room, nothing changed, and I push through on my laptop, my determination, and the principles I set for myself. No feedback. Just the work.
1
1
11
I want to start clean. One honest step, then the next. What is the one step that would make your next decision clearer? Next: The still voice.
1
8
A rate limit is not a punishment for using the API too much. It is the part of the design that says how much the system can honestly do, and it says it before the system falls over instead of after.
1
1
32
The limit protects both sides. It protects my system from one client taking everything, and it protects the client from building on a throughput I cannot keep. Say the number out loud before launch, not after the first outage.
1
15
Follow along if you build APIs other people depend on. I write about the backend decisions that only show up when real traffic arrives.
1
1
15