2026W35
A few interesting articles I read over the past few days
- You should never be angry at work — The line that got me was “you are not Linus Torvalds or Bryan Cantrill”, because the famous angry engineers can afford anger only since they already control the decisions, and that’s not the position most of us are arguing from. The mechanism he describes is the uncomfortable part, an angry colleague quietly becomes a problem to route around, so people start backchanneling and excluding you, which feeds exactly the frustration that started it. I’ve felt the caring-overflow thing he names, where anger is just competence spilling past the edge of the tube, and I want to believe his claim that he gets more done now while calm than he did when volatile. Still chewing on the redirect-your-caring-toward-shareholder-value advice, it sounds right as self-management but a little bleak as a description of what work should feel like.
- How we saved 100 terabytes of memory by optimizing 1.1.1.1’s DNS cache — The framing that reset my sense of scale is that at 250 billion cached entries, wasting one byte per entry costs 250 gigabytes, so this is a post about squeezing individual struct fields and it still adds up to 100 terabytes. Most of the wins come from one observation, cached DNS responses never change, which means the capacity field in a Vec is dead weight and swapping to Box slices reclaims 64 bytes an entry. My favorite move is inferring the owner name from the cache key instead of storing it, since it trades a tiny bit of read-time work for dropping the allocation on roughly 80% of records. The counterintuitive result is that shrinking memory made it faster too, insert throughput up 43% and lookup latency down 19%, because a tighter layout is just better for the CPU cache. Good reminder that “make it smaller” and “make it slower” are not the same axis.
- Creepy crawlies — What makes this infuriating is the sheer waste, the whole kernel repo is trivially cloneable, and yet the scrapers pick the single most expensive path, rendering 1.48 million commits across 922 forks as HTML page by page. The number that lands is that only 2% of the roughly 6 million daily requests are real developers, and 14 to 16 of 90 cores are permanently pinned just serving bots. The detail that actually worries me is the arms race, Anubis proof-of-work worked until the bots started solving difficulty-5 puzzles, which means the training data is worth enough to burn compute on. And the residential-proxy swarm through compromised IoT devices kills IP blocking as a defense, so there’s no clean answer here, just a slow tax on shared infrastructure that everyone else pays for.
- lessons in latency — The reframe I’m taking from this is latency as a budget, if the page needs to render in 100ms then serving it in 50ms buys your customer nothing, it just costs you more. That flips optimization once you’re inside the threshold, the goal stops being speed and becomes capacity and cost. The distinction I keep underselling is that average latency sets your throughput via Little’s Law, but it’s p99 that actually eats the budget, so two systems with the same 2ms average can behave completely differently when one has a 1% tail at 100ms. The storage point ties it together nicely, cheaper high-latency storage needs more replicas to hit the same throughput, so the “savings” quietly claw themselves back. Useful lens for arguing against optimization work that feels virtuous but moves no real number.
- Me divorcié de una vida de 10 años — Lo que más se me quedó es el reencuadre del título, no se divorció de una persona sino de un plan de vida entero, la casa comprada, el trabajo estable, el guión que sus padres le enseñaron. Su frase de que llevaba años intentando encajar en un molde que ya no era suyo describe algo muy concreto, tenerlo todo y aun así sentirse vacía. La semana de despedidas, desayuno de empresa, firma del divorcio y vuelo con cinco maletas es brutal por lo comprimida que fue. Y la idea final me parece la honesta, salir de una vida de diez años no termina al firmar ni al cerrar la maleta, termina mucho después cuando te reconoces en lo que estás construyendo.
- Your intellectual fly is open — The title is the whole argument, people notice the AI tells in your writing, the emoji bullets and the “it’s not just X but also Y” cadence, and no one points it out, so you keep walking around exposed. The cost he names isn’t ugliness, it’s trust, once a reader suspects the post is generated they stop believing it’s real, which poisons everything you actually did write. I appreciate that he isn’t anti-LLM, he draws the line at authorship, brainstorming and editing keep your voice central while letting the tool write for you erases it. Funny detail aside, calling LinkedIn the Gerald Ford of social networks is a good jab, it survived mostly because its rivals imploded. Sits a little pointedly next to me handing an agent my own reading notes, though the line he draws is exactly the one I’m trying to hold.