标签：#t

I rent six dedicated servers from a company in Germany. Together they have more cores, more memory, and more SSD than most production clusters I worked on a decade ago. I run my own Kubernetes on them. Not managed. Not EKS. Not GKE. The whole stack, from the immutable OS up to the workloads. People who hear this ask me why, and the question usually arrives in one of two tones. The dangerous tone is "that's amazing, how do I do it" . The responsible tone is "why on earth would you do that to yourself" . This post is for the second group. Why the cloud is the right answer for almost everyone Let me get this out of the way honestly: by every conventional metric, I should be using the cloud. Managed Kubernetes has become genuinely good. EKS has dramatically improved over the last three years. GKE has always been better than people gave it credit for. The serverless options are mature. The serverless databases are mature. The observability is mature. The bill is predictable in the way a Tuesday is predictable. Self-hosting violates almost every assumption that makes a startup productive. Time is the most expensive resource you have. The cloud sells you abstractions that turn that time into product. Running your own substrate means the time goes into the substrate. If you are trying to ship a product to customers — go use the cloud. Stop reading this post. It will only confuse you. What the cloud does not sell you Here is what the cloud will not sell you, even if you are willing to pay extra: control over your own roadmap. The cloud's roadmap is the cloud's. They decide which APIs deprecate. They decide which regions get the new feature. They decide what your egress bill looks like. They decide whether your monitoring vendor — sitting on top of their infrastructure — is allowed to charge you eight times what it would cost you to host the same software yourself. They decide whether the small ML company hosting your fine-tuned model gets acquired by someone with very differ

Humanistic Architecture Learning to See Problems Differently I had the opportunity to attend a class called “Humanistic Architecture” by Mr. Chakrit Riddhagni. The class was about applying humanistic principles to software development . Before talking about what I learned, I would like to share a little about my own perspective on software development. Personally, I have a quote and a belief about software development: it is both a science and an art. I see software development as something that has an artistic side, while being grounded in logic, with almost endless possibilities. “Crafting software requires artistry, guided by imagination, grounded in logic, endlessly enduring.” This quote has been one of my inspirations since I started working as a developer. What I mean by this is that I have always liked thinking about software development as a kind of literature . We are not simply writing code. We are solving problems. We are developing and creating something to solve a problem that either we or our customers are facing. Because of this, as a developer, I naturally work with problems every day. But… I never really thought much deeper about what a “problem” actually is. Usually, when we solve a problem for a customer or develop software for them, we receive an issue or a scope of work and then start working on it. We know that something is a problem, so we focus on solving it. But we do not always stop and ask What is the actual structure of this problem ? Is this really the problem? And does the solution we are building actually address the problem we are trying to solve? That changed when I attended the “Humanistic Architecture” class. One of the biggest things I gained from this class was a new perspective and a set of tools for defining what a “problem” really is. Anatomy of a Problem We can look at a problem through the Anatomy of a Problem , which consists of three parts: Current State — where we are now Gap — the difference between where we are and wher

The bill that started this I was paying $40/month for a PDF generation API to power a tiny internal invoicing tool for a client project. Forty bucks a month to convert some JSON into a PDF. That's it. That's the whole service. I finally sat down on a Saturday to see if I could kill that subscription. Three weekends later, not only did I kill it — the replacement is faster than the API ever was, because there's no network round-trip at all. This post is the log of how it went, in the order I actually hit the problems, not the order that makes me look competent. Attempt #1: jsPDF on the main thread (it worked, until it didn't) First pass was the obvious one — jsPDF running directly in the click handler: function generateInvoice ( data ) { const doc = new jsPDF (); doc . text ( data . clientName , 20 , 20 ); data . lineItems . forEach (( item , i ) => { doc . text ( ` ${ item . description } — $ ${ item . amount } ` , 20 , 40 + i * 10 ); }); doc . save ( ' invoice.pdf ' ); } Fine for a 3-line invoice. Once I tested with a 40-line-item invoice (a real client sent me one to test against), the tab froze for almost two full seconds. Not crashed — frozen. Scroll didn't work, buttons didn't respond, and on a mid-range Android phone it was closer to five seconds. The main thread doing synchronous PDF math while also being responsible for painting the UI is exactly the kind of thing that looks fine in a demo and falls apart the moment a real user pastes in real data. Attempt #2: move it to a Web Worker Web Workers get talked about like they're this exotic tool for WASM and video processing. They're also just... a really good fit for "expensive synchronous work that a user is waiting on." I'd never reached for one before this project, mostly out of habit. The tricky part isn't the worker itself, it's that jsPDF assumes it has access to document and window in a couple of code paths (font metrics, mostly), which don't exist inside a worker. I ended up switching to pdfkit compiled

A random game is easier to debug when the same inputs produce the same result. In HoopTrait, a browser basketball project, the Lab mode combines eight selected traits and generates a fictional career. The interesting engineering problem is keeping replay, sharing and validation consistent. This is a technical development note, not a claim that a game score predicts an athlete's real performance. Store the decisions, not just the result The Lab state records a seed, a dataset version and an ordered list of actions. An action is a pick or a reroll. Replaying those actions reconstructs the build. A seed alone is not a complete replay contract: changing the player pool or its order can change a seeded draw. A dataset version therefore matters alongside the random seed. For a future release, the same principle should apply to changes in the rules themselves. Test invariants across many runs The Lab test suite iterates through 1,000 seeds. For each seed it shuffles the order of the eight skills, uses the two allowed rerolls, and completes a build. It checks that: Eight distinct players were selected. All eight traits are present, and no player remains to be drawn after completion. The overall game score stays between 0 and 99 and matches the shared rating function. Packing and unpacking the share state returns the original state. Recomputing the fictional career returns the same output. The ten simulated seasons sum to the displayed career earnings. Those assertions catch different problems. A stable score does not prove that a shared link reproduces the same selections. A complete build does not prove that its season totals add up. Reject impossible histories A share payload is untrusted input, even in a client-side game. Negative or fractional seeds, duplicate skill picks, a third reroll, unknown action types, a mismatched dataset version and actions after completion are rejected. The tests also cover malformed encoded payloads and unexpected fields. Local state is usef

The PS5 controller has various longevity concerns to consider, including its battery life and overall useful lifespan.

For better coverage or more wired connections, mesh, access points and Ethernet switches usually make more sense than a second active router.

Samsung's Z Fold series has improved the foldable phone form factor over time, but there remains an issue that the rumored iPhone Ultra may not remedy.

My harness flagged the model for sending the wrong arguments. It compared what the model actually...

Volatile and non-volatile memory are both vital parts of modern computers and devices, but how do they work together?

This is a step-by-step guide to checking the health of your Windows hard drive and what you can do if you get a bad result.

To get started with competitive Pokémon battles, all you need is your phone. But to compete at the highest level, like at the Pokémon World Championships, you're going to need a Switch. And there doesn't seem to be a good reason why. Battle simulator Pokémon Champions originally launched on the Switch and Switch 2 in […]

The Fairphone 6 Plus feels like an extremely average midrange Android phone and I couldn't be more thrilled. The mission has always been admirable. Fairphone seeks out ethically sourced materials and offers a high degree of repairability for its devices. But the phones themselves required a lot of sacrifices, like putting up with underpowered processors […]

Overview In this article, I'll introduce the concept of Multi-Version Concurrency Control (MVCC) and explain how Postgres implements this protocol across different isolation levels. I'm assuming you already have a basic understanding of isolation levels, database locks, and concurrency in general. I won't cover those concepts here, so if you're not familiar with them, I highly recommend checking out A Straightforward Guide for Isolation Levels first before continuing. The goal of this article is to help you understand: What Multi-Version Concurrency Control is How Postgres implements MVCC across different isolation levels Multi-Version Concurrency Control High-Level Concept The idea behind MVCC is simple: it's a protocol designed to accomplish one goal — when two or more transactions run concurrently on the same data, the end result should look as if those transactions ran one after another, in sequence. Take a look at the diagram above. Two transactions are running concurrently, and we want the end result to look as if either the first transaction ran and committed before the second one started, or vice versa. MVCC guarantees there are only two possible outcomes — never a third. But in reality, these transactions are running at the same time, so this is exactly the core idea of MVCC: it's a protocol that gives us this guarantee even though the transactions genuinely overlap in time. Note that other protocols aim for the same goal, like Two-Phase Locking and Optimistic Concurrency Control. They all take different approaches, but they're all working toward the same thing. The core idea of MVCC is that whenever a transaction updates a row, it doesn't mutate the value in place. Instead, it creates a new record as the latest version and links it back to the old version. After the update, the row has a new version, and the old version is never changed. This version chain exists per record — every update to a row creates a new version, and since each version is linked to

Docker management, monitoring that goes deeper than a green dot, backups I have actually restored, and everything else that showed up once deploying stopped being the hard part. Moving off Vercel solved exactly one problem: deploying. Everything else I used to get for free, quietly, as part of the platform, I now had to go find and wire up myself. A month into running my own server, I had a list of ten tools taped to the inside of my head, each one solving a problem I did not know I had until it happened to me at a bad time. This is that list, in the order I actually needed them, with the mistake or the moment that made me install each one. I lean JS and Rust wherever I can, partly out of preference and partly because those are the tools that keep pace with how fast the rest of my stack moves. A couple of these are not JS or Rust at all, and I kept them anyway because they were simply the best tool for the job. 1. Dokploy, for everything I wrote about yesterday This is the one I already spent an entire post on, so I will keep it short here. Push to main, Dokploy builds the container, Traefik points a domain at it, done. Four apps running on one $24 droplet, and adding a fifth would not move the bill. If you deploy anything with Docker and are still doing it over SSH, start here. Everything else on this list assumes you already have a platform under you, not just a server. 2. Neon, for the database half of preview environments The first crack after Dokploy was previews. Dokploy gives every pull request its own preview URL, which is one of the nicest things about the whole setup, right up until every preview hits the same production database. I corrupted a batch of test data twice before I noticed what was happening. Neon branches Postgres the way git branches code, copy-on-write, so a preview PR gets its own preview database that costs almost nothing until it actually diverges from main. The storage engine underneath is written in Rust, and it quietly closed the othe

An AI agent that has no job description will invent one. That is the whole reason agents drift, and it is the reason most of the agents I have seen deployed inside Indian businesses are quietly switched off within a few months of going live. Nobody would hire a person, point them at the office, and say "handle things". Yet that is exactly how most owners deploy an agent. They connect it to WhatsApp or email or the accounts folder, give it a paragraph of instructions, and let it run. Then they are surprised when it starts answering questions it was never meant to answer, promising delivery dates it cannot know, or filing something that a human should have looked at first. The fix is not a better model. It is the same discipline you already use for people: defined duties, an escalation path, a probation period with a review date, and one named person who is accountable for it. What drift actually looks like Drift is not a dramatic failure. It is a slow widening of scope that nobody approved. A distributor in the FMCG trade sets up an agent to acknowledge incoming orders on WhatsApp and log them into a sheet. Week one, it does that. Week three, a retailer asks "when will my stock reach?" and the agent, being helpful, answers with a guess. Week five, a retailer asks for a discount, and the agent, having seen discounts mentioned in earlier messages, offers one. None of this was in the brief. All of it followed naturally from "be helpful to customers", which is what the owner wrote because they did not know what else to write. By the time the owner notices, the agent has made commitments in writing to twenty retailers, and the sales team is cleaning up after it. The agent did not malfunction. It did what an unsupervised new employee does: it filled the vacuum with its own judgement. The mistake was upstream, at the moment of deployment. The job description A job description for an agent is not a prompt. It is a one-page document the owner can read and sign off, written in

In a previous post , I covered why TypeScript types alone can't protect you from a backend that returns something you didn't expect, and how to build a small apiRequest boundary that validates both the outgoing request and the incoming response against Zod-style schemas before your application ever touches the data. That post answered one question: Is this data actually shaped the way I think it is? It left another question open: When the answer is no, or when the request fails for a completely different reason (like a timeout or a dropped connection), what does the rest of the app do with that failure? In practice, "the rest of the app" usually does something different depending on who's writing it: One component checks error.response?.status directly. Another checks error.code === "ECONNABORTED" . A form manually digs through the error to find field-level messages. A toast just displays whatever string happens to be on error.message . The app works, but every layer speaks a different error dialect. This post is Part 2: it takes the validation boundary from Part 1 and builds the missing piece on top of it, a single, normalized ApiError shape that every layer of the app can speak, plus the logging, messaging, and form-mapping that make it actually usable. Quick Recap: The Validation Boundary From Part 1, the apiRequest wrapper validates request payloads and response bodies against schemas, and throws one of two typed errors when something doesn't match the contract: export class ApiRequestValidationError extends Error { constructor ( public readonly url : string , public override readonly cause : unknown ) { super ( `API request input does not match the contract for ${ url } .` ); this . name = " ApiRequestValidationError " ; } } export class ApiResponseValidationError extends Error { constructor ( public readonly url : string , public override readonly cause : unknown ) { super ( `API response does not match the contract for ${ url } .` ); this . name = " ApiRespon

Prologue: Open the window and look at the sky 🌌 Right now, while you read this, multi-million euros satellites are passing over your head. Some of them are transmitting pictures of the Earth to anyone who cares to listen. Not to their partner companies. Not to governments. To anyone . The images go out unencrypted, raining down over whole continents on frequencies you can pull in with about 100 euros of hardware and a balcony facing the sky. That is true for Earth pictures, for weather data, aviation/maritime tracking, amateur radio, even videos from the International Space Station 🛰️! I didn't know that either... until I did! And when I did, I felt a new frontier opening up, like I could reach the stars for real, from my home. New space adventures and discoveries were expecting me, one satellite at a time. This is a story about reaching (weather imagery) satellites. It's also a story of how a "let me just try one thing" turned into a distributed ground station running across two countries, a dead laptop resurrected as a server, my dad on a ladder in Théding while I watch from Barcelona and, as I write this, a geostationary satellite drifting slowly across the sky toward the one spot that would let me receive it. The project isn't finished. I mean that literally: this article ends on a spacecraft that is still in motion. But I should introduce the cast first, because the difference between two kinds of satellite is the key to this whole project. Weather satellites come in two families. The first are polar orbiters 🛰️. They fly low, a few hundred kilometers up, and fast, looping North to South while the Earth turns underneath them, so they eventually pass over everywhere (over your home too! 👾). They're the ones I chase in this story: the Russian METEOR-M series, 800 km over our heads, broadcasting down in the VHF band around 137 MHz . The catch is they're only up there for about ten to fifteen minutes at a time. The satellite climbs out of the horizon, arcs over us,

Background When streaming AI chat responses, Server-Sent Events (SSE) are commonly used. They are also adopted by APIs from OpenAI and Anthropic, as well as by MCP server responses. In fact, I implemented several AI chat projects that modified responses from AI platforms while streaming them to the browser. In doing so, I encountered an issue where SSE did not work because of certain intermediary proxies and load balancers, such as AWS App Runner. After taking a closer look at the SSE specification, I no longer felt that using SSE was right when the purpose was not actually event notification. The API's block data itself is JSON. This is also the same format as the structured logging sent to services such as CloudWatch Logs today (I had already been working on structuring application logs as JSON). Moving from SSE to JSON Lines Chunked Communication What I came up with was a combination of Transfer-Encoding: chunked and Content-Type: application/jsonl (which is not defined by the IAEA). With this approach, even if a proxy or load balancer buffers the response and returns it as a single body rather than chunks, only streaming is lost; the final complete data remains unchanged. Because it is JSON Lines (NDJSON), all you need to do is split on line feeds (LF) and JSON-parse each line. It is also easy to inspect in browser developer tools. However, implementing this from scratch every time is a bit of work, so I implemented and published jsonl-webstream , an npm library of stream utilities for browsers and servers (Node.js). The library has zero dependencies . tilfin / jsonl-webstream Lightweight library for JSON Lines web stream between browsers and Node.js environments jsonl-webstream Lightweight library for JSON Lines web stream between browsers and Node.js environments Overview This library provides utilities for processing JSON Lines formatted data through the Web Streams API It enables efficient streaming of JSON Lines data with minimal memory overhead across brow

Modern price monitoring systems need to do more than tell you that a price changed. A single abnormal listing, a scraped error, or a temporary outlier can make a traditional threshold-based detector fire an alert when nothing meaningful happened. In this project, I built a lightweight real-time price anomaly detector in Python that combines: A rolling median baseline Median Absolute Deviation (MAD) Robust Z-scores Short-term percentage returns Trend confirmation Alert cooldowns The goal is simple: detect meaningful price movements without overreacting to noisy observations. Note: This project monitors retail prices from Google Shopping results through SerpApi. It is a retail-price monitoring example, not a financial exchange-data feed. What we're building The pipeline looks like this: ┌──────────────────────┐ │ SerpApi / Shopping │ └──────────┬───────────┘ │ ▼ ┌──────────────────┐ │ Price Extraction │ │ + Validation │ └────────┬─────────┘ │ ▼ ┌────────────────────┐ │ Rolling Price │ │ History │ └────────┬───────────┘ │ ┌───────────┼───────────┐ ▼ ▼ ▼ Median MAD Return % │ │ │ └───────────┼───────────┘ ▼ Robust Z-score │ ▼ Trend Confirmation │ ▼ Signal Engine │ ▼ Alert Cooldown The implementation is intentionally small and interpretable. The complete engine is built around a single PriceAlertEngine class and a compact AnomalyResult data structure. Why not just use standard deviation? A common first implementation is: price > mean + 3 * standard_deviation The problem is that standard deviation is sensitive to extreme observations. Suppose your historical prices are: 990, 995, 999, 1001, 1005 Then one bad observation such as: 1500 can distort the mean and standard deviation. That can move your detection boundary away from the actual market behavior you are trying to model. For a noisy retail environment, a more robust baseline is useful. That's where median and Median Absolute Deviation come in. 1. Building a rolling median baseline Instead of storing an unlimited stre

With the right setup, an old, unused phone can be repurposed into a security camera. Even better? It's free.