标签：#performance

In this blog post, we will see how to performance test a RAG (Retrieval-Augmented Generation) application properly, covering both speed and correctness, and how to wire both into a CI/CD pipeline so regressions get caught before they reach production. Performance testing a RAG application requires two separate testing gates: one for speed and one for answer quality. Traditional load testing tools measure response times but cannot detect hallucinations, where a model returns fast but factually incorrect answers grounded in fabricated context rather than retrieved documents. The guide demonstrates using k6 for load testing end-to-end latency and DeepEval for evaluating faithfulness and answer relevancy using an LLM-as-judge approach. Both gates are integrated into a GitHub Actions CI/CD pipeline so regressions in either performance or output quality are caught automatically on every pull request before reaching production. If you've come from a JMeter or k6 background like I have, your first instinct with a RAG endpoint is probably to point a load test at it and check response times. That gets you halfway there. A RAG app can return a fast, confident, completely wrong answer, and a plain load test will never tell you that. You need two testing surfaces, not one: performance and quality. This guide covers both, using a single running example throughout: a documentation assistant that answers "How do I run JMeter in non-GUI mode?" against a small knowledge base. Why RAG breaks traditional load testing assumptions A conventional API returns a complete response and you measure the round trip. A RAG endpoint does two expensive things before it answers: it retrieves context from a vector store or search index, then it streams a generated response token by token. That second part matters a lot. A single request can stream hundreds of tokens over several seconds, so "request duration" as a single number hides two very different problems: how long the model took to start answe

Todo dev brasileiro que já precisou consultar CNPJ conhece o dilema: ou você usa uma API que faz proxy da Receita (3 a 10 segundos por consulta, quando não cai), ou baixa o dump de dados abertos e monta a própria base — e descobre que "baixar um CSV" era a parte fácil. Eu montei a própria base. Este post é o diário honesto do que funcionou, do que quebrou e dos números reais — 217 milhões de linhas servidas em ~10ms de p50 dentro do datacenter, num Postgres de 1 vCPU. A arquitetura em uma frase Não consulte a Receita em tempo real. Ingira o dump mensal e sirva da sua infra. O resto é decorrência. Receita (dump mensal, ~6GB zip) ──▶ ingestão Go (COPY) ──▶ Postgres ──▶ API (chi) CGU (CEIS/CNEP, zip diário) ──▶ job diário ──┘ O dump da Receita: as pegadinhas que ninguém documenta O layout oficial existe, mas o que quebra parser de verdade é o que está fora dele: Encoding latin1 (ISO-8859-1) — acento vira lixo se você ler como UTF-8. Em Go: charmap.ISO8859_1.NewDecoder() num transform.Reader streaming. Decimal com vírgula ( "1000000,00" ) e datas YYYYMMDD onde 0 e 00000000 significam nulo. CNPJ quebrado em 3 colunas (básico 8 + ordem 4 + DV 2). A chave de junção entre empresas, estabelecimentos e sócios é o básico — errar isso custa um dia. As partições 0–9 não se alinham entre arquivos. O estabelecimento da partição 3 pode ser de uma empresa da partição 7. Foreign key rígida entre as tabelas = COPY quebrando no meio da carga. A solução: sem FK; a integridade vem da fonte. Bytes NUL ( 0x00 ) no meio dos dados. O Postgres rejeita NUL em text . Um strings.ReplaceAll(s, "\x00", "") no parser economizou três recargas. Desde jan/2026 o repositório é um Nextcloud do SERPRO+ com WebDAV público — dá pra listar meses com PROPFIND e baixar com o token do share como usuário. Adeus, scraping. COPY ou morte A diferença entre INSERT em lote e o protocolo COPY não é incremental — é outra categoria. Com pgx.CopyFrom e lotes de 50k: 28,1 milhões de empresas em 1m28s (~320k linhas/s) num

PostgreSQL provides several planner configuration parameters, such as enable_seqscan and enable_indexscan , that influence how execution plans are generated. These settings affect planning, not the execution of an already-generated plan. With prepared statements, this raises an interesting question. Should planner settings be applied before PREPARE, before EXECUTE, or both? Let's look at a simple example: a "tasks" table with a due date and a "done" status: \ c drop table if exists tasks ; -- a table of tasks with status (done or not) and due date create table tasks ( id bigint generated always as identity primary key , due timestamptz , done boolean ); -- insert 500 tasks, with 1% not done insert into tasks ( due , done ) select now () + interval '1 day' * n , 42 != n % 100 from generate_series ( 1 , 500 ) n ; -- index the todo (partial index) create index on tasks ( due , id ) where done = false ; vacuum analyze tasks ; With a partial index, I indexed only the tasks that are not yet done ( done = false ) because that's my most frequent query pattern: postgres =# explain select id , due , done from tasks where done = false and id > 0 order by due limit 1 ; QUERY PLAN --------------------------------------------------------------------------------------- Limit ( cost = 0 . 13 .. 3 . 60 rows = 1 width = 17 ) -> Index Scan using tasks_due_id_idx1 on tasks ( cost = 0 . 13 .. 17 . 47 rows = 5 width = 17 ) Index Cond : ( id > 0 ) ( 3 rows ) With partial indexes, the condition covered by the index is not even visible in the execution plan because the index itself enforces the condition. Prepared statement I decided to use a prepared statement with all values as parameters. It is probably not a good idea in this case. When a parameter can have only a few different values and you expect different cardinalities for each, you should probably define one query per value, using literals. I'm doing this to illustrate what can happen, with a simple, extreme example: postgres =# pr

Introduction We already covered picking the right model tier for the task and caching a large shared prefix in https://pg-blogs.netlify.app/posts/11-building-reliable-llm-apps-in-java/ . Those two lines were the tip of a bigger discipline: LLM cost is not a fixed line item, it's an engineering variable — one you can measure and shrink with the same rigor you'd apply to database query time or container memory. This post goes deeper: how input/output pricing actually works, the exact cache_control shape and how to prove a cache hit rather than assume one, the Batches API for work that isn't latency-sensitive, and model routing — using a cheap model to triage, escalating only the hard cases to a stronger one. The honest framing throughout: measure before you optimize. Every technique here has a cost of its own; applied to the wrong workload, "optimization" makes things slower or more expensive. Token Economics: Why the Prefix Is the Bill Anthropic (like every hosted LLM provider) prices input and output tokens separately, and output is always pricier — the model has to generate output autoregressively, one token informed by all the ones before it, while input can be processed in parallel. Representative pricing from the current model catalog: Model Input Output Claude Opus 4.8 $5.00 / MTok $25.00 / MTok Claude Sonnet 5 $3.00 / MTok $15.00 / MTok Claude Haiku 4.5 $1.00 / MTok $5.00 / MTok Two consequences follow directly: Long system prompts, tool definitions, and RAG context are read on every request , not written once. A 20K-token system prompt sent on every one of 10,000 requests is 200M input tokens — at Opus 4.8 rates, $1,000 before a single output token is generated. The shared prefix , not the user's question, is usually where the money goes. A verbose model wastes money twice — once on the extra output tokens themselves, and again because the next turn's messages history now carries that verbosity forward as input on every subsequent call. Trimming max_tokens an

Introduction https://pg-blogs.netlify.app/posts/10-building-reliable-llm-apps-in-python/ closed with a section on picking the right model per task and caching a shared prefix. That was the entry point into a bigger discipline: LLM spend is an engineering variable, not a fixed bill — one you can measure and reduce with the same rigor you'd apply to query latency or memory footprint. This post goes deeper on four levers: how input/output pricing actually works and why the prefix is usually where the money goes, the exact cache_control shape and how to prove a cache hit instead of assuming one, the Batches API for work that isn't latency-sensitive, and model routing — a cheap model triaging requests and escalating only the hard ones. The throughline is honest: measure before you optimize. Every lever here has its own cost; misapplied, it makes things slower or pricier, not cheaper. Token Economics: Why the Prefix Is the Bill LLM providers price input and output tokens separately, and output always costs more — generation is autoregressive (each token depends on every one before it), while input can be processed in parallel. Representative pricing from the current model catalog: Model Input Output Claude Opus 4.8 $5.00 / MTok $25.00 / MTok Claude Sonnet 5 $3.00 / MTok $15.00 / MTok Claude Haiku 4.5 $1.00 / MTok $5.00 / MTok Two things follow: A long system prompt, tool list, or RAG context is billed as input on every request , not written once. Send a 20K-token system prompt on 10,000 requests and that's 200M input tokens — at Opus 4.8 rates, $1,000 before the model has generated a single output token. The shared prefix , not the user's actual question, is usually the dominant cost. Verbose output costs twice — once directly (more output tokens billed at the higher rate), and again because the next turn's history carries that verbosity forward as input. Asking for concise output and setting a sane max_tokens is a cost control, not just a style choice. This is why the tw

Introduction Fixing N+1 queries with select_related / prefetch_related or selectinload (see the previous post ) gets you down to a small, sane number of queries per request. The next bottleneck is what each query costs once the table has millions of rows — and that is almost always about indexing. An index turns "scan every row" into "look it up directly." Skip it, and a query that's instant in development takes seconds once real data volume shows up in production. How Indexes Work: The B-Tree Intuition Without an index, a WHERE clause forces a sequential scan : the database reads every row and checks the condition — O(n) , cost grows linearly with table size. An index is a separate, sorted structure (almost always a B-tree ) mapping column values to row locations. Because it's sorted and balanced, finding a value is a tree walk: O(log n) . On a 10-million-row table, that's the difference between reading 10 million rows and roughly 23 tree nodes. This isn't free: Writes get slower — every INSERT / UPDATE / DELETE on an indexed column also updates the index. Storage grows — each index is a sorted copy of (part of) the data. An index trades write cost and storage for read speed. Indexing a column you rarely filter or sort on is pure cost, no benefit. Reading Query Plans: EXPLAIN ANALYZE Postgres' EXPLAIN ANALYZE shows what the planner actually did, not an estimate. Before an index , filtering orders by customer_id : EXPLAIN ANALYZE SELECT * FROM orders WHERE customer_id = 48291 ; Seq Scan on orders (cost=0.00..21453.00 rows=42 width=96) (actual time=0.021..118.442 rows=41 loops=1) Filter: (customer_id = 48291) Rows Removed by Filter: 1199959 Planning Time: 0.112 ms Execution Time: 118.471 ms Seq Scan means Postgres read all ~1.2 million rows and discarded all but 41. actual time is real elapsed time — 118ms for one lookup. After CREATE INDEX idx_orders_customer_id ON orders (customer_id); : Index Scan using idx_orders_customer_id on orders (cost=0.42..8.53 rows=42 wid

Introduction Fixing N+1 queries (see the previous post ) gets your Hibernate app down to a handful of queries per request. The next bottleneck is what each of those queries costs once your tables have millions of rows — and that is almost always a question of indexing. An index turns "scan every row" into "look it up directly." Get the index wrong — or skip it — and a query that took 2ms in development takes 4 seconds in production once real data volume shows up. How Indexes Work: The B-Tree Intuition Without an index, a WHERE clause forces a sequential scan : the database reads every row and checks the condition. That's O(n) — cost grows linearly with table size. An index is a separate, sorted data structure (almost always a B-tree ) that maps column values to row locations. Because it's sorted and balanced, finding a value is a tree walk: O(log n) . On a 10-million-row table, that's the difference between reading 10 million rows and reading roughly 23 tree nodes. The cost is not free: Writes get slower. Every INSERT / UPDATE / DELETE on an indexed column must also update the index structure. Storage grows. Each index is a copy of (part of) the data, sorted differently. An index is a trade: you pay on every write so that specific reads become fast. Indexing a column you rarely filter or sort on is pure cost with no benefit. Reading Query Plans: EXPLAIN ANALYZE Postgres' EXPLAIN ANALYZE shows what the planner actually did — not what you hope it did. Before an index , filtering orders by customer_id : EXPLAIN ANALYZE SELECT * FROM orders WHERE customer_id = 48291 ; Seq Scan on orders (cost=0.00..21453.00 rows=42 width=96) (actual time=0.021..118.442 rows=41 loops=1) Filter: (customer_id = 48291) Rows Removed by Filter: 1199959 Planning Time: 0.112 ms Execution Time: 118.471 ms Seq Scan means Postgres read all ~1.2 million rows and threw away all but 41 of them. actual time is the real elapsed time, not an estimate — 118ms for one lookup. After CREATE INDEX idx_orders

How to Compress Images in the Browser with Canvas API Every image you upload to a "free" online compressor is sent to a server — often without you knowing what happens to it afterward. For a tool that processes your private photos, that's a terrible design. Here's how to build (or use) an image compressor that runs entirely in the browser using the HTML5 Canvas API. No uploads, no server costs, and unlimited file sizes. The Core Technique: Canvas toBlob() The key API is HTMLCanvasElement.toBlob() : js const canvas = document.createElement('canvas'); const ctx = canvas.getContext('2d'); const img = new Image(); img.onload = () => { canvas.width = img.naturalWidth; canvas.height = img.naturalHeight; ctx.drawImage(img, 0, 0); canvas.toBlob((blob) => { const url = URL.createObjectURL(blob); }, 'image/jpeg', 0.8); }; img.src = 'your-image.jpg'; The second parameter is the MIME type (image/jpeg, image/png, image/webp, image/avif). The third is quality (0–1). Step-Down Resizing for Large Images If you're compressing a 6000×4000 px photo, drawing it at full resolution onto a canvas can eat 70+ MB of memory. Step-down resizing halves the dimensions repeatedly: function stepDownEncode(img, maxDim, quality) { let w = img.naturalWidth; let h = img.naturalHeight; let src = img; while (w > maxDim * 2 || h > maxDim * 2) { w = Math.floor(w / 2); h = Math.floor(h / 2); const temp = document.createElement('canvas'); temp.width = w; temp.height = h; temp.getContext('2d').drawImage(src, 0, 0, w, h); src = temp; } const canvas = document.createElement('canvas'); canvas.width = w; canvas.height = h; canvas.getContext('2d').drawImage(src, 0, 0, w, h); return new Promise((resolve) => { canvas.toBlob((blob) => resolve(blob), 'image/jpeg', quality); }); } This prevents memory crashes and actually produces better quality (step-down preserves more detail than a single jump). Comparing Real-World Results Format Avg Original Avg Compressed Avg Savings JPEG → JPEG (Q80) 3.2 MB 0.8 MB 75% PNG → We

Why I'm Building the Fast Series I'm building the Fast Series because creator software has gotten too complicated. Plenty of tools are powerful, but they make you fight the software before you can make anything. You want to record a tutorial, stream a game, clip a useful moment, compress a file, or turn an idea into a short video. Instead, you're digging through settings, codecs, plugins, device permissions, export presets, and cryptic error messages. That's the problem I keep running into, and the Fast Series is my attempt to solve it: practical Windows software where each tool does one job clearly and reliably. Not everything needs to be a giant all-in-one platform. Sometimes the better product is a small tool that opens quickly, gives you sensible defaults, explains what's happening, and gets out of your way. That's the direction I'm taking with Sturm Technologies. The Problem With Creator Tools There are already great tools for recording, streaming, editing, clipping, and compressing. OBS is powerful. Professional editors are powerful. FFmpeg is powerful. There are cloud tools, browser tools, AI tools, and creator suites that promise to do everything. But power is not the same thing as clarity. Most creators don't want to become experts in capture APIs, bitrate math, encoder settings, audio routing, or export pipelines. They want to make something and publish it. The pain usually shows up in small moments. You record a video and the audio is missing. You compress a file and it still doesn't meet the upload limit. You spend more time scrubbing a long video than actually clipping it. You hit an error and the app hands you a technical dump instead of telling you what to fix. That's where I think there's room for better software. Not bigger software. Better software. Start With FastCast The first product in the series is FastCast , a Windows recording and streaming app for people who want OBS-level practicality without OBS-level setup. FastCast focuses on screen cap

Book: The Complete Guide to Go Programming Also by me: Hexagonal Architecture in Go — the companion book in the Thinking in Go series My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools Me: xgabriel.com | GitHub You reach for unsafe.Pointer to skip a copy. Maybe a []byte you want as a string without the allocation, maybe a struct you want to reinterpret as another. The code compiles. go vet stays quiet. Tests pass. Then, weeks later, under GC pressure, a rare crash shows up in a place that has nothing to do with your change. The Go unsafe package documentation is one long doc comment. It lists the conversions that are valid and warns that everything else is not portable and not guaranteed to keep working. The trouble is that "everything else" includes a lot of code that looks obviously correct. The rules are narrow on purpose. There are effectively four patterns you are allowed to write, and the standard library stays inside all four. Here they are, in the shape you will actually use them in Go 1.23+. The one fact under every rule unsafe.Pointer is a pointer the garbage collector understands. It keeps the object it points at alive and it moves with the object if the runtime relocates a stack. uintptr is a plain integer. The GC does not treat it as a reference. Convert a pointer to a uintptr , store that integer in a variable, and as far as the runtime is concerned nothing points at that memory anymore. That single fact is behind three of the four patterns. Every legal conversion either avoids uintptr entirely or keeps it inside one expression where the compiler can prove the pointer stays live. Pattern 1: reinterpret *T1 as *T2 with the same layout The first pattern converts a *T1 to *T2 when the two types have the same memory shape. You take the address, run it through unsafe.Pointer , and cast to the target pointer type. The standard library does exactly this in math.Float64bits : func Float64bits ( f float64

Introduction In large-scale Unity development, GC Alloc can quietly become a real problem. At first, nothing looks wrong. But as the project grows and you add more enemies, UI, master data, events, states, notifications, logs, and other systems, small allocations that happen every frame begin to pile up. LINQ is especially convenient. var aliveEnemies = enemies . Where ( x => x . IsAlive ) . OrderBy ( x => x . DistanceToPlayer ) . ToList (); It is readable. But if this kind of code runs every frame, it can become a source of both GC Alloc and CPU overhead. Unity's official documentation also recommends reducing frequent managed heap allocations as much as possible, ideally getting close to 0 bytes per frame. https://docs.unity3d.com/2022.3/Documentation/Manual/performance-garbage-collection-best-practices.html For general GC Alloc best practices, this article refers to the Unity 2022.3 documentation, because the general guidance still applies. Unity 6-specific GC behavior is covered later using the Unity 6.0 documentation. This article assumes Unity 6.0 as the minimum Unity version and explains how to choose between regular LINQ and ZLinq in production code. Unity 6.0 uses the Roslyn C# compiler, and its C# language version is C# 9.0. However, some C# 9 features, such as init-only setters, are not supported. https://docs.unity3d.com/6000.0/Documentation/Manual/csharp-compiler.html The short version The point of this article is not to ban LINQ completely. Do not use LINQ in hot paths just because it is readable. Do not assume ZLinq solves everything just because you introduced it. Those are the two main ideas. A rough guideline looks like this: Area Guideline Editor extensions, build scripts, debug code Regular LINQ is usually fine Startup, loading, initialization LINQ can be fine, but measure when data size is large Update / LateUpdate / FixedUpdate Avoid LINQ by default Code that is not per-frame but still called frequently Consider ZLinq Code that materializes res

Introduction Most projects start with a single repository. Imagine you're building an e-commerce platform: a Next.js storefront for customers, a React Native mobile app, and a NestJS backend API. Splitting these into three repositories feels like the obvious, clean solution. ecommerce-web ecommerce-mobile ecommerce-api For the first few months, this works fine. Then the project grows, and the cracks start to show: You copy utility functions between repositories instead of importing them. You duplicate TypeScript interfaces across the frontend, mobile app, and API. Your frontend and backend drift apart because each repo defines its own version of the same models. Updating one shared component means editing it in three different places. Eventually, maintaining the project becomes harder than building new features. If that sounds familiar, you're not alone — it's exactly the problem monorepos were designed to solve. In this article, we'll cover: What a monorepo actually is, and how it differs from a multi-repo (polyrepo) setup Why engineering teams choose it How Turborepo makes monorepos fast instead of slow A practical, production-ready structure for a Next.js + React Native + NestJS monorepo Common mistakes and best practices Whether you work with React, Next.js, React Native, or NestJS, these concepts will help you build projects that scale without becoming a maintenance burden. The Problem With Multiple Repositories A typical multi-repo setup looks like this: web-app/ mobile-app/ backend-api/ shared-components/ Each repository has its own package.json , dependencies, CI/CD pipeline, Git history, and versioning strategy. It looks clean at first — but as the project grows, several problems appear. 1. Code duplication You write a helper function once: export function formatPrice ( price : number ) { return `$ ${ price . toFixed ( 2 )} ` ; } Both the web app and the mobile app need it, so instead of importing it, someone copies it. Now there are two versions. When one

Your portfolio is often the first impression a recruiter, client, or fellow developer gets of you. If it loads slowly, ranks nowhere on Google, or is a pain to update, it's working against you instead of for you. Here's how I approached optimizing mine — covering performance, SEO, and everyday usability. 1. Start With a Lightweight Foundation The biggest performance wins come before you write a single line of custom code. Pick a lean stack. Static site generators (Astro, Next.js with static export, Hugo, or even plain HTML/CSS/JS) ship far less JavaScript than a full SPA framework for a mostly-static portfolio. Avoid unnecessary UI libraries. A heavy component library for a five-page site adds kilobytes you don't need. Hand-roll simple components instead. Use system fonts or self-host web fonts. Pulling fonts from a third-party CDN adds an extra DNS lookup and render-blocking request. Self-hosting with font-display: swap avoids layout shift and speeds up first paint. 2. Optimize Images (This Is Usually the Biggest Win) Images are almost always the heaviest assets on a portfolio site. Convert images to WebP or AVIF — typically 30–50% smaller than JPEG/PNG at the same visual quality. Resize before upload. Don't serve a 4000px-wide photo in a 600px container. Use loading="lazy" on below-the-fold images so the browser doesn't fetch them until needed. Add explicit width and height attributes to prevent layout shift (this also helps your Cumulative Layout Shift score). <img src= "/project-thumb.webp" alt= "Project screenshot" width= "600" height= "400" loading= "lazy" /> 3. Minimize and Defer JavaScript Ship only the JS a page actually needs — code-split per route if your framework supports it. Defer non-critical scripts (analytics, chat widgets) with defer or load them after the page is interactive. Audit your bundle with a tool like source-map-explorer or your framework's built-in bundle analyzer to catch unexpectedly large dependencies. 4. Nail the SEO Basics Good perf

Introduction to Hydration and Rendering Strategies In the relentless pursuit of faster, more responsive web applications, developers have engineered a spectrum of hydration and rendering strategies . Each approach emerges as a response to specific performance bottlenecks, yet none is universally optimal. This section dissects the core mechanics of these strategies, their historical evolution, and the critical problem they aim to solve—balancing speed with practicality. The Problem: A Trade-Off Landscape At its core, the challenge is mechanical : how to deliver content to the user’s browser with minimal latency while maintaining interactivity. Traditional rendering methods (e.g., server-side rendering) prioritize initial load speed but often defer interactivity until JavaScript execution. Client-side rendering, conversely, delays the first paint but ensures seamless interactions post-hydration. The tension between these extremes has birthed hybrid strategies like incremental hydration and islands architecture , each addressing specific failure points in the rendering pipeline. Key Mechanisms Driving Strategy Evolution Advancements in Web Technologies : New APIs (e.g., Web Components, Streaming SSR) enable finer-grained control over rendering. For instance, streaming SSR reduces Time-to-First-Byte (TTFB) by sending HTML in chunks, but risks breaking the causal chain of DOM hydration if not synchronized with client-side scripts. User Expectations : Sub-second load times are no longer aspirational but expected. This pressure deforms traditional workflows, pushing developers toward pre-rendering or static site generation (SSG), which trade dynamic flexibility for speed by offloading rendering to build time. Competitive Pressure : Performance is a zero-sum game. Companies adopt strategies like partial hydration (hydrating only interactive components) to minimize JavaScript payload, but this risks breaking interactivity if the hydration boundary is misaligned with user int

Every data grid demo looks incredible with twenty rows. The columns line up. The hover state is tasteful. The checkbox has confidence. Someone scrolls three inches and everyone quietly agrees that software has advanced. Then the real product arrives. Fifty thousand rows. Twenty columns. Editable money. A custom status cell. Filters. Sorting. Horizontal scrolling. A user who pastes something suspicious from Excel. A product manager asking whether the total row can stay pinned while the server is slow. That is when a table stops being a table and starts becoming infrastructure. So we built a benchmark. Not a perfect benchmark. Those do not exist. A useful one. What we measured The fixture is intentionally boring: 50,000 deterministic rows 20 fixed-width columns 1,200 by 600 pixel viewport two editable columns sorting filtering virtual scrolling production bundles fresh browser contexts raw samples committed to GitHub No network requests. No paid-only feature tricks. No images. No grouping. No heroic demo code designed to make one library look blessed by destiny. The report measures: JS gzip : reachable JavaScript after gzip Ready median : navigation until the grid adapter mounts and two animation frames pass Scroll settle : one scripted vertical and horizontal jump plus animation frames Mounted cells : body cells in the DOM after the scroll Interaction health : heap, long tasks, estimated FPS, dropped frames Live benchmark: https://vitashev.github.io/react-data-grid-benchmark/ Source and raw samples: https://github.com/Vitashev/react-data-grid-benchmark The part most benchmarks get wrong Not every grid exposes the same surface. For example, MUI X Data Grid Community uses 100-row pagination for this workload. That is a valid product boundary, but it is not the same as continuously virtualizing 50,000 rows. So the ranked tables include only compatible continuous-scroll libraries. MUI remains in the fixture and raw data, but not in the leaderboard. That makes the benchma

Let's be honest—traditional A/B testing is broken. If you've ever implemented client-side testing, you know the drill. Your users load the page, wait for JavaScript to execute, and then—flicker—the content changes. It's jarring. It hurts your Core Web Vitals. And worst of all, your experiments might be measuring user frustration instead of genuine engagement. But what if you could run A/B tests with zero flicker? What if you could modify your HTML before it even reaches the browser? That's exactly what Cloudflare Workers make possible . The Server-Side Advantage A/B testing at the edge means making decisions about which variant a user sees before any HTML is sent to their browser . Instead of: Load the page Execute JavaScript Flicker Apply the variant Track the result You get: Decision made at the edge Correct HTML streamed immediately Zero flicker Better performance Companies like Ninetailed are already using Cloudflare Workers to achieve 2-3x cost savings compared to traditional serverless platforms, all while delivering personalized experiences with minimal latency . How It Actually Works The concept is surprisingly elegant. Your Cloudflare Worker intercepts incoming requests, checks for a cookie to maintain user consistency, and routes users to the appropriate variant . Here's a simplified version that's production-ready: javascript const NAME = "myExampleWorkersABTest"; export default { async fetch(req) { const url = new URL(req.url); // Determine which group this user is in const cookie = req.headers.get("cookie"); if (cookie && cookie.includes(`${NAME}=control`)) { url.pathname = "/control" + url.pathname; } else if (cookie && cookie.includes(`${NAME}=test`)) { url.pathname = "/test" + url.pathname; } else { // New user—randomly assign them (50/50 split) const group = Math.random() < 0.5 ? "test" : "control"; url.pathname = `/${group}` + url.pathname; // Fetch and modify the response to set a cookie let res = await fetch(url); res = new Response(res.body, res

Modern web application architectures typically run each layer in its own process, like a NodeJS server and database both running in their own processes. They communicate via a network connection or localhost socket. This separation introduces protocol overheads, TCP stack latency, and data serialization/deserialization costs on every single query. Planck is designed around the concept of Zero-Distance Architecture that co-locates data and application code. It combines the database engine and a WebAssembly application runtime into a single, unified process. By running your application code directly inside the database process, database calls become direct in-memory function calls rather than network round-trips. This article provides a practical guide to getting started with Planck. We will look at the core toolchain, walk through setting up a self-contained local benchmark, compare its performance against a NodeJS, ExpressJS, MongoDB stack, and look at how to build more complex features. The Toolchain: Planck, planctl, and Workbench Running and managing a zero-distance app requires three main components. Planck itself is the core binary. It functions as both the storage engine (a WiscKey-style, LSM-tree-based engine) and the WebAssembly host. Instead of running a database in one process and your application server in another, you run a single Planck process. It loads your compiled WebAssembly application directly into its memory, running it in the same process space as the database. To manage this runtime, you use planctl. This is the command-line tool for developers. It handles the compilation of your code, packages it, and deploys it to the Planck host. It also allows you to perform database operations, like creating stores and defining indexes, export/import, backup/restore directly from your terminal. Finally, there is the Workbench. This is a web console that comes built into the platform. It provides a visual dashboard to monitor your applications, view databa

I needed to convert 500 product images from one format to another. Server-based solutions quoted $15-50/month for batch processing. So I built a client-side solution using Web Workers and OffscreenCanvas. The Architecture The key insight: Canvas operations on large images block the main thread. The fix: Web Workers handle image decoding/encoding off the main thread OffscreenCanvas renders without DOM access — perfect for worker contexts Transferable objects pass image data between workers with zero-copy const worker = new Worker ( ' processor.js ' ); const canvas = new OffscreenCanvas ( 800 , 600 ); // Worker processes image, main thread stays responsive Real Performance Processing 500 images (average 2MB each) on a mid-range laptop: Server upload approach: 12 minutes (mostly upload time) Browser-local with Workers: 3 minutes 40 seconds Memory usage: Stable at ~400MB with proper cleanup The Tools I packaged this into webp2png.io for batch WebP conversion and svg2png.org for vector batch processing. For barcode generation, genbarcode.org uses similar worker-based rendering for bulk label generation. If you're processing more than 50 images, Workers + OffscreenCanvas is the way to go. Your server bill will thank you.

Erik Steiger discusses the operational pain of legacy PDF generation in regulated banking and manufacturing. He explains how transitioning from resource-heavy engines like Puppeteer and LaTeX to a serverless Rust architecture powered by Typst can drop render latencies below 2ms. He shares how applying Git and Docker concepts to template registries ensures ironclad compliance and rapid debugging. By Erik Steiger

Denormalization has been the standard approach to analytical data modeling for good reason. Moving joins, lookups, and business rules out of query time and into ingestion gives you the fastest possible reads for a known access pattern. For most of the past decade, it was often the practical default for latency-sensitive analytics. Earlier columnar engines and distributed query processors could execute joins, but many workloads paid for them through higher latency, higher compute cost, spill-to-disk, or distributed coordination overhead. That constraint has loosened. Modern columnar databases with advanced join algorithms have reduced the cost of runtime joins enough that normalization is now a genuinely viable option for many analytical workloads. Denormalization still delivers faster reads, but normalization can bring operational benefits: simpler pipelines, flexible schemas, and cleaner governance. Engineers can now make the decision based on their actual workload characteristics, rather than being forced into one approach by engine limitations. This guide is a decision framework for making that choice in ClickHouse. It starts with why denormalization became the default, explains what has changed in join performance, then compares the tradeoffs on both sides so you can decide where to denormalize, where to join, and where to use ClickHouse primitives that bridge the gap. For a broader evaluation framework covering latency, concurrency, ingest throughput, SQL flexibility, and cost across real-time OLAP options, see our guide to choosing a database for real-time analytics in 2026 . For a deeper comparison of how ClickHouse executes star schema joins against Druid, Pinot, and cloud DWHs, see our star schema and fast joins guide . TL;DR Denormalization and normalization are both valid modeling strategies. The right choice depends on your workload. Denormalization's tradeoffs are primarily operational : pipeline complexity, write-path overhead, data freshness lag, back