标签：#orm

If confirmed, it would be the fly's first breach of the US-Mexico border.

Google has shared details of its fleet wide large scale A/B experimentation system designed to standardize experiment assignment, exposure logging, and configuration propagation across distributed services. The approach enables consistent measurement across products, reduces experiment conflicts, and improves reliability of data driven decision making at scale. By Leela Kumili

There's a specific feeling you get after your third production caching incident. It's not panic....

Your API is fast. Your code is clean. Your architecture looks solid on paper. Then you hit 500,000 records and everything slows down. Queries that ran in 12ms now take 4 seconds. Your dashboards lag. Users start filing support tickets. Your on-call engineer is staring at a query plan at midnight wondering what went wrong. Nine times out of ten, the answer is indexing. Not missing indexes — wrong indexes. Indexes that exist but don't help. Indexes that actively hurt write performance without meaningfully improving reads. This is a breakdown of the most damaging database indexing mistakes in production SaaS systems — and how to fix them before they become incidents. Mistake 1: Indexing Everything "Just in Case" The most common mistake isn't under-indexing. It's over-indexing out of anxiety. New engineers especially fall into this pattern — add an index on every column that appears in a WHERE clause, just to be safe. Seems responsible. It isn't. Every index you add is a write tax. On every INSERT, UPDATE, and DELETE, PostgreSQL (or MySQL) has to update every index on that table. On a table with 8 indexes, every write touches 8 data structures. At low volume, this is invisible. At 10,000 writes per minute, it becomes your bottleneck. The fix: Audit your indexes regularly. In PostgreSQL: SELECT schemaname , tablename , indexname , idx_scan , idx_tup_read , idx_tup_fetch FROM pg_stat_user_indexes ORDER BY idx_scan ASC ; Any index with idx_scan = 0 or near zero hasn't been used since your last stats reset. That's a candidate for removal — not immediately, but after investigation. Mistake 2: Not Understanding Index Selectivity An index on a boolean column ( is_active , is_deleted ) is almost always useless. Here's why: selectivity measures how many distinct values exist relative to total rows. A boolean column has two values. If 95% of your rows have is_active = true , an index on that column tells the query planner almost nothing useful. It will often skip the index entire

This week's Java roundup for May 25th, 2026, features news highlighting: lifecycle changes with two of the JEPs that were targeted for JDK 27; the GA release of Koog 1.0; point releases of Hazelcast, Quarkus, Hibernate and JHipster; the eighth milestone release of Spring AI 2.0; and introducing Endive, a JVM-native WebAssembly (Wasm) runtime. By Michael Redlich

The Instagram accounts for the Obama White House and the Chief Master Sergeant of the U.S. Space Force were briefly defaced with pro-Iranian images and messages over the weekend, after instructions began circulating on Telegram showing how to trick Meta's "AI support assistant" bot into resetting account passwords.

This is a follow-up to SynaptoRoute: A Study in Local Semantic Routing . If you haven't read it, the short version is: SynaptoRoute is a zero-token semantic routing engine that classifies user queries into intents using local embeddings instead of LLM API calls. SynaptoRoute v0.3.0: Matching Semantic Router While Scaling to 50,000 Routes What Changed Since v0.2.0 When I published the first post, SynaptoRoute had just shipped dynamic batching and O(1) hot-reload. The throughput numbers were promising, but the accuracy story was incomplete. I had internal benchmarks but no comparison against a widely adopted baseline under identical, reproducible conditions. That gap is now closed. v0.3.0 is live on PyPI: pip install synaptoroute == 0.3.0 The Benchmarking Journey Getting to these numbers took multiple benchmark revisions. Early synthetic datasets produced catastrophic accuracy collapse and initially suggested that both SynaptoRoute and Semantic Router were performing poorly. After deeper investigation, the root cause turned out to be flaws in the dataset generation pipeline rather than limitations of the routing engines themselves. Several rounds of validation, failure analysis, threshold tuning, adversarial testing, and external benchmarking followed. All final results presented in this article come from independent public datasets with strict train/test separation, eliminating dataset leakage and benchmark inflation. That process was valuable because it forced the project to validate assumptions against real-world data instead of relying on synthetic benchmarks. The Benchmark That Actually Matters I evaluated SynaptoRoute against Semantic Router on two standard NLU datasets. Same embedding model ( BAAI/bge-small-en-v1.5 ). Same hardware. Same evaluation script. Same train/test splits loaded from HuggingFace. CLINC150 150 intents spanning 10 domains, plus an out-of-domain class. This is the standard stress test for intent routers. Metric SynaptoRoute Semantic Router

Shopify introduced GraphQL Cardinal, a new execution engine replacing depth-first traversal with breadth-first execution. The redesign improves large-scale GraphQL performance with up to 15x faster field execution, 6x lower GC overhead, and +4s P50 latency gains. It focuses on execution-layer efficiency and batched resolver processing for high-cardinality commerce queries. By Leela Kumili

An LLM-written kernel benchmarked 38% faster on a microbench. Here is what kernel-level validation showed it actually did at runtime. TL;DR Multi-agent LLMs are now writing CUDA kernels (RightNow AI’s AutoKernel, Meta’s KernelEvolve, a multi-agent system claiming 38% speedup on Blackwell). Source-level benchmarks measure clean throughput on a single isolated kernel. They do not measure SM occupancy under co-scheduling, DRAM bandwidth saturation, dispatcher off-CPU during a real serving workload, or NCCL wait correlation with sibling kernels. Kernel-level validation closes that gap: an eBPF trace of the same kernel running under the same workload as production answers all four questions in one capture. The kernel-writing wave Three pieces of work in April surfaced the same pattern: agents generate CUDA kernels, then quote a single throughput number against a baseline. RightNow AI’s AutoKernel (announced Apr 6) – LLM agents iteratively rewrite CUDA kernels for a target metric, claiming substantial speedups on selected microbenchmarks. Meta’s KernelEvolve – similar shape: agents propose kernel variants, rank by throughput, keep the best. Multi-agent system on Blackwell (Apr 29 reports) – claims a 38% speedup on a public kernel benchmark using a coordinated agent setup. All three are real research, all three produce real kernels, and all three report numbers that come from microbenchmarks. The microbench setup is exactly what you want for the optimization loop. It is not what you get in production. What microbenchmarks do not see Run an LLM-generated kernel under nvprof or nsight-compute on an otherwise-idle GPU and the throughput number is real. Put the same kernel in front of a vLLM serving workload and four properties change immediately: SM occupancy under co-scheduling. The kernel that achieves 95% SM occupancy in isolation will achieve 40-50% with three other kernels sharing the same SMs. The optimizer never sees this regime. DRAM bandwidth saturation. A kernel tha

Modern search engines do more than match keywords. When you search for "cozy mystery set in Paris" and get results for "atmospheric detective novel in France" that's vector search at work: documents and queries are converted into lists of numbers, called embeddings, and the search engine finds the documents whose numbers are closest to the query's. Manticore Search supports this natively. Under the hood, it uses a data structure called HNSW: a graph that connects nearby vectors, so it can find nearest neighbors quickly without scanning every document. That makes vector search fast enough to run on millions of documents in milliseconds. But HNSW has an inefficiency. Early in the traversal, almost every distance computation finds a better candidate than the ones already in the result set. As the search goes on, those improvements become rarer, but the algorithm keeps traversing the graph until it exhausts its exploration budget. By that point, the result set has often already converged, and the remaining work does little or nothing to improve it. Early termination fixes this by detecting that point and stopping early. The effect becomes more noticeable as k grows, where k is the number of nearest neighbors the query asks Manticore to return. Returning more neighbors requires more graph exploration, and much of that extra work happens after the result set has already stabilized. That also makes early termination more valuable, because it has more unnecessary work to cut. This gets more pronounced with vector quantization . Quantization compresses stored vectors to save memory, which slightly lowers search precision. To recover it, Manticore uses oversampling : it fetches 3x more candidates than requested, then rescores them using the original full-precision vectors. With the default 3x oversampling, HNSW explores many more candidates per query. Large k values often come from this kind of candidate expansion: an application may ask the vector index for hundreds or thous

Overview PDF text extraction is a common pre-processing step in data pipelines — ingesting research papers, legal documents, or reports before embedding or indexing. Both pypdf and PdfPig are pure managed-code parsers: no native binaries, no OCR, no system PDF renderer. They implement the same PDF specification operations in their respective languages. This makes the benchmark unusually clean: the performance difference is entirely due to language execution speed, not library architecture differences. Benchmark Setup 200 recent arXiv PDFs (mixed technical papers, 5–40 pages each). Tested on subsets of 10, 50, 100, and 200 files. Both libraries extract all text from all pages; output is validated for page-count agreement and character-count agreement within 15% (pypdf and PdfPig decode whitespace and encoding tables slightly differently). Results PDFs Pages Python (pypdf) .NET (PdfPig) Speedup 10 ~120 ~0.9 s ~230 ms 3.9× 50 ~600 ~4.2 s ~810 ms 5.2× 100 ~1,200 ~8.5 s ~1.4 s 6.1× 200 ~2,400 ~17 s ~2.7 s 6.2× The speedup grows slightly with corpus size, suggesting pypdf has a per-document startup cost that compounds as PdfPig's JIT gets warmer. Why PdfPig Is Faster PDF parsing is byte-heavy: every page is a stream of PostScript-like operators (move, show text, set font, etc.). Each operator must be lexed, looked up in a dispatch table, and executed against a graphics state machine. In Python, each operator dispatch is a Python method call — the CPython bytecode interpreter has overhead per call regardless of what the method does. In .NET, the JIT compiles the dispatch loop to native code the first time it runs; subsequent pages pay only the cost of the actual work. Additionally, PdfPig's content-stream parser operates on ReadOnlySpan<byte> — zero-copy slicing through the raw page bytes with no intermediate string allocations. pypdf builds Python string objects for each token. Key Code // PdfPig — zero-copy span-based page extraction public Result Extract ( string path )

Overview PageRank is the canonical graph algorithm. NetworkX implements it in pure Python — its dict-of-dict adjacency representation means every power-iteration step dispatches millions of Python attribute lookups. When the graph has 1.8 million nodes and 28.5 million edges (Wikipedia category hyperlinks), those lookups dominate the runtime. The .NET replacement uses a CSR (Compressed Sparse Row) matrix — two flat int[] arrays for the graph structure — and TensorPrimitives for the SIMD-accelerated normalization step inside each iteration. Benchmark Setup Five SNAP datasets of increasing size: Dataset Nodes Edges wiki-Vote 7,115 103,689 soc-Epinions1 75,879 508,837 web-Stanford 281,903 2,312,497 web-Google 875,713 5,105,039 wiki-topcats 1,791,489 28,511,807 Algorithm: power-iteration PageRank, damping=0.85, tol=1e-6. Both implementations converge to identical top-10 node rankings. Results Dataset Python (NetworkX) .NET (CSR) Speedup wiki-Vote (103k edges) ~0.8 s ~100 ms ~8× soc-Epinions1 (508k edges) ~8 s ~600 ms ~13× web-Stanford (2.3M edges) ~120 s ~5 s ~24× web-Google (5.1M edges) ~5.5 min ~12 s ~28× wiki-topcats (28.5M edges) ~47 min ~60 s ~47× The speedup grows with graph size because NetworkX's Python dispatch cost scales with edge count, while the CSR inner loop is a tight JIT-compiled SIMD pass. Why CSR Beats NetworkX NetworkX represents each node's neighbors as a Python dict. Iterating the adjacency in one power-iteration step means: Calling G.neighbors(node) — a Python method call Iterating a dict — unboxing int keys, chasing heap pointers Accumulating a float into another dict value — another boxing step That happens for every edge, every iteration, roughly 50–80 times to convergence. CSR collapses the graph to two arrays: rowPtr[n+1] (where each node's neighbors start) and colIdx[edges] (the neighbor list). Iterating neighbors of node v is a tight C loop from rowPtr[v] to rowPtr[v+1] . No Python objects, no dict hashing, no pointer chasing. Key Code // P

There is a fascinating psychological phenomenon in modern software engineering: the relentless pursuit of the upgrade. As frontend developers, we are conditioned to believe that speed and efficiency come from adopting the newest technologies. We migrate from Webpack to Vite to shave seconds off our build times. We transition between UI libraries in search of better reconciliation algorithms. We constantly audit our CI/CD pipelines. We treat performance as a destination we must reach by continuously adding or swapping out the moving parts of our toolchain. Yet, amidst this endless cycle of optimization, we consistently overlook the most sophisticated, highly optimized piece of software in our entire stack. It is the software you are using to read this article right now: the web browser. The Underappreciated Engine The modern browser is an absolute marvel of engineering. Over the past decade, teams of the world's most talented systems engineers have engaged in a fierce arms race to optimize browser engines like V8, SpiderMonkey, and JavaScriptCore. Today’s browsers feature Just-In-Time (JIT) compilation, sophisticated garbage collection, and massively parallelized rendering pipelines. They are capable of executing highly complex, interactive applications with a level of fluidity that was unimaginable a few years ago. However, when we evaluate developer tools—specifically the online IDE or the browser-based code editor—there is a stark contrast. The environments we use to write and test our code rarely reflect the speed of the engine they run inside. Why the Standard Web IDE Misses the Mark If you want to quickly prototype a component or isolate a bug, you will likely reach for a frontend playground or a popular Replit alternative. What happens next is often a masterclass in friction. The environment feels heavy. The interface is cluttered with features you didn't ask for. As you type, the live code editor experiences micro-stutters. The instant live preview isn't actu

Notes on Serving LLMs with TensorRT-LLM and Triton 2026-05-31 · LLM serving / NVIDIA stack These are working notes on taking an open-weights LLM from a Hugging Face checkpoint to a production-style serving endpoint on the NVIDIA stack — TensorRT-LLM for the engine, Triton Inference Server for the deployment surface — and benchmarking it honestly against vLLM on multi-GPU hardware. They follow the harness in trtllm-triton-serving (4× H100, NVLink). The goal is to move from "I use vLLM" to "I can stand up the NVIDIA inference stack on real multi-GPU hardware and reason about the trade-offs." 1. The serving pipeline The path from checkpoint to endpoint has four stages. Each one is a place where a decision affects latency, throughput, or accuracy: Checkpoint — a Hugging Face model. Engine build — compile to a TensorRT-LLM engine for a fixed tensor-parallel degree, precision, and batching policy. Model repository — wrap the engine in a Triton tensorrt_llm -backend model repo. Serving + load test — trtllm-serve (or Triton) exposes an OpenAI-compatible endpoint; a load generator drives it under controlled concurrency. The key mental shift from vLLM: TensorRT-LLM does ahead-of-time compilation . vLLM is a runtime that takes the model and serves it; TensorRT-LLM builds an engine specialized to your GPU, TP degree, and precision first. That build is where the performance comes from, and also where the rigidity comes from. 2. Tensor parallelism (TP) For a model that doesn't fit on one GPU — or to cut latency — TensorRT-LLM shards each layer across GPUs. On a 4× H100 NVLink box, TP=4 means every forward pass does an all-reduce across the four GPUs over NVLink. The all-reduce is not free. On this fabric it tops out around 77 % of the NVLink budget (see the separate NVLink-wall notes ). For prefill (large tensors) you're bandwidth-bound and TP helps. For decode (one token at a time) you're pinned against the small-message latency floor, and past a point more TP makes decode slowe

c_std: A Leak-Free, Cross-Platform Standard Library for Modern C Bringing the comfort of the C++ STL and Python's standard library to C17 — without leaving C A technical white paper. Executive summary C is still the substrate of the computing world — kernels, databases, language runtimes, embedded firmware, and the inner loops of nearly everything else. Yet the moment you step away from the kernel and try to write ordinary application code in C, you feel the gap: no growable vector, no hash map, no JSON parser, no string type that doesn't invite a buffer overflow. You either pull in a grab-bag of mismatched third-party libraries, each with its own conventions and failure modes, or you re-implement the same dynamic array for the hundredth time. c_std is an attempt to close that gap deliberately and coherently. It is a single, consistent library — written in pure C17 — that reimplements a large slice of the C++ Standard Library (containers, algorithms, smart pointers) alongside many Python-style conveniences ( json , regex , random , statistics , csv , config , even turtle graphics). It targets Windows and Linux from one source tree, compiles cleanly under -Wall -Wextra , and — this is the part I care about most — is verified leak-free under Valgrind , module by module, example by example. This paper explains the design philosophy, the architecture, and the engineering discipline that makes a library like this trustworthy enough to build on. 1. The problem: C's missing middle Every C programmer knows the two extremes. At the bottom, the language itself: pointers, malloc , memcpy , raw arrays. At the top, whatever the platform hands you — <windows.h> or POSIX, OpenSSL, a JSON library someone wrapped a decade ago. The middle — the layer the C++ STL and Python's batteries-included standard library occupy — is missing. That missing middle has a real cost. It shows up as: Re-invention. Teams write their own vector, their own string builder, their own linked list, each subt

Google Cloud's automated systems suspended Railway's production account without notice, triggering an eight-hour platform-wide outage affecting 3 million users. The cascade took down workloads across all providers including AWS and bare metal because Railway's control plane was hosted on GCP. Railway is demoting GCP to backup-only status. By Steef-Jan Wiggers

22 Astro Best Practices: The Bookmark-Worthy Tips At QuotyAI I'm using Astro to build landing pages and blog posts, so I have hands-on experience how to use it properly and how to vibe-code without headache. Astro is the best framework for content sites right now - #1 in developer satisfaction in the State of JS 2025 survey, with Cloudflare backing it since January 2026. But like any tool, it rewards people who use it the way it was designed. This is the reference I wish I had when I started. Whether you're building your first Astro project or vibe-coding a blog at 2am, these are the habits worth forming from day one. Heads up on versions: This article covers Astro 6.x (released March 2026) and Astro 6.4 (released May 2026). Some APIs from older tutorials are now deprecated - those are called out explicitly below. Always check the upgrade guide when moving between majors. 🖼️ Assets & Media 1. Use <Image /> instead of <img /> Astro's built-in <Image /> component does a lot of work at build time that plain <img> tags leave on the table: it converts images to WebP, generates the right width and height attributes to prevent layout shift, and compresses everything without you touching a single config file. --- import { Image } from 'astro:assets'; import hero from '../assets/hero.png'; --- <!-- ✅ Optimized: converted to WebP, compressed, no layout shift --> <Image src={hero} alt="Hero image" /> <!-- ❌ Skips all of that --> <img src="/hero.png" alt="Hero image" /> For art-direction scenarios (different images at different breakpoints), reach for <Picture /> instead. 2. Use the Astro 6 Built-in Fonts API Almost every website uses custom fonts, but getting them right is surprisingly complicated - performance tradeoffs, privacy concerns, self-hosting, fallback generation, and preload hints. Astro 6 added a built-in Fonts API that handles all of it for you. Configure your fonts in astro.config.mjs : // astro.config.mjs import { defineConfig , fontProviders } from ' astro/conf

Si vous utilisez PostgreSQL, vous avez probablement déjà dû choisir entre une clé primaire BIGSERIAL et un UUID. Depuis des années, la version 4 (aléatoire) est le choix par défaut quand on veut un identifiant unique et distribué. Mais en 2026, une alternative plus récente s’impose : UUID v7, qui intègre un timestamp et promet de meilleures performances pour les index. Dans cet article, je vous explique concrètement ce qui change, avec des benchmarks PostgreSQL et des exemples de code, pour que vous puissiez décider en connaissance de cause. UUID v4 : le standard aléatoire et son problème d’index Un UUID v4 est constitué de 122 bits aléatoires. Cette absence totale de tri est sa force pour l’unicité, mais elle devient un handicap dans un index B‑tree, qui est la structure utilisée par PostgreSQL pour les clés primaires. Lorsque vous insérez un nouvel UUID v4, il a autant de chances de se retrouver au début de l’index qu’à la fin. Résultat : l’index se fragmente, les pages se remplissent mal, et les performances d’écriture se dégradent à mesure que la table grossit. J’ai reproduit un test simple sur PostgreSQL 16 avec 10 millions de lignes, en utilisant une table dont la seule différence est la colonne id : -- Table UUID v4 CREATE TABLE events_v4 ( id UUID DEFAULT gen_random_uuid () PRIMARY KEY , payload JSONB , created_at TIMESTAMPTZ DEFAULT now () ); -- Table UUID v7 (généré côté application, voir plus bas) CREATE TABLE events_v7 ( id UUID PRIMARY KEY , payload JSONB , created_at TIMESTAMPTZ DEFAULT now () ); Après insertion, voici les mesures : Type de clé Taille de l’index Fragmentation Latence moyenne d’insertion BIGINT ~214 Mo 0 % ~0,8 ms/ligne UUID v4 ~428 Mo (2×) 99 % ~4,8 ms/ligne UUID v7 ~428 Mo (2×) ~2 % ~1,1 ms/ligne Ce qui frappe, c’est la fragmentation quasi nulle de l’UUID v7. L’index reste compact et les insertions sont presque aussi rapides qu’avec un BIGSERIAL. L’UUID v4, lui, est plus de quatre fois plus lent à l’insertion sur ce volume. UUID v7 :

Over the last few weeks I built a website audit tool and ran it on 200+ small business and service websites — dental practices, plumbing companies, landscapers, law firms, real estate agents. Not Fortune 500 pages optimized by dedicated teams. The sites that actually serve local customers. I expected some issues. I did not expect this. Here's the raw data, the patterns I found, and what you can actually do about it. The Scorecard I grade sites across five dimensions on a 0–100 scale. Here are the averages from 200+ audits: Dimension Average Score Worst Score Speed 56 11 SEO 68 29 Mobile usability 61 18 Accessibility 52 8 Security 70 17 SEO and security tend to be passable (automated checks from Google Search Console and automatic SSL help). Speed and accessibility are consistently neglected — probably because the feedback loop is invisible. A slow or inaccessible site loses visitors silently, and the owner never knows why. Finding #1: 67% of Sites Ship >50% Unused CSS This was the single most surprising data point by far. When a browser loads a page, it downloads every byte of CSS, parses it, builds style rules for every selector, and only then paints. If 60% of those rules never get applied (because they target a contact form behind a click, or a mobile menu at 768px+), the browser still processed them. Worst case: a dental practice shipped 287 KB of CSS. Only 31 KB was used on first paint. That's 256 KB of unnecessary render-blocking weight that delayed First Contentful Paint by roughly 1.4 seconds. The fix: If you're using Tailwind, make sure tree-shaking is enabled. If you're writing vanilla CSS, open DevTools > Coverage tab > Reload. Anything over 40% unused is worth addressing. Most bundlers handle this — you just need to turn it on. Finding #2: Average Image Payload Is 1.8 MB — Way Too High Average image payload across all scanned sites: 1.8 MB per page. Only 34% serve WebP or AVIF (modern formats that cut file size by 30-50%). Only 28% serve responsive sizes

The Problem We Were Actually Solving I still remember the day our treasure hunt engine started to show its weaknesses. We had been using a custom-built solution written in Java, and it had served us well until our user base grew exponentially. The engine, which relied heavily on recursive searches and dynamic memory allocation, began to cause performance issues and occasional crashes. Our team was under pressure to find a solution that would allow our server to scale without sacrificing the user experience. After some research, I became convinced that Rust was the answer to our problems. Its focus on memory safety and performance seemed like the perfect fit for our needs. What We Tried First (And Why It Failed) Our first attempt at solving the problem was to simply translate our Java code into Rust. We thought that the language's built-in features would automatically solve our performance and memory issues. However, we quickly realized that this approach was not going to work. The Rust compiler was complaining about lifetime issues and borrow checker errors, which we did not fully understand at the time. We spent weeks trying to fix these issues, but our code was still not stable. I recall one particularly frustrating error message from the Rust compiler: error: cannot borrow self.list as mutable because it is also borrowed as immutable. It was then that I realized we needed to take a step back and rethink our approach. The Architecture Decision We decided to start from scratch and redesign our treasure hunt engine with Rust's strengths in mind. We chose to use a graph-based data structure, which allowed us to take advantage of Rust's ownership model and avoid common pitfalls like null pointer dereferences. We also made use of the crossbeam crate for parallelism and the tokio crate for async I/O. This new design required us to think differently about our problem domain, but it ultimately led to a more efficient and scalable solution. I was impressed by the level of