标签：#Python

GitHub热门项目 | The 30 Days of Python programming challenge is a step-by-step guide to learn the Python programming language in 30 days. This challenge may take more than 100 days. Follow your own pace. These videos may help too: https://www.youtube.com/channel/UC7PNRuno1rzYPb1xLa4yktw | Stars: 65,108 | 150 stars today | 语言: Python

There is a persistent myth that to build a worthy code assistant, you absolutely must use GPT or Claude. This is false. You don't need a 1-trillion parameter model. You need a small local model and extremely rigorous engineering around it. This is the direction history is taking for companies. As Mark Zuckerberg mentioned, the future isn't a single omniscient model, but "every company having its own specialized AI" . And this specialization necessarily involves fine-tuning and local deployment (or on sovereign servers) to guarantee data security. The thesis behind the construction of Vibrisse Agent can be summed up in one sentence: Small models, Great tools. In this article, I will detail the technical stack and concrete engineering solutions I implemented to tame a local model and make it reliable in production: LangGraph, Ollama, FastAPI, React (no build step, with embedded custom CSS) , all running on a machine with 32 GB of RAM. For the curious who want to run the agent on their machine right now: // MacOs / Linux curl -sSL https://agent.vibrisse-studio.dev/install.sh | bash // Windows irm https://agent.vibrisse-studio.dev/install.ps1 | iex Architecture: Why a State Machine (LangGraph)? At first, when building an LLM application, we tend to think in sequential chains: Input -> Prompt -> Tool -> Output . The problem is that if one node fails, the whole chain stops without us being able to catch the error or understand the context of the crash. That's where LangGraph comes in. Vibrisse's architecture isn't a chain, it's a state machine . Every node in the graph has a very precise responsibility, shares a global conversation state, and uses conditional transitions to move to the next node. I implemented the Supervisor / Worker pattern: The Supervisor analyzes the user's intent. It does nothing else but route. It dispatches the task to specialized Workers (the RAG Worker, the Search Worker, the Ghost Worker...). If a Worker fails or needs more information, it can se

Il y a un mythe persistant selon lequel pour construire un assistant de code digne de ce nom, il faut absolument utiliser GPT ou Claude. C'est faux. Vous n'avez pas besoin d'un modèle à 1 trillion de paramètres. Vous avez besoin d'un modèle local de taille réduite et d'une ingénierie extrêmement rigoureuse autour de lui. C'est d'ailleurs le sens de l'histoire pour les entreprises. Comme l'évoquait Mark Zuckerberg, l'avenir n'est pas à un modèle omniscient unique, mais à "chaque entreprise avec sa propre IA spécialisée" . Et cette spécialisation passe obligatoirement par le fine-tuning et le déploiement local (ou sur serveurs souverains) pour garantir la sécurité des données. La thèse derrière la construction de Vibrisse Agent tient en une phrase : Small models, Great tools. Dans cet article, je vais détailler la stack technique et les solutions d'ingénierie concrètes que j'ai mises en place pour dompter un modèle local et le rendre fiable en production : LangGraph, Ollama, FastAPI, React (sans build step, avec CSS custom embarqué) , le tout tournant sur une machine avec 32 Go de RAM. Pour les curieux qui souhaitent lancer l'agent sur leur machine dès maintenant : // MacOs / Linux curl -sSL https://agent.vibrisse-studio.dev/install.sh | bash // Windows irm https://agent.vibrisse-studio.dev/install.ps1 | iex L'Architecture : Pourquoi une Machine à États (LangGraph) ? Au début, quand on construit une application LLM, on a tendance à penser en chaîne séquentielle : Input -> Prompt -> Outil -> Output . Le problème, c'est que si un nœud échoue, toute la chaîne s'arrête sans qu'on puisse rattraper l'erreur ou comprendre le contexte du plantage. C'est là qu'intervient LangGraph . L'architecture de Vibrisse n'est pas une chaîne, c'est une machine à états . Chaque nœud du graphe a une responsabilité très précise, partage un état global de la conversation, et utilise des transitions conditionnelles pour passer au nœud suivant. J'ai implémenté le pattern Supervisor / Worker : L

Every Indian data scientist hits the same wall. You need district-level population data. You go to censusindia.gov.in. You find hundreds of inconsistent Excel files with merged headers, footnote rows, and zero documentation. You spend a full day just loading the data before doing any actual analysis. I fixed that. Once. For everyone. What I built indiaset/census-2011 India's Census 2011 district data, clean, typed, and ready for pandas. 640 districts · 29 columns · 0 missing values Validated against official India total · LGD codes attached Load it in 4 lines from huggingface_hub import hf_hub_download import pandas as pd path = hf_hub_download ( repo_id = " indiaset/census-2011 " , filename = " census_2011_districts_final.parquet " , repo_type = " dataset " ) df = pd . read_parquet ( path ) print ( df . shape ) # (640, 29) What's in it Column Description state_code Census 2011 state code state_name Official state/UT name district_code Census 2011 district code district_name District name as per Census lgd_code LGD permanent district code district_name_lgd District name as per LGD pop_total Total population pop_male Male population pop_female Female population pop_under6_total Children under 6 years pop_sc Scheduled Caste population pop_st Scheduled Tribe population literate_total Literate persons literate_male Literate males literate_female Literate females illiterate_total Illiterate persons workers_total Total workers workers_male Male workers workers_female Female workers non_workers_total Non workers literacy_rate Literate / Total × 100 sex_ratio Females per 1000 males workforce_participation Workers / Total × 100 The validation The most important test - do all 640 district populations sum to India's official total? print ( df [ ' pop_total ' ]. sum ()) # 1210854977 ✅ — exact match, zero discrepancy What the data actually shows Most literate district → Pathanamthitta, Kerala : 88.74% Least literate district → Alirajpur, Madhya Pradesh : 28.77% Literacy gap acro

So here's what happened: i Wish I Knew AI Recommendation Sooner — Here's the Full Breakdown Last quarter I burned through about three billable hours debugging a recommendation pipeline for a Shopify client. The thing was — it shouldn't have taken that long. I had the data. I had the API keys. What I didn't have was a clear-eyed picture of what AI recommendation systems actually cost in 2026 when you're paying the bills yourself. If you freelance like I do, every line item matters. My "office" is a kitchen table, my "PM" is a Slack ping at 11pm, and my CFO is whatever's left in my checking account after software subscriptions. So when I say I've been digging into the numbers on AI recommendation systems for the last six weeks, I mean I've been doing it the way I do everything: with a calculator open in one tab and a client invoice in the other. This post is the writeup I wish I'd had before I started. Consider it the field guide for anyone building recommendation features on a budget, on a deadline, or just for fun. Why I Even Cared About Recommendation Systems I took on a small retainer back in February for an indie e-commerce shop that sells specialty coffee beans. They wanted "AI-powered product recommendations" on their storefront — you know, the classic "customers who bought this also bought..." thing, but smarter. The owner had been quoted $15,000 by a "full-service AI agency" to build it. He doesn't have $15,000. He has $15,000 in revenue per month and a wife who is deeply skeptical of his side-hustle energy. So he came to me. And I said yes, because I'm a sucker and also because I knew it should cost a tiny fraction of that quote. The math was never going to support five figures for a recommendation widget. Not when the underlying API calls are fractions of a cent. That's when I started really paying attention to the pricing landscape. The 184-Model Elephant in the Room Here's the thing nobody tells you when you start shopping for LLMs: there are a lot of the

I'm not here to trash Sololearn. Sololearn taught millions of people how to code. It was one of the first apps to make programming education feel mobile-native. That's a real achievement. I respect it. But I'm building Codino — a Python learning app — and I'd be lying if I said I didn't study Sololearn carefully before writing a single line of code. I looked at what they got right. I looked at where users complained. And I made decisions based on both. This is that honest breakdown. What Sololearn Got Right 1. The Community Feel Sololearn built a genuine community. The code playground where users share their projects, comment on each other's code, and get likes — that was smart. Learning feels less lonely when other people are doing it alongside you. It created a social loop that kept people coming back even when they weren't actively doing lessons. I haven't built this yet in Codino. The leaderboard is a start, but a full community layer is something I'm thinking about for a future update. 2. Multi-Language Support Sololearn didn't bet on just one language. Python, JavaScript, C++, SQL, HTML — they covered everything. That gave them a massive addressable audience. Codino is Python-only right now. That's intentional — going deep on one language is better than going shallow on ten. But I understand why multi-language eventually matters for scale. 3. The Code Playground The ability to write and run real code inside the app — without going to a browser — was ahead of its time when Sololearn launched it. That feature alone brought back users who had finished all the lessons. Codino has a full offline IDE powered by Sora Editor. I'd argue ours is actually more capable — real syntax highlighting, autocompletion, offline Python execution — but Sololearn deserves credit for proving this feature matters. 4. Bite-Sized Lessons That Actually Work Sololearn understood that people learn on the bus, in bed, waiting in line. Their lessons are short, digestible, and don't demand 45

Last year I was on a team that pushed 40 million events per day through Kafka. We had consumer lag alerts, rebalancing incidents, and a whole runbook for when the broker got behind. I understood how to operate Kafka. But I did not understand how Kafka works. So I built a tiny one. No dependencies. No Zookeeper. No JVM. Just Python and the core ideas. Here is what I learned. The Three Things Kafka Actually Does People say "Kafka is a message queue." That is not quite right. Kafka is a distributed commit log . It has three jobs: Accept writes from producers and append them to a log Let consumers read from any offset in that log Remember where each consumer group is up to That third one is the thing that makes Kafka different from a traditional queue. A queue forgets a message once it is consumed. Kafka remembers. You can replay. You can have 10 different consumer groups reading the same topic at different speeds. The code to implement this is smaller than you think. brokelite: A Message Broker in 120 Lines import threading import time from collections import defaultdict from typing import Dict , List , Tuple class Partition : """ Append-only log for one partition of a topic. """ def __init__ ( self ): self . _log : List [ Tuple [ int , bytes ]] = [] # (offset, message) self . _lock = threading . Lock () self . _next_offset = 0 def append ( self , message : bytes ) -> int : with self . _lock : offset = self . _next_offset self . _log . append (( offset , message )) self . _next_offset += 1 return offset def read_from ( self , offset : int , max_count : int = 100 ) -> List [ Tuple [ int , bytes ]]: with self . _lock : return [ ( off , msg ) for off , msg in self . _log if off >= offset ][: max_count ] def __len__ ( self ): return self . _next_offset class Topic : """ A topic is just N partitions. """ def __init__ ( self , name : str , num_partitions : int = 3 ): self . name = name self . partitions = [ Partition () for _ in range ( num_partitions )] def route ( self , k

Local SEO rankings are not the same everywhere. Your website may rank in position 2 in Austin, position 8 in Dallas, and not appear at all in Chicago. That is why checking one generic Google result is not enough for local SEO. If you are working on local SEO , agency reporting, competitor monitoring, or location-based search analysis, you need to track rankings by city. A simple workflow looks like this: Keyword + city → SERP API → local search results → ranking check → CSV report In this tutorial, we’ll build a basic Python script that tracks local SEO rankings by city using a SERP API. We will: Define target keywords Define target cities Send city-specific searches to a SERP API Extract organic results Check where a target domain appears Save the ranking data to CSV This is not a full SEO platform, but it gives you the core logic behind many local rank tracking tools. Why city-level rankings matter Google search results are location-sensitive. A query like: best digital marketing agency may return different results in: New York Austin London Singapore Sydney This matters even more for local intent keywords, such as: dentist near me plumber in Chicago coffee shop in Austin real estate agent in Miami For these searches, the result page can include: organic results local packs Google Maps results ads business directories review sites service pages location-specific landing pages If you only check one location, you may miss what users actually see in other cities. For local SEO, ranking data without location context is incomplete. Why use a SERP API? You could try to check Google rankings manually. But that does not scale. You could also try to scrape Google directly, but that brings a lot of maintenance work: changing page layouts CAPTCHA blocked requests proxy handling inconsistent HTML location mismatch parser updates retry logic A SERP API gives you structured search results in JSON. Instead of parsing raw HTML, you get data like this: { "query" : "plumber in Aust

I open-sourced MarketEye today. For anyone who missed the first post: MarketEye is a self-hosted competitor price monitor I built because I didn't want to pay $99/month for Prisync. The code is now up on GitHub under MIT license. GitHub: github.com/dachengzi065-gif/marketeye Why open source? Three reasons: 1. People actually asked for it. After my first post here, a few people DM'd me asking to see the code. They're developers too — they want to modify it, extend it, make it their own. That's fair. Selling source code to devs is like selling ice to eskimos. 2. Trust. A closed-source price tracker that "runs on your machine" — you either trust the author or you don't. Open source removes that doubt. You can read every line, check what data leaves your machine (nothing), and build it yourself if you want. 3. Longevity. Self-hosted tools have a dirty secret: if the developer disappears, you're stuck with a broken tool. Open source changes that. Even if I get hit by a bus tomorrow, you can fork the repo and keep going. What this means for the $49 version The Gumroad package still exists. It includes: The same code, pre-packaged Email support (I'll help you set it up) A clear conscience subscription (you're paying for convenience, not software) But honestly? If you can run pip install , just clone the repo. It's free. What's next I'm actively working on: Docker image (one-command deploy) More scrapers (plugins for different sites) Discord/Telegram bot alerts (requested by several people) PRs welcome. Issues welcome. Feedback welcome. 👉 github.com/dachengzi065-gif/marketeye

Most AI demos work perfectly on a laptop. But production AI systems can become fragile when everything is handled inside one synchronous API call. A user sends a request. The API extracts text. The API chunks the content. The API generates embeddings. The API stores data. The API waits for everything to finish. This may look simple in a demo, but it quickly becomes a problem in real systems. The problem with one giant API call In many AI applications, the API is expected to do too much. For example, in a document processing or RAG pipeline, one request may trigger multiple heavy steps: text extraction chunking embedding generation indexing summarization database updates If all of this happens inside one synchronous request, the API becomes slow and fragile. If one downstream step fails, the complete request may fail. If traffic increases suddenly, the API may become overloaded. This is why event-driven architecture becomes useful for AI workloads. A better approach: API + Kafka + workers Instead of making the API do everything, we can split the workflow into smaller services. The API accepts the request and publishes an event. Background workers consume events and continue the processing asynchronously. A simple flow looks like this: User Request ↓ FastAPI ↓ Kafka / Redpanda Topic ↓ Python Worker ↓ Next Processing Stage In my practical demo, I am using: FastAPI Redpanda Python workers Docker Compose Kafka-compatible messaging Why Redpanda? Redpanda is Kafka-compatible, which makes it useful for local demos and event-driven architecture experiments. It allows us to work with Kafka-style topics, producers, and consumers while keeping the setup simple for development. What this architecture gives us This approach helps with: decoupling services handling bursty workloads moving long-running tasks to background workers improving scalability isolating failures building production-style AI pipelines This pattern is especially useful for AI systems involving: document proce

Intro If you work with real-time financial tick data via WebSocket long connections, you’ve probably run into frustrating hidden issues: Connections show as online, but data stops flowing. Adding/removing trading symbols triggers connection storms. Minor network glitches break your data pipeline without obvious error logs. Today I’ll share a practical Python solution built around single-connection dynamic subscription and heartbeat timeout detection . We use alltick api as our example throughout this post. I’ll cover real production pain points, architecture, full runnable code, common bugs and optimization results. All code can be directly used in your projects. Real Production Pain Points & Background My team maintains real-time data pipelines for stocks, forex and cryptocurrencies. A core requirement is to add or remove monitored trading symbols on demand. At first, we created one independent WebSocket connection for each symbol. This simple approach caused a lot of problems in production: Bulk symbol updates create frequent connection creation and destruction, leading to reconnection storms and high network load. Ghost subscriptions: Data keeps coming even after you send an unsubscribe request. False-alive sockets: The connection status stays connected, but tick data stops completely. Your analysis logic runs on invalid data. We refactored the architecture to use one single persistent WebSocket connection for all symbols, plus an independent thread for heartbeat monitoring. After the upgrade, all hidden connection issues are gone, and the system runs stably under high-frequency data traffic. 4 Common WebSocket Issues in Fintech Let’s break down the most frequent problems when using WebSocket for financial streaming. 1. Connection Flood & Reconnection Storms Rebuilding connections every time you update symbols causes endless handshakes and closures. It wastes system resources and makes the entire data stream unstable. 2. Undetectable False-Alive Sockets Network j

If you've ever needed a list of local businesses - every dentist in Manchester, every plumber in Leeds - with their contact details , you've probably hit the same wall I did: Google's Places API is rate-limited, costs money once you scale, and annoyingly doesn't return email addresses at all. Copy-pasting from Maps by hand is soul-destroying past the first ten rows. Most "scrapers" give you the name and a phone number, then stop right where the value starts: the email . This guide shows a practical way to pull structured business data straight from Google Maps and auto-enrich each result with emails, extra phones, and social links - no Google API key, exportable to JSON/CSV/Excel, and callable from code. What you actually get per business { "name" : "Ringway Dental - Cheadle" , "address" : "187 Finney Ln, Heald Green, Cheadle SK8 3PX" , "phone" : "0161 437 2029" , "website" : "https://www.ringwaydental.com/" , "rating" : 5 , "reviewsCount" : 598 , "category" : "Dental clinic" , "lat" : 53.37 , "lng" : -2.22 , "emails" : [ "reception@ringwaydental.com" ], // ← enriched from the website "socialLinks" : { "facebook" : "..." , "instagram" : "..." }, "extraPhones" : [ "..." ] } The first block (name → coordinates) comes from Maps. The emails / socialLinks / extraPhones are the bit that makes a list actually usable for outreach - they're crawled from each business's own website. The fast way: a ready-made Actor Rather than build and babysit the scraping yourself, I packaged this as an Apify Actor: Google Maps Scraper . You give it search terms + locations; it returns enriched rows. Input: { "searchTerms" : [ "dentists" ], "locations" : [ "Manchester, UK" ], "maxPlacesPerSearch" : 50 , "scrapeContacts" : true , "relatedEmailsOnly" : true } That's it. scrapeContacts: true turns on the website crawl for emails/socials; relatedEmailsOnly keeps only emails that belong to the business's own domain (so you don't get random gmail noise). Call it from code (Python) Every Apify Act

When you double-click WP Maintenance Manager, it opens a browser tab — and the entire UI lives inside that tab. No native window is created. It's an unusual structure for a first-time user, and the natural question is: "why a browser?" That choice was an intentional design decision when building a Python desktop application. Here's the comparison that led to it, and the side effects of the choice. Four realistic options For a WordPress maintenance automation tool, four implementation styles were practical: Approach UI Distribution size Dev cost Per-OS extra work Native (Swift / WPF) OS-native windows Small–medium High (separate impl per OS) Heavy PyQt / PySide Qt widgets Medium (~80 MB) Medium Light Electron Chromium-embedded web UI Large (~150 MB+) Medium Light Local Flask + system browser System browser tab Small (~50 MB) Medium Light PyQt was a serious early candidate. A Python-only stack is appealing, but widget styling drifts subtly between OSes, Qt's layout system demands constant attention, and resolving Qt plugins under PyInstaller is fiddly. Dev velocity was not where it needed to be. Electron is the industry-standard choice for cross-platform UI, with the big benefit that HTML/CSS-based UIs are quick to write. But the distribution is well over 100 MB, and memory consumption is heavy. For a tool that often runs in the background, that overhead is too much to justify. Why local Flask + browser won The final structure was Flask (Python's lightweight web framework) + the system browser for UI. The decision rested on three axes: 1. The backend had to be Python anyway SSH connections via fabric / paramiko , browser automation via playwright , encryption via cryptography — every library at the core of WordPress maintenance lives in the Python ecosystem. Writing the backend in another language wasn't really an option. If Python is already required on the backend, putting the UI in Python too keeps distribution simple. 2. HTML/CSS/JS makes UI iteration fast Flask r

Meet Perri Comic Generator , a lightweight, single-panel comic creator that merges LLM-driven storytelling with real-time diffusion models. By pairing an Gradio frontend with a high-performance backend, Perri orchestrates a seamless pipeline: it takes a simple story seed, structures it into a panel description, generates the art, and burns the dialogue right onto the final image. The best part? It achieves all of this without massive, resource-heavy infrastructure. Every AI model under Perri's hood is under 32 billion parameters , proving that you don't need giant, compute-heavy models to build something amazing. Here is a look inside the architecture and tech stack that powers Perri. The Technical Architecture Perri is built using a clean separation of concerns, splitting the heavy lifting of generation away from the user interface. 1. The Frontend ( app.py ) Built using Gradio 6.16.0 , the frontend provides a sleek, user-friendly interface for inputting story seeds. To match the creative spirit of comics, the UI utilizes a custom theme, incorporating a vintage aesthetic complete with star-twinkle CSS overlays. The frontend's main jobs are: Capturing the user's initial prompt. Shipping the payload to the backend infrastructure via secure API requests. Decoding the backend's response—a Base64-encoded JPEG—and rendering it within the Gradio image component. 2. The Backend Orchestrator ( orchestrator.py ) The orchestrator acts as the brain of the operation, executing three distinct phases in the lifecycle of a single comic panel: Script Generation: It refines the user's raw prompt into a highly structured visual script and dialogue snippet using meta-llama/Meta-Llama-3-8B-Instruct . Image Generation: It passes the visual description to stabilityai/sdxl-turbo to synthesize the retro comic art. Dialogue Overlay Composition: Instead of relying on separate text captions, the orchestrator dynamically draws the generated dialogue directly onto the JPEG image, ensuring an au

A connected Gmail grant starts with an OAuth consent screen and ends with a refresh token you have to babysit; a Nylas Agent Account starts and ends with one POST request. Same API surface afterward — same messages endpoints, same webhooks, same calendar — but the provisioning story couldn't be more different, and that difference is what makes these hosted mailboxes such a natural fit for Python automation, agents, and test harnesses. Agent Accounts are in beta, and the official quickstart gets you from nothing to a sending-and-receiving mailbox in under 5 minutes using curl. Here's the whole flow as a Python script. Step 0: prerequisites You need an API key (run nylas init with the CLI, or use the Dashboard) and a domain. The fast path for testing: register a *.nylas.email trial subdomain from the Dashboard — no DNS records, instantly usable. Custom domains need MX and TXT records published at your DNS provider, with automatic verification once they propagate; save that for production. import os import requests BASE = " https://api.us.nylas.com " HEADERS = { " Authorization " : f " Bearer { os . environ [ ' NYLAS_API_KEY ' ] } " , " Content-Type " : " application/json " , } Step 1: provision the account POST /v3/connect/custom with "provider": "nylas" . No refresh token — just an email address on a registered domain: resp = requests . post ( f " { BASE } /v3/connect/custom " , headers = HEADERS , json = { " provider " : " nylas " , " settings " : { " email " : " test@your-application.nylas.email " }, }, ) resp . raise_for_status () grant_id = resp . json ()[ " data " ][ " id " ] print ( f " Agent Account live: { grant_id } " ) Save that grant_id — per the docs, you'll use it in every subsequent call. The mailbox works with every existing endpoint from this moment on. If you want policies or mail rules applied, add a top-level workspace_id to the same request body; the account inherits the workspace's limits, spam settings, and rules. Omit it and the account lands i

Notes from building a memory layer that forgets on purpose. Most "memory-enabled" agents don't remember anything. They re-read. Every turn, the whole conversation gets pasted back into the prompt, and we call that memory because the model can answer questions about earlier turns. It's a good trick. I used it for months. It also falls apart the moment real people start using the thing, and it falls apart in three separate ways. The first is the one everyone notices: it's expensive and noisy. You re-send every prior turn on every request. The single line you actually care about - "I'm allergic to peanuts" - is buried under a thousand lines of small talk, and you pay for all of it, every time. The second is quieter and worse. Transcript-stuffing has no idea what stale means. If someone told your agent "I'm vegetarian" in March and "I eat fish now" in May, you've just handed the model both facts with equal weight. Now it has to guess which one is current. Sometimes it guesses wrong, and there's nothing in the system that even thinks that's a problem. The third one is the reason I stopped treating this as a side quest. When you finally add summarization to control the cost from problem one, the summarizer is free to drop whatever it wants to save tokens. Including the allergy. I spent years around fintech, where the wrong record surviving (or the right one quietly vanishing) is how people get hurt, so this landed hard: forgetting an allergy to save 40 tokens isn't a cost bug. It's a safety bug wearing a cost bug's clothes. So the question I actually wanted to answer wasn't "how do I make my agent remember more." It was: how do I build something where acting on a fact the user already retracted and silently dropping a fact that must survive are impossible by construction, not just unlikely if the prompt is good that day. Everyone has already solved one third of this The encouraging part is that you don't have to invent much. The discouraging part is that every existing sy

Why I modeled a load balancer after Emperor Penguin huddles A few months ago I was reading about how emperor penguins survive Antarctic winters. Temperature drops to -40°C, wind hits 120km/h, and somehow these birds make it through. Not because they're individually tough. Because they rotate. Cold penguins on the outside push inward. Warm ones from the center move out to rest. Nobody coordinates this. No penguin is in charge. It emerges from one simple rule: if you're cold, push in. If you're warm, you'll get pushed out eventually. I couldn't stop thinking about this. I was working on a service mesh at the time and dealing with the usual problem — one slow server quietly dragging down the whole cluster. Round robin doesn't care. Least connections helps but not always. Weighted approaches need manual tuning that goes stale immediately. The penguin thing kept nagging at me. What if servers had a "temperature"? What if hot servers rotated out to rest? That's HuddleCluster. The basic structure Two rings: Inner ring (deque): Active servers. Requests go to them round-robin. Simple, fair, zero overhead for normal traffic. Outer ring (min-heap): Resting servers. Keyed by temperature — coolest server sits at the top, ready to rotate back in first. When a server in the inner ring runs hot past a threshold, it moves out. When an outer ring server cools down, it comes back in. That's the entire rotation logic. About 50 lines of Python. What is "temperature"? This took me a while to get right. My first attempt was just raw latency. That was bad. A server handling one slow database query looks terrible even when it's completely healthy. I needed something more composed. Current formula: pythontemperature = EMA( 0.7 * relative_latency_anomaly + 0.1 * cpu_score + 0.1 * memory_score + 0.1 * (error_rate + connection_score) ) Three decisions here worth explaining. EMA over simple moving average EMA weights recent measurements more heavily. If a server just had a bad spike but recovere

GitHub热门项目 | Archipelago Multi-Game Randomizer and Server | Stars: 1,460 | 3 stars today | 语言: Python

GitHub热门项目 | Local MTProto proxy server for partial bypassing of Telegram loading | Stars: 7,484 | 102 stars today | 语言: Python

Pipeline & Prompts | Byte size guides on DevOps, Cloud and AI AI in the Stack #2 ⚡ Byte Size Summary RAG inserts a retrieval layer between your existing runbooks and an LLM — answers come from your documentation, not generic training data, with source citations included. This article builds a complete FastAPI service with /ingest , /query , and /health endpoints, using OpenAI embeddings and Chroma as the vector store. Everything is cloneable from GitHub. The goal is not to replace your runbooks. It is to make them queryable at the moment an incident is happening. I have never met a platform team with bad runbooks. I have met plenty of platform teams where the runbooks exist, are reasonably well written, are stored somewhere sensible — and are still completely useless at 2am when something is on fire. Not because the content is wrong. Because nobody can find the right one fast enough. The search in Confluence returns fourteen results and none of them are titled the way the engineer is thinking about the problem. The person on call is junior and doesn't know the runbook exists. The runbook was written for a slightly different version of the service and nobody updated it. The runbook problem is not a writing problem. It is a retrieval problem. That is exactly the problem RAG was built to solve — and it is one of the highest-ROI first applications of AI in a platform engineering context. Not because it is technically impressive. Because it closes a gap that costs your team hours every month. This article builds a working pipeline. By the end you will have a FastAPI service that takes a natural language question — "why is my pod stuck in CrashLoopBackOff after a config change?" — and returns an answer grounded in your actual runbooks, with the source document cited. Everything is in the GitHub repo agentic-devops What RAG Is — Without the Hype RAG stands for Retrieval-Augmented Generation. Instead of asking an LLM a question and hoping its training data contains the answ