标签：#process

Netflix migrated most of its batch workloads onto Kueue, an open-source cloud-native batch job execution system that has outgrown its homegrown solution over the years. The company mapped the capabilities previously created in-house to Kueue’s functionality and also benefited from new features that would have been costly to incorporate into its homegrown solution. By Rafał Gancarz

📺 Prefer to watch? 90-second YouTube Short · 💬 Telegram Originally published on software-engineer-blog.com . You run code concurrently all the time. But "concurrent" hides a critical choice: are you spawning separate processes or threads inside the same process? That choice decides whether one crash takes down your entire system or stays contained, and whether you're copying data between isolated worlds or racing to read the same memory. Mental model: A process is its own house; threads are roommates sharing one. Processes: Isolation at the Cost of Weight When you start a process, the operating system hands it its own private address space. That address space is walled off. Your process can't touch another process's memory—the OS enforces it at the CPU level. If your process crashes, it corrupts only its own memory. The kernel cleans it up. Every other process keeps running untouched. This is why browsers put each tab in its own process. One tab runs malicious JavaScript, spins into an infinite loop, or has a memory leak—that tab's process dies. The rest of your browser lives. You close the dead tab and open a new one. Your other tabs don't even hiccup. But isolation isn't free. Each process carries: Its own copy of the heap, stack, and memory pages Its own file descriptor table, open sockets, and kernel resources OS overhead to track and protect it Spawning a process is expensive—milliseconds on modern hardware, but measurably heavier than a thread. And if two processes need to share data, they can't just read the same memory. One process must copy data into a pipe or socket, send it across, and the other process must copy it out and into its own memory. That's overhead on every exchange. Threads: Speed and Sharing, With a Trap Threads live inside a single process and share that process's entire memory. The kernel doesn't wall them off from each other. When you spawn a thread, you're not duplicating the heap, the file descriptors, or the kernel state—you're just cr

`California's AB 723 has been in force since January 1, 2026. It amends Business & Professions Code § 10140.8 and it applies to any real estate listing image that has been digitally altered. Virtual staging is the obvious case, but the definition is wider than that. The rule has two parts: A statement that the image has been altered, "reasonably conspicuous" and placed on or adjacent to the image. A link to a publicly accessible URL, or a QR code, that includes and clearly identifies the original, unaltered image. Everyone builds the first part. It is a text label on a photo, an afternoon of work. The second part is a small piece of infrastructure: a permanent public URL, per image, that outlives the tab the agent had open when they exported. I build a virtual staging product, so I had to ship both. This is how the second part is put together, and the one thing I got wrong. What counts as altered Worth getting right before writing any code, because it decides which of your features need the label and which do not. Subsection (b)(2) carves out ordinary photo editing. Covered: Adding furniture, rugs, art or decor Removing furniture, clutter or personal items Changing paint, flooring or wall finishes Sky replacement and day to dusk Greening or reshaping lawns and landscaping Anything that changes the facade or the property itself Not covered: Exposure, lighting, white balance, color correction Sharpening Straightening, cropping, angle In the codebase that line is a set, and the two omissions are deliberate: ts export const TOOLS_ALTERING_LISTING_IMAGES = new Set([ "virtual-staging", "sky-replacement", "day-to-dusk", "grass-greener", "declutter", "object-remover", ]); image-enhancer is out because exposure and white balance are precisely what the statute excludes. A floor plan generator is out because a diagram is not an altered photograph. Attaching a legal claim to a feature the law does not cover is not a harmless extra: it is the fastest way to make the rest of your

ImgProxy is an open-source image-processing server — resize, convert, and transform images on the fly via URL parameters, ideal as a caching layer in front of a CDN or web app. This guide builds ImgProxy from source on Ubuntu, runs it as a systemd service behind Nginx with TLS, walks through its URL processing options, and secures it with signed URLs. Prerequisites: an Ubuntu server, a domain A record (e.g. imgproxy.example.com ), non-root sudo user. Install ImgProxy ImgProxy uses libvips for image processing; this builds it from source with Go. $ sudo add-apt-repository ppa:dhor/myway $ sudo apt update $ sudo apt install libvips-dev -y $ sudo snap install --classic --channel = latest/stable go $ git clone https://github.com/imgproxy/imgproxy.git $ cd imgproxy $ sudo CGO_LDFLAGS_ALLOW = "-s|-w" go build -o /usr/local/bin/imgproxy Create the environment config: $ sudo touch /usr/local/bin/imgproxy.env $ sudo nano /usr/local/bin/imgproxy.env IMGPROXY_BIND = :8080 IMGPROXY_NETWORK = tcp IMGPROXY_READ_TIMEOUT = 10 IMGPROXY_WRITE_TIMEOUT = 10 IMGPROXY_WORKERS = 2 IMGPROXY_REQUESTS_QUEUE_SIZE = 0 IMGPROXY_QUALITY = 100 IMGPROXY_PREFERRED_FORMATS = webp,jpeg,png,gif,avif IMGPROXY_LOG_FORMAT = "pretty" IMGPROXY_LOG_LEVEL = "INFO" IMGPROXY_WATERMARK_URL = https://example.com/watermark.png IMGPROXY_WATERMARK_OPACITY = 1 Key settings: IMGPROXY_WORKERS should be ~2× your vCPU count; IMGPROXY_REQUESTS_QUEUE_SIZE=0 means unlimited queueing; IMGPROXY_WATERMARK_URL points at whatever image you want overlaid when watermarking is enabled. Point ImgProxy at the config and test: $ export IMGPROXY_ENV_LOCAL_FILE_PATH = /usr/local/bin/imgproxy.env $ cd $ imgproxy WARNING [2024-05-28T00:40:42Z] No keys defined, so signature checking is disabled WARNING [2024-05-28T00:40:42Z] No salts defined, so signature checking is disabled INFO [2024-05-28T00:40:42Z] Starting server at :8080 Stop it with Ctrl+C once verified, then set it up as a service. Run ImgProxy as a systemd Service $ sudo useradd

Async processing qua message queue: vì sao đẩy việc nặng ra khỏi request path, và cái giá phải trả bằng eventual consistency Async processing là mô hình tách một request thành hai giai đoạn: request handler nhận việc, xác nhận với client, rồi giao phần xử lý thật cho một worker chạy ngoài request path — thường qua một message queue (RabbitMQ, AWS SQS, Kafka, Redis Streams, hoặc queue trên nền Redis như BullMQ/Sidekiq). Lý do dev gặp nó trong việc thật rất cụ thể: một endpoint gọi payment provider mất 3s, gửi email confirm mất 1s, resize ảnh mất 5s — nếu làm tuần tự trong request, p99 latency của endpoint là tổng các con số đó, và một downstream chậm hoặc chết đủ để làm timeout hết thread pool của app server. Đẩy vào queue thì request trả về trong vài chục ms; nhưng đổi lại, cái "xong" mà client thấy không còn nghĩa là việc đã thực sự hoàn thành. Cơ chế hoạt động Ba thành phần: producer (thường là API server) đóng gói việc thành message rồi publish vào broker; broker (RabbitMQ/SQS/Kafka…) giữ message trong queue có persistence tuỳ cấu hình; consumer/worker poll hoặc được push message, xử lý, rồi ack để broker biết xoá. Nếu worker chết trước khi ack, broker redeliver — đây là gốc của semantic at-least-once : mỗi message được giao ít nhất một lần, có thể nhiều lần. Exactly-once trong hệ phân tán chỉ đạt được ở lớp application bằng cách consumer viết idempotent, không phải bằng cấu hình broker. Ví dụ với RabbitMQ + Node ( amqplib ): // producer — trong HTTP handler const ch = await conn . createConfirmChannel () await ch . assertQueue ( ' image.resize ' , { durable : true }) app . post ( ' /upload ' , async ( req , res ) => { const jobId = crypto . randomUUID () const payload = Buffer . from ( JSON . stringify ({ jobId , s3Key : req . body . key })) await ch . sendToQueue ( ' image.resize ' , payload , { persistent : true , // ghi xuống disk, sống sót broker restart messageId : jobId , // để consumer dedupe contentType : ' application/json ' , }) // đợi broker confirm đ

The AI Implementation Process I Use With Every Client Most AI projects do not fail at the model. They fail in the six weeks before anyone writes a prompt, and in the six weeks after the demo lands in a Slack channel and nobody knows who owns it. I have run enough of these now (from one-off automations to multi-agent content systems running unattended) that the process has converged into something stable. This is the version I actually use. It has five phases: scoping, POC, integration, evaluation, operations. Each phase has an exit criterion. If we cannot meet the exit criterion, we do not move forward. That single rule has saved more projects than any clever architecture choice. Phase 1: Scoping (1 to 2 weeks, fixed price) Scoping ends with a written document that names the workflow being automated, the system of record it touches, the success metric in hours or dollars, the data we have access to, and the smallest possible first slice. No model is chosen yet. No code is written. If we cannot produce that document, the engagement stops here and the client keeps the document. The hardest part of scoping is resisting the urge to solve the interesting problem. Clients almost always describe the AI-shaped fantasy ("an agent that handles all support tickets") when the real opportunity is narrower and uglier ("triage tier-1 tickets that mention billing, route to the right queue, draft a reply for human approval"). The narrower version ships. The fantasy does not. I run scoping as three sessions: Workflow walkthrough. Someone who actually does the work shows me their screen for an hour. I record it. I take timestamps. The point is to find the moments where a human is doing pattern matching that an LLM can do, and to find the moments where they are doing judgment that an LLM should not do. Data audit. Where does the input live? Where does the output need to go? What is the auth story? If the data is locked inside a SaaS product with no API and no export, that is the projec

Atlassian details the Forge billing platform built for usage-based pricing across its cloud ecosystem. It processes large-scale usage events with correct attribution, deduplication, and aggregation using a streaming pipeline, idempotent processing, and layered storage to enable accurate billing, near real-time visibility, and reliable reconciliation across distributed services. By Leela Kumili

AMD's stripping of TSME from consumer CPUs appears to be a deliberate, covert move.

Shareable blog post edition: https://andymaleh.blogspot.com/2026/06/andys-laws-of-ai-in-software-engineering.html Law #1: "The more Software Developers use AI, the more valuable Software Engineers who do not use AI become." Software Engineers who are masters at delivering Software without using AI will actually have increased job security the more Software Developers in the worldwide Software Development community rely on AI to deliver Software without having true mastery over Software Engineering. As more Software Developers become fully dependent on AI to build Software without truly understanding how AI gets work done, Software Engineers who do understand what is going on under the hood will dwindle and become more valuable than ever. In other words, they will have a competitive advantage over Software Developers who can only deliver Software features with AI as well as Software Developers who have not mastered Software Engineering. Also, there will always be a need for Software Engineers who can maintain the Software of AI itself. Law #2: "Software Developers benefit from AI in direct proportion to how weak they are in Software Engineering" The weaker Software Developers are at Software Engineering the more they benefit from AI. After all, AI learns from Master Software Engineers and then applies its learnings in code generation done for lower-level Software Developers who lack mastery in Software Engineering. So, users of AI simply place themselves lower in the expertise hierarchy to be on the receiving end of what Master Software Engineers feed AI with their code. This explains why many experts like Linus Torvalds do not find AI very useful while devs who have zero degrees and qualifications feel like they get a lot from AI. A beneficial thing to learn from this law is that it is more valuable for a Software Developer to hone in their Software Engineering skills (including the completion of university degrees) than to hone in their AI usage skills because if t

There is a version of organizational life that feels, from inside, like maturity. Meetings reach conclusions. Decisions move forward without extended debate. The leadership team operates with visible coherence. Escalations are rare. When concerns are raised, they are quickly absorbed into the existing framework and resolved without disruption. Everything functions exactly as designed. That is what makes it difficult to notice that something essential has stopped occurring. The more stable the system appears, the more completely it has eliminated the conditions under which instability would be visible. The Epistemic Function of Conflict Conflict in organizations is not primarily a social problem. It is an epistemic mechanism. When a decision is challenged, something precise occurs: the decision's internal logic is forced into the open. Its assumptions are made explicit. Its evidence is tested against contrary interpretation. The challenge does not guarantee a better outcome — but it generates information the unchallenged decision never produces. Conflict is not disruption of the system. It is how the system verifies itself against reality. Remove the disagreement, and the system continues deciding. It simply stops testing whether its decisions are sound. The absence of challenge feels like confidence. It is blindness — a blindness that is, from inside, indistinguishable from clarity. What Fills the Vacuum When legitimate conflict disappears, the space does not remain empty. It fills with the performance of conflict. Meetings still contain discussion. Questions are still asked. Concerns are occasionally raised. But the texture has changed in ways that experienced practitioners feel before they can articulate. Questions are asked to signal engagement rather than to probe assumptions. Concerns are framed to demonstrate awareness rather than to force resolution. Debate occurs within the boundaries of what the system has already decided is acceptable to debate. The ritual

Disclosure: I work on Passlens, a browser-first passport and ID photo maker. This post is about the product decisions behind that workflow, not a neutral review of every tool in the space. A passport photo looks simple until you try to make one that an upload form will actually accept. It is a headshot, yes, but it is also a small chain of constraints: physical size, pixel size, background, head position, print scale, and whatever the destination country's portal decides to reject that week. That is why generic photo editors feel slightly wrong for this job. They can crop. They can resize. They can export. The hard part is not any one of those actions. The hard part is keeping all of them tied to the document rule the user picked. The unit problem For developers, document photos are awkward because two units matter at the same time. A user may need a 2x2 inch passport photo. A visa portal may ask for 600x600 pixels. A print sheet may need 35x45 mm photos at 300 DPI. These are not the same request, but people often treat them as if they are. If the app only thinks in pixels, the print can come out the wrong physical size. If it only thinks in millimetres or inches, the digital upload can be rejected for the wrong pixel dimensions. A good workflow has to keep both ideas alive: the document size and the export target. That is the main reason Passlens keeps presets and print layouts as first-class pieces of the workflow instead of treating them as labels on a crop box. The crop is not the output Another small trap: the crop the user sees is not always the final output. For a digital upload, the crop usually becomes one image file. For printing, the same crop may become several photos arranged on 4x6, A4, or Letter paper with spacing, margins, and optional cut marks. If that print sheet is scaled by the browser or printer dialog, the whole thing is wrong. So the editor needs to separate three things: the face and shoulder crop the finished document-photo size the print s