How a 10-Person Team Built an IoT Platform Serving 30 Million Connections

·1 views

DGIOT, an open-source industrial IoT platform, was developed by a small team of 10 engineers to handle up to 30 million concurrent connections. The project began in 2021 after China's largest oil producer, Daqing Oil Field, approached the team with a complex data infrastructure challenge involving 928 gateways and over 114,000 sensor points across 16 oil fields. The existing setup suffered from high latency and incompatible protocols, prompting the need for a scalable, unified solution. DGIOT's architecture now processes 652 million data points while maintaining 99.9999% uptime. The platform has since grown to a community of 70,000 developers who have contributed integrations beyond the original team's scope.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Why Writing Formal Specs Before Code Catches Bugs AI and Humans Miss

A software engineering article argues that developers should document every system guarantee in a formal specification before writing a single line of code. Using TLA+, a decades-old formal specification language, the author demonstrates how a simple to-do list app's rules — such as valid task state transitions — can be precisely defined in just a few lines. The approach gained credibility when AWS engineers reported in a CACM paper that writing TLA+ specs uncovered bugs in already-reviewed, production-bound systems. The author warns this practice is especially urgent now that AI agents generate plausible-looking code that can pass human review while violating core system rules. A lightweight Rust-based toolchain called tlatools is presented as a practical way to automatically check whether generated code conforms to a written specification.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Alan Babychan Launches Feature-Rich Portfolio with Dual Themes and Audio

Developer Alan Babychan has launched a major update to his personal portfolio at alanbabychan.online, treating it as a full product rather than a static showcase. The portfolio features two distinct visual themes — a Matrix theme and a Paper theme — along with an interactive audio system, custom cursor, and micro-interactions. Accessibility, user experience, and performance were core priorities, with Google Analytics 4 integrated for tracking. Babychan is also working on a personal AI assistant that will let visitors explore his projects and skills through conversational interaction. He is actively seeking feedback from fellow developers on both the design and functionality.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

npm, Yarn, pnpm, or Bun: Which Package Manager Suits Modern Web Dev Best?

A developer comparison of the four major JavaScript package managers — npm, Yarn, pnpm, and Bun — highlights key differences in speed, disk efficiency, and ecosystem support. pnpm emerges as the preferred choice for its content-addressable storage system, strict dependency management, and strong balance of performance and reliability. Bun is noted as a promising all-in-one tooling solution with impressive install speeds, but its ecosystem maturity and production readiness remain considerations. npm is recommended for beginners and compatibility-critical projects, while Yarn remains a solid option for teams with established monorepo workflows. The analysis concludes that raw speed alone should not drive the decision, with overall reliability, maintainability, and developer productivity being equally important factors.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How the Kernel Trick Solves Hard ML Problems by Changing the Space

The kernel trick, central to Support Vector Machines (SVMs), avoids brute-force computation by transforming a problem into a space where a simple linear boundary can separate data. Crucially, the data never actually moves into that high-dimensional space — kernel functions compute the required inner products directly from original coordinates. SVMs determine their decision boundary using only the closest data points, called support vectors, ignoring all other training examples. Unlike ensemble methods that average across all data, SVMs are defined entirely by these edge cases. A key practical warning: feature scaling in SVMs is not optional hygiene but a correctness requirement, since unscaled inputs silently distort the margin calculation without any obvious error.

0 comments Read more at DEV Community