Git at any scale

Git at any scale
26 minutes by Vicent Martí

Git hosting is hard because Git was designed for distributed use, but most teams rely on a central server. Early approaches like distributed filesystems failed due to Git's random read patterns across packfiles. GitHub's Spokes system improved things with consistent replication using three-phase commit, but struggles with large monorepos and scale.

Automate your next bug bash with AI
sponsored by Momentic

Your coding agents are going faster, but testing is still a bottleneck. Mo explores your web, iOS, or Android app like a manual tester and reports every issue with a video recording and repro. For a Fortune 1000 file-collaboration platform, Mo caught 27 bugs, without writing or maintaining a single test script. Hire Mo for your next release.

Regressive JPEGs
4 minutes by Maurycyz

JPEG files can store image data across multiple scans, loading low-quality previews first and refining them over time. By stripping out certain markers and combining multiple images into one file, each scan can overwrite the previous image data, creating a crude animation. Browsers stop rendering after around 90 scans, so each frame must use a minimal single DC-only scan. The result is a working video packed into one image file, with playback speed controlled entirely by network delay.

How Cloudflare saved 100 terabytes of memory
12 minutes by Sebastiaan Neuteboom

Cloudflare's DNS platform stores over 250 billion cache entries, so even single-byte savings matter at scale. Five memory optimizations cut the per-entry footprint by 56%, freeing around 100 terabytes of RAM across the fleet. Changes included dropping unused capacity fields, storing record data as raw bytes in one contiguous buffer, and inferring repeated domain names instead of storing them. Beyond saving memory, insert throughput rose 43% and lookup latency fell 19%.

A quick look at zero-knowledge proofs
12 minutes by Max Bernstein

Zero-knowledge proofs let one party convince another that it has solved a hard problem, without revealing the solution itself. Graph 3-coloring is a classic example: the prover shuffles colors, locks them in hashed boxes, and lets the verifier check one random edge per round. Repeating this many times makes cheating statistically very unlikely. The same idea extends to other hard problems like Sudoku, or anything reducible to a known NP-complete problem.

Agentic code quality
9 minutes by Addy Osmani

AI agents can write massive amounts of code, making traditional human code review impossible to scale. Quality now depends on automated constraints like unit tests, security scans, and complexity checks that act as gatekeepers throughout the pipeline. These constraints create back-pressure, catching bad output early before it reaches production. The key is placing tight constraints where quality matters most, while relaxing them elsewhere to keep delivery fast.

And the most popular article from the last issue was:

• AI is removing the middle class of software engineering