Happy Monday! ☀️
Welcome to the 43 new hungry minds who have joined us since last Monday!
If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.
📚 Software Engineering Articles
🗞️ Tech and AI Trends
👨🏻💻 Coding Tip
Use TypeScript branded types to keep IDs with different meanings from getting mixed up.
Time-to-digest: 5 minutes
Cloudflare’s D1 is a serverless SQLite database built to support millions of small databases. This model powers multi-tenant platforms and AI-assisted development tools, but introduces a hard question: how do you manage metadata and routing for millions of databases without making every query slower?
The answer lies in moving coordination away from the hot path. D1 experienced 15x growth in live databases over one year and 18x growth in daily active databases—growth that forced the team to rethink their control plane architecture.
The challenge: Scale metadata operations and database lifecycle management to support millions of instances without centralizing bottlenecks that slow down query execution.
Implementation highlights:
Distributed creation workflow – Database creation runs near Postgres for metadata writes (milliseconds) but initializes the backing Durable Object near the user’s location, avoiding long-distance blocking operations.
Workers KV as edge cache – Critical routing metadata (database ID to Durable Object ID mappings) lives in globally distributed KV, so data plane queries can route directly without hitting central databases.
Periodic heartbeats for self-healing – Each Durable Object heartbeats Postgres every 30 seconds to detect deleted databases and updated limits, allowing the system to self-correct without strong consistency requirements.
Hyperdrive for efficient pooling – Central Postgres queries use Hyperdrive’s connection pooling and read replicas, reducing latency and connection overhead for metadata operations.
Eventual consistency between planes – Control plane operations (CRUD) tolerate central single-region components; only data plane queries (the hot path) must avoid centralized bottlenecks.
Results and learnings:
Failure-driven design – Retrying database creation with fresh Durable Object IDs and using heartbeats to detect orphaned objects prevents subtle split-brain scenarios.
Load testing drove priorities – Periodically testing above expected growth revealed which components needed optimization, preventing overengineering uncontended paths.
Separation of concerns pays off – Splitting operations into data plane (fast), control plane (tolerates latency), and background async (eventual consistency) gave each tier the right consistency guarantees.
Cloudflare keeps central coordination off the query path. Workers KV caches routing data at the edge, while heartbeats repair stale state. Queries can proceed without a round trip to Postgres.
ARTICLE (review velocity tradeoff)
Speeding up code review with small-change
ESSENTIAL (pattern matching speedrun)
Aho-Corasick algorithm for simultaneous substring matching
ARTICLE (json compression wins)
Packing binary is fun
ARTICLE (training data determinism)”
Bringing Elastically Deterministic Data Loading to TorchTitan with Zephon
ARTICLE (code review evolves)
Code Review After Code Gets Cheap
ARTICLE (sandbox all the things)
Perplexity Open-Sources SPACE, a Firecracker MicroVM Sandbox for AI Agents
ARTICLE (bit twiddling wizardry)
Superoptimizing an FPU emulator
ARTICLE (deflate from scratch)
tugz: a new deflate library with three applications
Want to reach 200,000+ engineers?
Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.
💾 Anthropic Built Scion, A Custom Source Control System, To Handle 30,000+ Daily Code Merges (8 min)
Brief: Anthropic built Scion, an in-house version control system, to replace git after code merges surged from 2,500 to over 30,000 per day driven by Claude’s autonomous coding capabilities.
Brief: The Deno team is joining Cloudflare to make Workers and Durable Objects easier to self-host, combining work on celld and workerd.
Brief: Muse Wonder brings Active Memory and Active Focus to agent tasks and is available through Meta’s API and Muse Code. Wonder Full is limited to Muse Code subscribers.
Brief: TanStack Start 1.0 is stable. The team says its full-stack framework is ready for production.
Brief: OpenAI released prompt-to-voice for all API customers, enabling creation of custom voices from text descriptions for reuse in GPT-Live sessions.
Brief: Google opened Gemini 4 Argon to developers on the paid Gemini API tier, targeting long-running work such as code migrations and research.
This week’s tip:
Use branded types to keep IDs apart
TypeScript treats two string IDs as interchangeable. Give each one a distinct brand so the compiler catches mix-ups before they reach an API call.
Wen?
Use brands at boundaries where several IDs share a primitive type. Keep the casts in a small set of constructors so callers cannot invent IDs throughout the codebase.
“Try not to become a man of success, but rather become a man of value.” – Albert Einstein
That’s it for today! ☀️
Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!
If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦
Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.
Icons by Icons8.
*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).