Serving the For You feed

2026-04-24 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Intermediate, quick

Summary

The Bluesky "For You Feed," operated by spacecowboy and serving approximately 72,000 users, demonstrates a highly efficient and decentralized architecture. This custom feed, a core feature of Bluesky's AT Protocol, runs on a single Go process utilizing SQLite on a personal "gaming" PC with 16 cores, 96GB RAM, and 4TB NVMe storage. It consumes the Bluesky firehose, storing 90 days of relevant data, which currently amounts to 419GB in SQLite. Recommendations are generated based on user likes, identifying content liked by individuals with similar preferences. Public internet traffic is managed by a $7/month OVH VPS, communicating with the home server via Tailscale. The total operational cost is $30/month, comprising $20 for electricity, $7 for the VPS, and $3 for domain names. The system is estimated to scale to 1 million daily active users with its most efficient algorithm.

Key takeaway

For AI Engineers or MLOps teams considering deploying custom recommendation engines, this architecture demonstrates that significant user loads can be served from highly cost-effective, even consumer-grade, hardware. You should evaluate the potential of SQLite and a decentralized setup with secure tunneling (like Tailscale) to drastically reduce infrastructure costs, especially for services that can tolerate a single point of failure or are designed for community-driven hosting.

Key insights

Decentralized social feeds can be efficiently served from consumer-grade hardware with minimal operational costs.

Principles

• Leverage existing hardware for cost efficiency.
• SQLite can manage large, active datasets.
• Decentralized services enhance platform resilience.

Method

The method involves consuming a platform firehose, storing relevant data in SQLite, generating recommendations based on collaborative filtering (likes), and using a low-cost VPS with Tailscale for secure public access.

In practice

• Use Tailscale for secure server-to-VPS links.
• Consider SQLite for high-volume data storage.
• Optimize algorithms for low-cost scaling.

Topics

• Bluesky Custom Feeds
• AT Protocol
• Go Programming Language
• SQLite Database
• Tailscale VPN

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Designing Scalable Product Catalog APIs With Cassandra, GraphQL, and AI — Scalable retail systems require balancing GraphQL flexibility, AI unpredictability, and Cassandra's structured query demands.
• datasette-tailscale 0.1a0 — An experimental Datasette plugin integrates with Tailscale, enabling secure, private network access to Datasette instances via a sidecar.
• This Keyboard — ThePrimeagen promotes community platforms, a backend dev course, and Turso's production-ready SQLite.
• From Tailnet to platform (Interview) — Tailscale is expanding its secure networking platform with new identity, application, and AI-focused gateway features.
• Uber’s Hive Federation Decentralizes 16K Datasets and 10+ PB for Zero-Downtime Analytics at Scale — Decentralizing monolithic data warehouses via pointer-based migration enhances scalability, security, and operational autonomy.
• Building a Production RAG Ingestion Pipeline on AWS: Unstructured.io, S3 Vectors, and a Private VPC — Self-hosting Unstructured.io within a private AWS VPC enables robust, secure, and cost-effective RAG ingestion.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.