Arc Notes Weekly #103: Trio
This week, why pgvector might not cut it in production, how bloom filters fail at scale, and why LLMs can't replace hands-on learning. Plus, slashing cloud costs and Postgres as a lakehouse system.
This week, uncover why dedicated vector databases often beat pgvector for production workloads despite Postgres’s appeal, how bloom filters create clever but ultimately flawed full text search systems, and why LLMs can’t replace the essential learning loop that builds true software development expertise.
Sponsor Spotlight: QCon San Francisco — Enterprise AI Architecture Architecting AI systems? Evaluating LLM infra or planning migrations? QCon San Francisco (Nov 17–21) dives deep into real-world AI architecture, platforms, and patterns—before they go mainstream. Learn from early adopters. Make smarter decisions. Learn more at qconsf.com →
Enjoy this week's round-up!
— Mahdi Yusuf (@myusuf3) or LinkedIn
👋🏾 You are reading Architecture Notes - Your Sunday newsletter, which curates best system design and architecture news from around the web. We would appreciate you sharing it with like-minded people.
Articles
Why Pgvector Might Not Be the Best Choice for Production
While pgvector is a popular choice for vector search in Postgres, Alex Jacobs highlights the significant challenges it poses in production environments, such as memory-intensive index rebuilds and suboptimal query planning. Discover why dedicated vector databases might be a simpler and more efficient solution!A Clever Yet Flawed Approach to Full Text Search
Explore the intriguing use of bloom filters for building a space-efficient full text search index, and discover why scaling this technique to large document corpuses might not be the best idea. The article delves into the challenges of query performance and the limitations of bloom filters compared to traditional inverted indexes, offering a fascinating look at the balance between space efficiency and practicality.Navigating Travel’s Hidden Bureaucratic Maze
Discover how a simple error fare to Iceland uncovers the complex web of travel regulations and the ingenious app that helps you navigate them without a hitch! From visa requirements to tax residency implications, this article dives into the intricate dance of international travel logistics.Why Ditching the Cloud Saved Me Big Bucks
Discover how one developer slashed his AWS bill from $1,400 to under $120 a month by moving off the cloud, and why he believes many are being scammed by cloud providers. This article dives into the backlash he faced and the real cost of cloud services.Enterprise AI Architecture at QCon San Francisco (Nov 17-21)
Whether you’re architecting multi-model serving infrastructure, evaluating RAG pipeline investments, or planning GenAI platform strategy, QCon San Francisco 2025 connects you with enterprise architects who’ve shipped it.
Learn from Netflix’s AI platform architect, Pinterest’s GenAI search lead, and Coinbase’s ML infrastructure engineer. Sessions cover autoscaling inference for hundreds of model variants, data access patterns for AI workloads, and balancing reliability with innovation.
Hear what worked, what cost too much, and what to avoid from 60+ senior practitioners who adopted first.
Why Context Engineering Is Crucial for LLMs
Discover why prompts alone aren’t enough for LLMs to tackle complex engineering problems. Learn how context engineering, which combines code, docs, and conversations, empowers models to deliver accurate and relevant solutions in real time.Why Composition Became the Darling Over Inheritance
Explore the shift from inheritance to composition in software design, tracing its roots to the “Gang of Four” and the evolving landscape of programming languages. Discover why composition offers flexibility and runtime adaptability, challenging the static nature of inheritance.Unlocking the Hidden Power of URLs in Web Development
Discover how URLs can be more than just web addresses—they can be powerful state management tools! Learn how to use URLs to store, share, and recover application states without relying on databases or cookies. This article explores the overlooked potential of URLs as state containers and offers practical tips for frontend engineers.Why LLMs Can’t Replace the Learning Loop in Software Development
Discover why the assembly line metaphor fails in software development and how LLMs, while useful for brainstorming and setup, can’t replace the essential learning loop that drives true expertise. Unmesh Joshi shares insights on using LLMs judiciously and the importance of hands-on learning in building robust systems.Projects
TOON
TOON is a new, compact, and human-readable format designed to optimize token usage when passing structured data to Large Language Models. It combines the best of YAML and CSV to provide a schema-aware JSON alternative, perfect for uniform arrays of objects. Dive into the benchmarks and see how TOON can save you tokens and money!pg_lake
Discover how ‘pg_lake’ transforms Postgres into a powerful lakehouse system, enabling seamless transactions and fast queries on Iceberg tables, while directly accessing raw data files in object stores like S3. Dive into the world of PostgreSQL extensions and explore the integration with DuckDB for enhanced performance!
Architectural Debt: More Than Just Technical Debt
Architectural debt isn’t just about the code; it’s the structural decisions that haunt you later. Dive into how enterprise architects tackle the complexities of application interactions, data flow, and strategic misalignments that can cost organizations dearly.