Meta Puts Idle Compute to Work 💤, GitHub Counts Stars Faster ⭐, Claude Dashboard Security 🔐

Meta uses idle computing capacity to improve ad recommendations by ranking more ads or routing requests to more powerful models. ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR Data 2026-10-15 Efficient search with OLTP (Sponsor) Put the duct tape away. You don't have to tack on a standalone search engine to your database to solve your AI-agent-scale search problem. Lakebase Postgres now has a fast, scalable and serverless built-in search engine. Separating storage from compute means you don't have to keep your entire search index in RAM just to keep queries fast. Compute scales with actual search activity, while your index stays in lower-cost storage. Instead of chaining together multiple APIs for vector search, keyword search and filtering, you can do it all in one SQL query, against the same live operational data. Quench your need for Postgres speed Start building 📱 Deep Dives Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations (16 minute read) Meta uses idle computing capacity to improve ad recommendations by ranking more ads or routing requests to more powerful models. When that capacity disappears, the system quickly scales back this extra work while protecting normal service speed and reliability. From 2,500 requests to 24: Rebuilding star history without the stargazers (7 minute read) GitHub built a star history API that returns daily counts without exposing users' identities, cutting React's full history from about 2,500 requests to 24. It reduces database work by filtering stars by date before counting them, rather than repeatedly processing every star. Zero-downtime Postgres upgrades with logical replication (25 minute read) Logical replication offers a practical path to near-zero-downtime PostgreSQL major upgrades, with live rollback enabled by reverse replication. A Docker-based demonstration ran old and new clusters in parallel, freezing writes and waiting for replication to catch up before cutover. It achieved 181 ms of write interruption during upgrade and 179 ms during rollback, without data loss. CASTLE: training natural language search before the first real query (6 minute read) Cold-start natural-language search often lacks both real queries and relevance labels. Airbnb's CASTLE jointly generates synthetic queries and labels from de-identified booking sessions. Its queries matched real guest query patterns much better than baselines, and human reviewers agreed with its relevance labels >90% of the time. Carefully governed behavioral data can bootstrap search training before meaningful traffic exists. 🚀 Opinions & Advice Need to Know: Implementing row-level security for Claude Dashboards via Model Context Protocol (11 minute read) This proof of concept shows how to add row-level security to Claude Dashboards using an MCP server to authenticate users and query DuckDB for only the data they can access. Access rules are enforced before data reaches Claude, allowing the same dashboard to show different results for each user. Four PostgreSQL arguments that came out even (2 minute read) PostgreSQL micro-optimization alternatives often don't matter. For 1,000 IDs, IN, ANY, and VALUES joins perform similarly; prefer parameterized ANY for stable query text. Short text, enums, and smallints use near-identical storage; mostly-NULL columns are cheap. Don't blindly rewrite but EXPLAIN on your own data. Rewinding the exponential: How we keep hardware usage down (9 minute read) Cloudflare held back exponential hardware growth by shrinking data earlier and moving duplication later: split monolithic schemas, switch to zstd, send each event once then replicate after sampling, sample at the edge, and pre-aggregate known queries. Optimize bytes, not just compute: prune fields, compress measured workloads, push reductions upstream, and replicate compacted data. 💻 Launches & Tools Browse ready-made AI assets. Run on live data in minutes. (Sponsor) Precisely AI Studio is a self-serve library of AI apps, agents, skills, prompts, and MCP connectors organized by industry and data domain. Connect to live, governed data and go from browsing to running in minutes. Try it free, no commitment, no strings. Try AI Studio free See what's inside SQLBuild: safe refactoring for warehouse SQL (GitHub Repo) SQLBuild is an open-source SQL and Python transformation framework for safely changing warehouses. It checks contracts and lineage before execution, runs graph-level tests and production diffs, and lets teams rename incremental models or columns while preserving history and compatibility views. A local DuckDB playground makes it easy to evaluate. Snowflake, Postgres, BigQuery, and Databricks adapters are available. Phoenix (GitHub Repo) Phoenix is a self-hosted, open-source AI observability platform for tracing, evaluating, and debugging LLM applications. It captures OpenTelemetry traces; offers retrieval and response evaluations, versioned datasets, and experiments; and supports prompt testing and replay. Framework- and provider-agnostic instrumentation makes it useful for teams that need to diagnose agent behavior and measure changes before production. XERJ (GitHub Repo) XERJ is an Apache-2.0, Elasticsearch-compatible local search engine that automatically indexes code, logs, PDFs, SQLite, and CSVs into typed datasets for agent retrieval, incident analysis, and RAG. Its autoindex workflow avoids schema setup and returns ranked passages with file and line locations. The project defaults to local operation, although optional reranking can send text to an external provider. 🎁 Miscellaneous Redesigning source control at Anthropic (17 minute read) Anthropic replaced Git with Scion, an in-house version control system built in five months, after daily code merges grew from 2,500 to over 30,000. By moving operations to servers and fetching only the files needed, it reports roughly 100 times faster checkouts and 200 times lower CPU and memory requirements than its previous Git mirror infrastructure. Parsing JSON at tens of gigabytes per second… in Python (5 minute read) Daniel Lemire's benchmark processed 7.6 GB of JSON in about 0.12 seconds using fastsimdjson on free-threaded Python across 64 cores, reaching roughly 66 GB/s. The speed comes from running threads in parallel and creating Python objects only for the fields needed, with data already decompressed in memory. ⚡ Quick Links Jev handed our LLM the wrong page a tenth as often as our embeddings (20 minute read) Jev picked the right page first 88% of the time (vs. 70% for embedding search) across 168 test questions, and cut irrelevant results to the AI receptionist from 33% to 3% by answering "none" when no page fit. What gets TOASTed in Postgres (8 minute read) Postgres's TOAST compresses large values or moves them to a separate table. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://sparklp.co/32815a84/11 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of data engineering professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here, create your own role or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! TLDR is one of Inc.'s Best Bootstrapped businesses of 2025. If you have any comments or feedback, just respond to this email! Thanks for reading, Joel Van Veluwen, Tzu-Ruey Ching & Remi Turpaud Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR Data isn't for you, please unsubscribe.