Fable Evals Performance ✅, Airbnb’s Evolved Data Architecture 🏘️, PostgreSQL Differential Privacy 🎭

Open original

This newsletter rounds up deep dives on modern data tooling—from Fable 5’s performance on complex analytic tasks to Airbnb’s flexible multi-product data architecture and PostgreSQL Anonymizer’s new local differential privacy features. It also links to articles on lakehouse governance with Gravitino, vector storage advances, streaming runtimes, and orchestration reactions to Dagster’s price hike.

  • Claude Fable 5 beats other frontier models by 10–15% on Hex's messy, long-horizon eval sets designed to test judgment and assumption clarity.
  • Airbnb's offline data pipeline enforces plain-vanilla modeling, uniform IDs, and strict namespaces to let teams build domain-specific features without breaking cross-product consistency.
  • PostgreSQL Anonymizer 3.1 adds true Local Differential Privacy via GRRM, letting users tune epsilon for surveys and categorical data alongside six PII-masking methods.
  • Capital One cut cloud-dormancy analysis from nine months to ten days using AI-generated Spark SQL with human validation.

On June 16, Ben Brook will walk you through wiring first-party data into every AI workflow. Transcend’s platform embeds permissions in the systems that process customer records, so every model and data product draws from an up-to-date “source of truth.” Their report, based on interviews with 220 IT and business leaders, breaks down why most AI projects stall and how to avoid those pitfalls immediately. QuestDB’s engine combines vectorized C++ kernels, Java batch work and JIT filtering under a tuple-at-a-time SQL interface. Flip a GROUP BY or tweak a WHERE clause and you could switch from SIMD-driven scans to frame-based parallelism. Airbnb rebuilt its offline pipeline to serve multiple products: they enforce plain-vanilla modeling (no hybrids), uniform IDs and strict namespaces, so teams can build domain-specific features without breaking cross-product consistency. And Apache Gravitino now lets you govern Iceberg tables and multimodal Lance datasets with a single metadata layer—watch out for config quirks like enum casing and client-jar mismatches. Claude Fable 5 outperforms other frontier models by 10–15 percent on Hex’s messy, long-horizon eval sets. The team wrote new benchmarks to measure judgment, assumption clarity and semantic-to-raw-data cross-checks. Meanwhile, high-update lakehouses need mutation-friendly metadata. Apache Hudi’s Merge-On-Read metadata table uses append-first writes plus deferred compaction to cut write costs and scale indexing better than Copy-On-Write. Gorgias ran into overlapping windows and wrong aggregates when building billing on Flink; they solved it by aligning keys across operators and adding small replay delays only on historical reprocesses. Loon—the new storage engine behind Milvus 3.0 beta and Zilliz Vector Lakebase—lets you update scalars, vectors and object refs separately via versioned manifests and hybrid file formats. Streamling, built in Rust with Arrow and DataFusion, targets transactional streaming on one node, supporting Kafka, Postgres, ClickHouse and TypeScript/WASM transforms with checkpointing and exactly-once delivery. PostgreSQL Anonymizer 3.1 now offers six PII-masking methods and true Local Differential Privacy via GRRM, letting you tune ε for surveys and categorical data. Finally, Salesforce Data 360’s Zero Copy moved from query federation to Iceberg file federation, slashing cross-system compute while keeping governance intact. Capital One cut cloud-dormancy analysis from nine months to ten days by feeding asset inventories through AI-generated Spark SQL and human validation.

Ask about this article