AWS Glue Zero‑ETL now replicates DynamoDB tables to Apache Iceberg on S3, letting you query semi‑structured data with Athena without custom pipelines. Built‑in schema unnesting and partition controls flatten nested attributes and optimize query performance, dramatically reducing development effort for analytics and ML workloads.
The post shows three ways to speed up Amazon Redshift queries on S3‑backed Apache Iceberg tables: using external schemas to shorten SQL, materialized views that cache results in Redshift, and S3 Tables compaction to align file layout with query patterns. Together they cut scan time and lower costs for dashboards and ad‑hoc analysis.
Databricks' Lakebase Change Data Feed (CDF) lets teams enable CDC on all tables in a minute, exposing operational data directly to the lakehouse via Unity Catalog. This eliminates the need for separate extraction pipelines, providing governed, real‑time feeds for streaming, materialized views, and AI workloads.
Tableau’s new Microsoft 365 app lets users embed live Tableau visualizations directly into Word documents, PowerPoint slides, and Teams chats, keeping data up‑to‑date without leaving the workflow. The integration aims to reduce screenshot copying and ensure decisions are based on current insights.
Databricks unveiled an LLM inference platform that can serve any frontier model, including open‑source and proprietary ones, while handling over 120 trillion tokens a month. The system tackles reliability and latency at scale with multi‑tenant GPU clusters, custom kernels, and fault‑tolerant design, keeping high‑throughput agents responsive despite spiky demand.
Snowflake warns that mismatched definitions in a financial services’ semantic layer let AI models silently misinterpret data, creating drift and governance gaps. Without shared conceptual standards, models can produce confident yet wrong outputs, exposing firms to compliance and operational risk.
Meta unveiled SilverTorch, a unified neural‑network retrieval system that replaces traditional microservice pipelines with an “Index as Model” architecture. In internal tests it delivered up to 23.7× higher request throughput and 20.9× better compute cost efficiency, while improving recommendation accuracy across feed and video apps.
Databricks now extends its built‑in prompt‑caching feature, previously limited to proprietary models, to open‑source LLMs on its platform. By reusing KV caches for repeated prompts, it cuts compute, latency, and cost for batch and real‑time inference, all without extra configuration.
Databricks’ Lakebase architecture makes Postgres stateless and stores all data in zone‑redundant object storage, allowing instant compute replacement without costly hot‑standby replicas. This design lets the service sustain cloud capacity shortages and massive agent‑driven workload spikes, delivering higher availability at lower cost.
DuckDB 1.5.3, while a patch release, adds the Quack client‑server protocol as a core extension and extends DuckLake support for Quack‑based catalogs. It also brings new AWS extension capabilities like IRSA web‑identity and IAM authentication for RDS/Aurora, plus HTTPS proxy enhancements.
You’re caught up.
That was all 10 stories in the 28 May issue.