DuckDB previews client-server mode with Cyanoptera
DuckDB 2.0, codenamed Cyanoptera, adds a client/server mode with a CONNECT statement, targeted for autumn 2026. SageMaker HyperPod lets data scientists create Ray clusters from Studio instead of writing Kubernetes manifests.
Artificial Intelligence··Midday
DuckDB Cyanoptera answers over the network
The preview of DuckDB version 2.0, codenamed Cyanoptera, adds a client/server mode through a new protocol extension and a CONNECT statement, along with multi-version concurrency control for multi-tenant deployments. General availability is targeted for autumn 2026. InfoQ counts more than 10,000 commits since version 1.5. The analytical engine that lived inside a single process now, in this preview, answers clients over the network from outside that machine.[1]
A VARIANT type and vector joins arrive
The release also brings a VARIANT type that shreds semi-structured JSON into columnar storage, a PEG-based parser replacing the PostgreSQL-derived grammar, BEFORE and AFTER triggers, approximate nearest-neighbour similarity joins for vector work, asynchronous input and output against object stores such as Amazon S3, and a versioned C interface with a stable binary layer for extensions. The approximate nearest-neighbour joins put vector search on Cyanoptera's SQL path; VARIANT shreds JSON into columns instead of carrying it row by row.[1]
SageMaker HyperPod starts a Ray cluster from the console
AWS has added managed Ray support to Amazon SageMaker HyperPod, so data scientists can create and run Ray clusters from the SageMaker Studio console rather than writing Kubernetes manifests or kubectl commands. The capability is available in every AWS Region where SageMaker HyperPod on EKS already runs. The release covers a built-in Ray Dashboard with Amazon Managed Grafana views for Ray Core, Data, Train and Serve; remote job submission through a toolkit package that authenticates with IAM; JupyterLab and Code Editor spaces that attach to a cluster as worker nodes contributing no compute; automatic node recovery with hung-job detection and tiered checkpointing; and Ray Serve inference with SageMaker JumpStart model loading and a managed tiered key-value cache for long-context serving.[2]
