Ray-2.55.0

The Ray Project, maintainers of the popular open-source distributed computing framework, has launched Ray 2.55.0. This release is a substantial upgrade focused on its Ray Data component, introducing a new DataSourceV2 API that provides a modern scanner/reader framework with improved file listing and partitioning capabilities. A standout feature is the support for GPU-accelerated data shuffling using RAPIDS 26.2, which can dramatically speed up preprocessing for machine learning workloads. The update also integrates a new Kafka datasink, migrating to the `confluent-kafka` library and supporting datetime offsets, making it easier to build real-time data pipelines that feed directly into AI training jobs.

Beyond new connectors, version 2.55.0 brings significant performance optimizations and operational improvements. It introduces queue-based autoscaling policies integrated with task consumers and enables autoscaling specifically for GPU-intensive stages, allowing infrastructure to scale efficiently with workload demands. The release also patches a critical remote code execution (RCE) vulnerability in Arrow extension type deserialization from Parquet files. Other enhancements include better resource management for object store usage, a new ExecutionCache for streamlined data caching, and the addition of a vLLM metrics export with a dedicated Grafana dashboard for monitoring large language model inference workloads within the Ray ecosystem.