AWS Glue 6.0 Cuts Pricing 30% and Adds Full Apache Iceberg v3 Support
AWS Glue 6.0 is now generally available across all regions where Glue operates, bringing 30% lower pricing than previous Glue versions alongside a modernized runtime built on Apache Spark 4.1, Python 3.12, and Scala 2.13. No API changes are required to upgrade — existing jobs can move to the new version via the --glue-version parameter through the AWS CLI, SDK, Glue Studio, or SageMaker Unified Studio, and an auto-upgrade feature is available for existing jobs.
The headline feature is full Apache Iceberg v3 support, built on Iceberg 1.11.0. That includes the new VARIANT data type with shredding for semi-structured JSON, logs, and event data; Geometry and Geography types for spatial processing; nanosecond-precision timestamps aimed at IoT and financial workloads; and unknown-type handling to keep schema evolution from breaking existing pipelines. On the ETL side, Spark Declarative Pipelines reduce orchestration boilerplate and Arrow-native Python UDFs/UDTFs eliminate serialization overhead between the JVM and Python runtimes. Glue 6.0 also claims single-digit millisecond latency for stateless streaming scenarios.
For teams running data lakes on Iceberg, this is a significant upgrade: the combination of lower cost and native v3 feature support (particularly VARIANT and nanosecond timestamps) removes several workarounds that Iceberg v2-era pipelines needed for semi-structured and high-precision time-series data.
Read more — AWS News Blog
AWS Weekly Roundup: EC2 Application Status Checks and IAM Role Manager
AWS's August 17 weekly roundup introduced two features aimed at reducing operational toil. Amazon EC2 Application Status Checks add application-level health monitoring on top of the existing instance-level checks, detecting problems like a stopped web server, a missing Docker daemon, networking misconfiguration, or failed traffic on a network interface — issues that a healthy-looking EC2 instance can otherwise mask for minutes or hours before someone notices.
The second feature, AWS IAM Role Manager, automates IAM role setup for AWS services by creating default roles or reusing existing ones that already match the required permissions, launching with support for six AWS service consoles. For teams that have historically hand-rolled IAM roles per service integration — a common source of both over-permissioned roles and broken deployments — this removes a recurring manual step without giving up visibility into what permissions get granted.
The same roundup also covered a next-generation AWS VPN Client built on the OpenVPN3 architecture with CLI support and admin controls, maintaining backward compatibility with existing Client VPN endpoints, plus new SageMaker JumpStart foundation models including NVIDIA's Nemotron 3.5 Lightning and Qwen3.5-122B-A10B.
Read more — AWS News Blog
Google Cloud Extends Compute Flex CUDs to G2 and G4 GPU VMs
Google Cloud expanded Compute Flex Committed Use Discounts (CUDs) to cover G2 (NVIDIA L4) and G4 (NVIDIA RTX PRO 6000) GPU VM families. Flex CUDs let customers lock in predictable savings while retaining the flexibility to shift spend across VM families and regions, and the expanded coverage means a single spend commitment can now span general-purpose compute, GKE, Cloud Run, and these two GPU families together.
This matters for teams running mixed workloads — for example, training or inference on L4/RTX Pro GPUs alongside general-purpose serving on standard VMs — since previously that flexibility required either committing separately per resource type or forgoing the discount on the GPU portion of the workload. Teams with existing Flex CUD commitments should check whether reallocating budget to include G2/G4 usage now yields a better effective discount rate than their current mix.
Read more — Google Cloud Blog
Rapid Bucket Brings Google's Colossus Storage to PyTorch via GCSFS
Google Cloud detailed how Rapid Bucket — a zonal, high-performance Cloud Storage class — integrates with PyTorch through the gcsfs fsspec interface, replacing REST-based storage access with persistent bidirectional gRPC streams backed by Google's Colossus architecture. The result is a claimed 23% training speed improvement and 4.8x read throughput gain, with GCSFS's adaptive concurrent prefetching boosting single-file throughput by up to 5x and scaling to 21 GiB/s when paired with Rapid Bucket.
Unlike standard regional or multi-region buckets, Rapid buckets are zonal, so they can be co-located with the GPU or TPU compute consuming them, cutting latency substantially. Adoption requires zero code changes beyond switching to a Rapid Bucket and upgrading gcsfs to version 2026.3.0 or later — a standard fsspec.open() call picks up the speed gains automatically. Rapid Storage already backs data preparation workflows for Dask, Pandas, Hugging Face Datasets, and Ray Data, and checkpointing for PyTorch Lightning, torch.dist, and Weights & Biases, making this a low-friction win for teams already training on GCP.
Read more — Google Developers Blog
Links & Sources
- AWS Glue 6.0 now available with 30% lower price and full Apache Iceberg v3 support
- AWS Weekly Roundup: EC2 application status checks, IAM role manager, OpenAI Daybreak on Bedrock, and more
- Expanded coverage for Compute Flex CUDs
- Speeding Up AI: Bringing Google Colossus to PyTorch via GCSFS and Rapid Bucket
