Google Cloud adds pause & resume to Dataflow batch jobs

Google Cloud has introduced new Dataflow features for batch processing and AI inference, including general availability of Pause/Resume for batch jobs and support for G4 virtual machines with NVIDIA RTX PRO 6000 Blackwell GPUs.

The changes target customers running long batch workloads and AI inference tasks on Dataflow, part of Google Cloud's data processing and machine learning portfolio.

Dataflow is used to build batch and streaming pipelines for analytics and AI workloads. One of the main issues for customers running large batch jobs has been the need to restart an entire workload after a failure, even when part of the work has already been completed.

Pause/Resume for Dataflow batch jobs addresses that problem by letting failed long-running jobs restart from where they stopped rather than from the beginning. It also lets users pause lower-priority batch jobs and resume them later, which could help teams reassign scarce GPU or TPU resources to more urgent tasks.

Batch recovery

Some Dataflow customers run batch workloads that last several days. In those cases, restarting after a failure can increase cloud spending and delay engineering work.

Google Cloud said the Pause/Resume feature is based on internal technology and is now generally available. The company described it as a way to reduce wasted compute and improve productivity for teams managing long-running data preparation or model-related workloads.

The second part of the update expands the hardware customers can use for inference inside Dataflow. Support for G4 virtual machines based on Nvidia's RTX PRO 6000 Blackwell Server Edition gives users access to newer GPU infrastructure for machine learning tasks.

According to Google Cloud, the Nvidia RTX PRO 6000 Blackwell GPU offers higher performance than the Nvidia L4 GPU, along with 96GB of vGPU memory and 1.6 TB/s of bandwidth. In practical terms, that could allow inference within a Dataflow job on models with more than 70 billion parameters.

Inference push

The announcement reflects wider competition among cloud providers to attract AI workloads that require both data processing and model execution in the same managed environment. For customers, one advantage is the ability to run inference within existing pipelines instead of moving data between separate systems.

Google Cloud said the new GPU support works with existing Dataflow machine learning features, including RunInference, right fitting and GPU-enabled autoscaling. Those tools are designed to let customers adjust resources for inference jobs while keeping the service managed.

That matters for users who want to combine data preparation, model evaluation and inference in one workflow. As AI systems move into production, demand has grown for cloud services that can handle continuous data movement as well as the computing needs of large models.

Dataflow has long been part of Google Cloud's managed data processing strategy, particularly for Apache Beam-based workloads spanning batch and streaming jobs. The latest additions suggest the company is sharpening the product for customers that want to use it more directly in AI pipelines rather than only as a data engineering service.

By adding a resume function for interrupted batch jobs, Google Cloud is also addressing a practical concern for teams running expensive, long-duration workloads. In AI development, where jobs may prepare training data or run large-scale feature engineering tasks over several days, restarting from scratch can carry both financial and operational costs.

At the same time, support for newer Nvidia hardware highlights the pressure on cloud vendors to keep pace with rapid changes in AI infrastructure. Customers increasingly want managed services that provide current-generation chips without requiring them to build and tune their own infrastructure stack.

Google Cloud said the new support allows users to carry out AI inference directly in Dataflow jobs on models with more than 70 billion parameters while continuing to use native Dataflow machine learning features.