Use Dataflow Prime

Dataflow Prime is a serverless data processing platform for Apache Beam pipelines. Based on Dataflow, Dataflow Prime uses a compute and state-separated architecture. In the following cases, Dataflow Prime might improve pipeline efficiency:

Your pipeline would benefit from Vertical Autoscaling.

Dataflow Prime supports both batch and streaming pipelines. By default, Dataflow Prime uses Dataflow Shuffle and Dataflow Runner v2 for batch pipelines.

SDK version support

Dataflow Prime supports the following Apache Beam SDKs:

Apache Beam Python SDK version 2.21.0 or later
Apache Beam Java SDK version 2.30.0 or later
Apache Beam Go SDK version 2.44.0 or later

To download the SDK package or to read the Release Notes, see Apache Beam Downloads.

Dataflow Prime features

The following is the list of supported Dataflow Prime features for different kinds of pipelines:

Vertical Autoscaling (memory). Supports streaming pipelines in Python, Java, and Go.
Right fitting (resource hints). Supports batch pipelines in Python and Java.
Job Visualizer. Supports batch pipelines in Python and Java.
Smart Recommendations. Supports both streaming and batch pipelines in Python and Java.
Data Pipelines. Supports both streaming and batch pipelines in Python and Java.

The features Job Visualizer, Smart Recommendations, and Data Pipelines are also supported for non-Dataflow Prime jobs.

Vertical Autoscaling

This feature automatically adjusts the memory available to the Dataflow worker VMs to fit the needs of the pipeline and help prevent out-of-memory errors. In Dataflow Prime, Vertical Autoscaling works alongside Horizontal Autoscaling to scale resources dynamically.

For more information, see Vertical Autoscaling.

Right fitting

This feature uses resource hints, a feature of Apache Beam. By using resource hints, you can specify resource requirements either for the entire pipeline or for specific steps of the pipeline. This feature lets you create customized workers for different steps of a pipeline. Right fitting lets you specify pipeline resources to maximize efficiency, lower operational cost, and avoid out-of-memory and other resource errors. It supports memory and GPU resource hints.

Right fitting requires Apache Beam 2.30.0 or later.

For more information, see Right fitting.

Job Visualizer

This feature lets you see the performance of a Dataflow job and optimize the performance of the job by finding inefficient code, including parallelization bottlenecks. In the Google Cloud console, you can click any Dataflow job in the Jobs page to view details about the job. You can also see the list of steps associated with each stage of the pipeline.

For more information, see Execution details.

Smart Recommendations

This feature lets you optimize and troubleshoot the pipeline based on the recommendations provided in the Diagnostics tab of the job details page. In the Google Cloud console, you can click any Dataflow job in the Jobs page to view details about the job.

For more information, see Recommendations and diagnostics.

Data Pipelines

This feature lets you schedule jobs, observe resource utilizations, track data freshness objectives for streaming data, and optimize pipelines.

For more information, see Working with Data Pipelines.

Quota and limit requirements

Quotas and limits are the same for Dataflow and Dataflow Prime. For more information, see Quotas and limits.

If you opt for Data Pipelines, there are additional implications for quotas and regions.

Unsupported features

Dataflow Prime does not support the following:

Designating specific VM types by using the flag --worker_machine_type or --machine_type for Python pipelines and --workerMachineType for Java pipelines.
Viewing or using SSH to log into worker VMs.
The class OrderedListState for Java pipelines.
Flexible Resource Scheduling (FlexRS).
Using VPC Service Controls with Vertical Autoscaling. If you enable Dataflow Prime and launch a new job within a VPC Service Controls perimeter, the job uses Dataflow Prime without Vertical Autoscaling.
NVIDIA Multi-Process Service (MPS).
Java pipelines that meet the following requirements can use the MapState and SetState classes:
- use Streaming Engine
- use Apache Beam SDK versions 2.58.0 and later
- don't use Runner v2

All pipeline options not explicitly mentioned previously or in the feature comparison table work the same for Dataflow and Dataflow Prime.

Before using Dataflow Prime

To use Dataflow Prime, you can reuse your existing pipeline code and also enable the Dataflow Prime option either through Cloud Shell or programmatically.

Dataflow Prime is backward compatible with batch jobs that use Dataflow Shuffle and streaming jobs that use Streaming Engine. However, we recommend testing your pipelines with Dataflow Prime before you use them in a production environment.

If your streaming pipeline is running in production, to use Dataflow Prime, perform the following steps:

Stop the pipeline.
Enable Dataflow Prime.
Rerun the pipeline.

Enable Dataflow Prime

To enable Dataflow Prime for a pipeline:

Enable the Cloud Autoscaling API.

Enable the API

Dataflow Prime uses the Cloud Autoscaling API to dynamically adjust memory.
Enable Prime in your pipeline options.

You can set the pipeline options either programmatically or by using the command line. For supported Apache Beam SDK versions, enable the following flag:

Java

--dataflowServiceOptions=enable_prime

Python

Apache Beam Python SDK version 2.29.0 or later:

--dataflow_service_options=enable_prime

Apache Beam Python SDK version 2.21.0 to 2.28.0:

--experiments=enable_prime

Go

--dataflow_service_options=enable_prime

Use Dataflow Prime with templates

If you're using Dataflow templates, you can choose to enable Dataflow Prime in one of the following ways:

For jobs launched from the Create job from template page:
1. Go to the Create job from template page.
  
  Go to Create job from template
2. In the Additional experiment field, enter enable_prime.
For jobs launched from a template through the command line interface, pass the --additional-experiments=enable_prime flag.
To enable Dataflow Prime when you create a template, set the --experiments=enable_prime flag.

Use Dataflow Prime in Apache Beam notebooks

If you're using an Apache Beam notebook, you can enable Dataflow Prime programmatically using PipelineOptions:

options = pipeline_options.PipelineOptions(
    flags=[],
    dataflow_service_options=['enable_prime'],
)

To learn more about setting Dataflow options in a notebook, see Launch Dataflow jobs from a pipeline created in your notebook.

Feature comparison between Dataflow and Dataflow Prime

The following table compares the available features for both variants of Dataflow.

Feature	Dataflow Prime	Dataflow
Runner v2	Default feature with no option to turn off for batch jobs and optional for streaming jobs	Default feature with an option to turn off for batch jobs and optional for streaming jobs
Dataflow Shuffle for batch jobs	Default feature with an option to turn off	Default feature with an option to turn off
Streaming Engine for streaming jobs	Optional feature for Java pipelines and always on for Python pipelines	Optional feature for Java pipelines and always on for Python pipelines since 2.45.0
Horizontal Autoscaling	Default feature with option to turn off	Default feature with option to turn off
Vertical Autoscaling	Default feature with option to turn off	Not applicable
Right fitting	Optional feature	Optional feature
Billing	Serverless billing	Standard billing

What's next

Read about Dataflow quotas.
Learn how to set pipeline options.
See available pipeline options for Java and Python pipelines.
Learn more about autotuning features for Dataflow Prime.
Learn more about Dataflow GPUs.