Roboflow Batch Processing¶

Roboflow Batch Processing is a fully managed solution powered by Workflows that allows you to process large volumes of videos and images without writing code. It offers an easy-to-use UI for quick tasks and a comprehensive API for automating data processing—fitting both small and large workloads.

With configurable processing workflows, real-time monitoring, and event-based notifications, Roboflow Batch Processing helps you efficiently manage data processing, track progress, and integrate with other systems—making it easy to achieve your goals.

Getting started¶

To get started with Roboflow Batch Processing, first build and test your Workflow. Once it's ready, select the data you want to process and upload it using the UI or CLI tool. Then, initiate the processing and let Batch Processing handle the rest. Once the job is completed, retrieve the results and use them as needed. That’s it — you no longer need to write code or run processing on your machine.

UI Interface¶

The Roboflow platform provides a UI interface to interact with Roboflow Batch Processing, making it easy and accessible to try out the feature or process a small to moderate amount of data.

When creating a job, you can choose between image and video processing, select the appropriate Workflow, and adjust settings to fine-tune the system’s behavior. Key options include:

• Machine type – Choose between GPU and CPU based on processing needs. For Workflows using multiple or large models, a GPU is recommended, while smaller models or external API-based tasks can run efficiently on a CPU.

• Predictions visualization & video outputs – Enable this option if you want to generate and retain visual outputs of your Workflow.

• Video frame rate sub-sampling – Skip frames for faster, more cost-effective video processing.

• Maximum job runtime – Set a limit to help control costs and prevent excessive processing times.

CLI¶

By installing inference-cli you gain access to the inference rf-cloud command, which allows you to interact with managed components of the Roboflow Platform — including batch-processing and data-staging, the core components of the Roboflow Batch Processing offering.

The typical flow of interaction with the CLI is as follows:

First, ingest data into the platform. For images, use the following command:

inference data-staging create-batch-of-images --images-dir <your-images-dir-path> --batch-id <your-batch-id>

for videos:

inference data-staging create-batch-of-videos --videos-dir <your-images-dir-path> --batch-id <your-batch-id>

Then, you can inspect the details of staged batch of data:

inference rf-cloud data-staging show-batch-details --batch-id <your-batch-id>

Once the data is ingested - you can trigger batch job.

For images, use the following command:

inference rf-cloud batch-processing process-images-with-workflow --workflow-id <workflow-id> --batch-id <batch-id>

For videos:

inference rf-cloud batch-processing process-videos-with-workflow --workflow-id <workflow-id> --batch-id <batch-id>

Command will display the ID of the job, which can be used to check the job status:

inference rf-cloud batch-processing show-job-details --job-id <your-job-id>

This allows you to track the progress of your job. Additionally, the command will provide the ID of the output batch, which can be used to export results.

inference rf-cloud data-staging export-batch --target-dir <dir-to-export-result> --batch-id <output-batch-of-a-job>

That's it - you should be able to see your processing results now.

inference rf-cloud --help
inference rf-cloud data-staging --help 
inference rf-cloud batch-processing --help 

Service Pricing¶

The service charges usage based on the runtime of the underlying compute machines, starting at 4 credits per GPU hour and 1 credit per CPU hour. You can find the specific rates for your workspace on our pricing page.

We cannot provide an exact cost estimate for processing 1,000 images or 1 hour of video, as this depends entirely on the complexity of the chosen Workflow. However, we offer benchmark results to help you better understand potential costs.

Known limitations¶

• Batch Processing service cannot run Custom Python blocks.

• Certain Workflow blocks requiring access to env variables and local storage (like File Sink and Environment Secret Store) are blacklisted and will not execute.

• Service only works with Workflows that define singe input image parameter.

Technical Details of Batch Processing¶

• Data is stored in the data-staging component of the system — both your input images / videos and the processing results. The expiry time for any piece of data submitted to data staging is 7 days.

• When you upload data, you create a data batch in data staging. Batch processing jobs accept input batches marked for processing (each batch is processed by a single job).

• A single batch processing job contains multiple stages (typically processing and export). Each stage creates an output batch that you can retrieve later. We advise using export stage outputs, as they are optimized for network transfer (content is compressed / packed into an archive).

• A running job in the processing tage can be aborted using both the UI and CLI.

• An aborted or failed job can be restarted using the mentioned tools.

• The service automatically shards the data and processes it in parallel under the hood.

• Parallelism is applied at two levels:

  • The service automatically adjusts the number of machines running the job based on data volume, ensuring sufficient throughput (values reaching 500k–1M images per hour should be achievable for certain workloads).

  • A single machine runs multiple workers processing tasks (chunks of data) that belong to the job. This option can be configured by the user and should be adjusted to balance processing speed and costs.