Reducing Processing Time by 11x and Cost by 90% Using AWS Step Functions at Vertex Pharmaceuticals

Vertex scientists use ML to quicken the time-consuming, meticulous processes that are involved in analyzing microscope images of biologic samples from experiments. After designing an experiment, they manually trained the base ML model about what they hoped to identify from
the images, labeling each area of interest by hand. After progressively fine-tuning the ML model on several images, the scientists ran the model across the rest of the images as a batch.

In the legacy system architecture, all scientists shared a single pool of compute resources. They sometimes had to wait up to 4 hours for results as their images queued for processing. “The image segmentation workload is incredibly spiky, and we continuously had to scale our compute fleet to match peak demand across scientific groups,” says Roberto Iturralde, senior director of software engineering at Vertex. “Under heavy load, performance degraded; under low load, we were overprovisioned.” To accommodate unpredictable demand, Vertex had to keep its servers running 24-7, accumulating cost.

Working alongside AWS, Vertex redesigned its system in April 2023 using AWS Step Functions Distributed Map and AWS Lambda, a serverless, event-driven compute service that lets engineers run code for virtually any type of application or backend service without provisioning or managing servers. Vertex uses infrastructure as code to model and deploy its system through its continuous integration and deployment pipelines so that Vertex engineers can safely and quickly deploy changes to the image segmentation application.

As an early adopter of AWS Step Functions Distributed Map, Vertex can run up to 10,000 parallel workflows. It can also iterate from millions of items in various formats that are stored in Amazon Simple Storage Service (Amazon S3), an object storage service offering scalability, data availability, security, and performance. “Since the new system went into production, we have hardly had to think about it,” says Iturralde. “Because it’s serverless, it just runs.”

Scientists interact with a web application running on AWS Fargate, a serverless, pay-as-you-go compute engine that lets engineers focus on building applications without managing servers. Using the web application, scientists fine-tune the parameters of the ML model that segments the image and test the model on individual images. Using AWS serverless technologies, the subsequent batch of images now process automatically. The automated process begins with a call to Amazon API Gateway, a fully managed service that makes it simple for developers to create, publish, maintain, monitor, and secure APIs at nearly any scale. The AWS Step Functions
workflow uses the input from Amazon API Gateway to determine which images to process and which parameters to pass back to the ML model. Using AWS Step Functions Distributed Map, the system processes all the images in parallel. Finally, the system automatically writes the output to Amazon S3, recording the metadata in a database for review.

Using AWS Step Functions, Vertex accelerated performance by 11 times because workflows now take minutes instead of hours. “Initially, some of our scientists questioned if the system was working, because it was going so fast compared with the old one,” says Iturralde.

Using the scalability of AWS Step Functions Distributed Map and AWS Lambda, dozens of scientists can work simultaneously and the system runs across multiple AWS Availability Zones. “Using AWS Step Functions, we greatly improved the robustness and reliability of our image segmentation solution,” says Karthik Ghantasala, director of cloud architecture and strategy at Vertex. “Our scientists analyze their experiments quickly and cost effectively.” Because the
serverless system runs only when it’s in use, Vertex has cut compute costs by 90 percent.

Vertex engineers have improved observability as thousands of images are processed across multiple experiments. “Using AWS Step Functions, we can pinpoint where exactly an error has caused a workflow to pause,” says Ghantasala. “We have also reduced operational overhead.” AWS Step Functions features native automatic retry and error-handling functionality, restarting from the failed step rather than from the beginning, which saves additional time and cost. To handle the observability, the team incorporates AWS monitoring tools such as Amazon CloudWatch, a service that monitors applications, responds to performance changes, optimizes resource use, and provides insights into operational health.