Architecture of Asset Compute Service

The Asset Compute Service is built on top of serverless Adobe I/O Runtime platform. It provides Adobe Sensei content services support for assets. The invoking client (only Experience Manager as a Cloud Service is supported) is provided with the Adobe Sensei-generated information that it sought for the asset. The information returned is in JSON format.

Asset Compute Service is extendable by creating custom applications based on Project Firefly. These custom applications are Project Firefly headless apps and do tasks such as add custom conversion tools or call external APIs to perform image operations.

Project Firefly is a framework to build and deploy custom web applications on Adobe I/O runtime. To create custom applications, the developers can leverage React Spectrum (Adobe’s UI toolkit), create microservices, create custom events, and orchestrate APIs. See documentation of Project Firefly.

The foundation on which the architecture is based includes:

  • The modularity of applications – only containing what is needed for a given task – allows to decouple applications from each other and keep them lightweight.

  • The serverless concept of Adobe I/O Runtime yields numerous benefits: asynchronous, highly scalable, isolated, job-based processing, which is a perfect fit for asset processing.

  • Binary cloud storage provides the necessary features for storing and accessing asset files and renditions individually, without requiring full access permissions to the storage, using pre-signed URL references. Transfer acceleration, CDN caching, and co-locating compute applications with cloud storage allow for optimal low latency content access. Both AWS and Azure clouds are supported.

Architecture of Asset Compute Service

Figure: Architecture of Asset Compute Service and how it integrates with Experience Manager, storage, and processing application.

The architecture consists of the following parts:

  • An API and orchestration layer receives requests (in JSON format) which instruct the service to transform a source asset into multiple renditions. These requests are asynchronous and return with an activation id (aka “job id”). Instructions are purely declarative, and for all standard processing work (e.g. thumbnail generation, text extraction) consumers only specify the desired result, but not the applications that handle certain renditions. Generic API features such as authentication, analytics, rate limiting, are handled using the Adobe API Gateway in front of the service and manages all requests going to I/O Runtime. The application routing is done dynamically by the orchestration layer. Custom application can be specified by clients for specific renditions and include custom parameters. Application execution can be fully parallelized as they are separate serverless functions in I/O Runtime.

  • Applications to process assets that specialize on certain types of file formats or target renditions. Conceptually, an application is like the Unix pipe concept: an input file gets transformed into one or more output files.

  • A common application library handles common tasks like downloading the source file, uploading the renditions, error reporting, event sending and monitoring . This is designed so that developing an application stays as simple as possible, following the serverless idea, and can be restricted to local filesystem interactions.

On this page