Application modernization for massively parallel applications

  Part 6 of this article on Application Modernization covered the migration process. This section focuses on specialty applications.

Every organization must determine its own roadmap to application modernization. Fortunately, patterns and best practices continue to help and provide guidance. This section describes the application modernization for a representative case study for those applications that do not conform to cookie cutter web applications.

When we take a specialty application that involves massively parallel compute intensive applications that provide predictions, the initial approach is one of treating the model as a black box and working around the dependencies it has. But the modernization effort does not remain constrained by the technology stack that are the dependencies of the model. Instead, this is an opportunity to refine the algorithm and describe it with a class and an interface that lends itself to isolation and testing. This has the added benefit of providing testability beyond those that were available until now. The algorithm can also be implemented with design patterns like the Bridge design pattern so that the abstraction and the implementation can vary or the Strategy design pattern which facilitates a family of algorithms that are interchangeable.

Microservices developed in an agile manner with Continuous Integration and Continuous Deployment pipeline provide an unprecedented opportunity to compare algorithms and fine tune them in a standalone manner where the investments for infrastructure and data preparation need to be considered only once.

Algorithms for massively parallel systems often involve some variation of batched map-reduce summation forms or a continuous one-by-one record processing streaming form. In either of these cases, the stateless form of the microservice demonstrates superior scalability and reduced execution time than other conventional forms of software applications. The leap between microservice to serverless computing is one that can be taken for lightweight processing or where the model has already been trained so that it can be hosted with little resources.

Parallel computing works on immense size of data and the considerations for data modernization continue to apply independent of the application modernization.