Architecture

RnD Center for BFSI Sector, RISE Group, IIT Madras

Data Centric Architecture

Data is at the heart of any financial system and the underlying
data architecture allows data to be represented in multiple forms (structured and unstructured), thereby allowing any type of querying and information retrieval.

In the ever changing world of data analytics , where the nature of information changes rapidly, this is the only viable architecture approach.

Comprehensive Query/Visualization modules and Rule/ML engines are overlaid on the data store complex, allowing any arbitrarily complex business need to be met.

Key Aspects on architecture

The key aspect of the architecture is decoupling data flow and data transformation from compute.

Given the dynamic nature of analytics, the data store and the related query engine can vary from application to application

A canonical architecture is proposed to cater to a wide variety of applications.

- A collection of adaptors gather data from various sources at predefined intervals or based on event triggers. This decouples the rest of the system from various legacy and transactional systems

- The data lake/data staging area aggregates the output of this adaptor into a centralized data store. This store is based
on a late binding philosophy for maximum flexibility.

Specialized Stores

Specialized analytics requires data stores optimized for the specific types of queries envisaged for that application.

Stores may be column optimized, graph data optimized (either
in Neo4J/Dgraph or in tabular stores with graph query engines) or optimized for approximate queries (like BlinkDB).

The architecture allows unlimited number of such derived data stores to be created to cater to the specific needs of such applications.

So applications then become merely plugins in the compute layer that can use specific store to get optimized query performance.

For extreme performance, GPU based stores like MapD
and FPGA acceleration will be supported.