Patterns and implementations for a banking cloud transformation

Cosmos DB
Event Hubs
Kubernetes Service

This article covers the patterns and implementations the commercial software engineer (CSE) team used when they created the Banking system cloud transformation on Azure.


Saga architecture

Orchestration-based Saga on Serverless Architecture

Download a Visio file of this architecture.


Contoso Bank had an on-premises implementation of an orchestration-based saga. In their implementation, the orchestrator is a finite state machine (FSM). The CSE team identified the following challenges in the architecture design:

  • Implementation overhead and complexity on the stateful orchestrator to handle with states management, timeouts, and restarts in failure scenarios.

  • Observability mechanisms for tracking the saga workflow states per transaction request.

The proposed solution below is a saga pattern implementation through an orchestration approach using a serverless architecture on Azure. It addresses the challenges by using:

For more information, see Pattern: Saga on

Saga pattern

Saga is a pattern suitable for distributed transaction management, commonly applied to financial services. A new scenario has emerged where operations are distributed across applications and databases. In the new scenario, customers will need a new architecture and implementation design to ensure data consistency on financial transactions.

The traditional atomicity, consistency, isolation, and durability (ACID) properties approach is no longer suitable. It's because the data of operations are now spanned into isolated databases. Using a saga pattern addresses this challenge by coordinating a workflow through a message-driven sequence of local transactions to ensure data consistency.

KEDA architecture

EFT-Processor Autoscaling with KEDA Kafka topic trigger

Download a Visio file of this architecture.

For more information on KEDA scalers, see the following KEDA documents:

  • Azure Event Hubs Trigger: Compatibility for reading Azure blob storage URI for Java applications. It uses the Event Processor Host SDK, allowing the ability to scale Java consumers that read advanced message queuing protocols (AMQP) protocol messages from Event Hubs. Earlier the Event Hubs scaler worked only with Azure Functions.

  • Apache Kafka topic trigger: Support for SASL_SSL Plain authentication, allowing the ability to scale Java consumers that read Kafka protocol messages from Event Hubs.


  1. The CSE team deployed the application on the Azure Kubernetes Service (AKS) cluster. The solution needed to scale out the application automatically based on the incoming message count. The CSE team used a Kafka scaler to detect if the solution should activate or deactivate application deployment. The Kafka scaler also feeds custom metrics for a specific event source. The event source in this example is an Azure Event Hub.

  2. When the number of messages in the Azure Event Hub exceeds a threshold, KEDA triggers the pods to scale out, increasing the number of messages processed by the application. Automatic scale down of the pods occurs when the number of messages in the event source falls below the threshold value.

  3. The CSE team used the Apache Kafka topic trigger. It gives the solution the ability to scale the EFT Processor service if the process exceeded the maximum number of messages consumed under an interval.

KEDA with Java support

Kubernetes Event-driven Autoscaler (KEDA) determines how the solution should scale any container within Kubernetes. The decision is based on the number of events that it needs to process. KEDA, which has different kinds of scalers, supports multiple types of workloads, supports Azure Functions, and is vendor-agnostic. Go to Autoscaling Java applications with KEDA using Azure Event Hubs to explore a working sample.

Load testing architecture

Load Testing Pipeline with JMeter, ACI and Terraform

Download a Visio file of this architecture.

The solution provisions JMeter agents as Azure Container Instances (ACI) instances. It uses the remote testing approach. In the approach, a JMeter controller configures all workers using its own protocol and combines all load testing results. Finally, it generates the resulting artifacts like dashboard and logs.

The CSE team created a Python script to convert the JMeter test results format (.jtl file) to JUnit format (.xml file). The script allowed the integration of JMeter results with the Azure Pipelines test results.


The CSE team structured the load testing framework into two Azure Pipelines:

  1. A pipeline that builds a custom JMeter Docker container and pushes the image to Azure Container Registry (ACR). This structure brings flexibility for adding any JMeter plugin.

  2. A pipeline that validates the JMeter test definition (.jmx file), dynamically provisions the load testing infrastructure, runs the load test, publishes the test results and artifacts to Azure Pipelines, and destroys the infrastructure.

For more information about the load testing pipeline solution, see Implementation reference for JMeter load testing pipeline solution.

Load testing framework

The load testing framework used during the engagement is now open-sourced on GitHub. The framework is a flexible and scalable cloud load and stress testing pipeline solution. It uses Apache JMeter as the open-source load/performance tool and Terraform to dynamically provision and destroy the infrastructure on Azure.

The solution creates a great experience for developers and testers. Integrating and combining JMeter testing results on Azure Pipelines through test results and artifacts improves the experience. See Load Testing Pipeline with JMeter, ACI, and Terraform.

Scenario details

This scenario helps you better understand big-picture patterns and implementations in the banking industry, when moving to the cloud.

Next steps

Learn more about the component technologies:

Explore related architectures: