Optimising AWS to Azure Migration: How We Successfully Migrated our AI Product

Azure offers several AI tools and services under its Cognitive Services suite, each tailored for specific applications. For an AI product focused on intelligent document processing like ours, services like Azure AI Vision and Azure Form Recognizer are invaluable. These tools automate and enhance data extraction, classification, and verification processes. They also significantly improve efficiency and accuracy in managing financial documents such as invoices, delivery notes, and bank cheques.

The integration with OpenAI in Azure OpenAI Service provides REST API access to OpenAI's powerful language models, including GPT-4 and GPT-4o. This allows AI products to leverage the advanced capabilities of GPT-4 without needing to set up separate infrastructure or manage deployments.

The architecture of Azure Web App offers high flexibility and efficiency in resource consumption with Scale out and Scale up features, especially compared to AWS Lambda Functions, which have a memory limit of 10,240 MB. For AI products and services, this means more scalable and adaptable resource management, enabling complex and demanding applications.

‍

Additionally, Azure Web App Plans allow shared computing resources among multiple apps. They enhance cost management and make it easier to scale AI applications without incurring prohibitive costs.

Transitioning from fully serverless AWS Lambda Functions to server-based Azure Web App has been crucial for cost management. The approach is particularly beneficial when handling a high volume of document processing requests, as it allows for more manageable resource allocation. The server-based architecture in Azure ensures that AI projects can maintain high performance and scalability while optimising operational expenses.

Pedant was our first product on Azure and it required us to get familiar with a new cloud platform. We dedicated time to learning and adapting to Azure’s environment, ensuring a smooth transition and effective use of its services. Our efforts laid a solid foundation for successfully implementing Pedant and leveraging Azure's capabilities for optimal performance.

Azure offers several managed computing services, including Container Instance, Web App, Container App, Function, and AKS. Selecting the best option for each microservice in our project posed a significant challenge. To address this, we conducted thorough evaluations of each service and focused on their specific functionalities and capabilities.

Transitioning Pedant's AI section from serverless computing to server-based services was essential to achieve stability. This shift allowed us to overcome the capacity limitations inherent in serverless computing. It ensured that Pedant could scale seamlessly with growing demand. By adopting server-based services, we enhanced the platform's ability to handle intensive document processing tasks efficiently.

Initially, Azure Web App presented a challenge as it did not support running multiple containers simultaneously for deployment, especially when including sidecar containers. However, Azure introduced Docker Compose as a preview feature to support this function.

‍

Following extensive testing and quality assurance, we found it to be stable and well-suited for our production environment. This update resolved our initial deployment limitations and provided flexibility.

Azure Container App did not support exposing multiple ports for HTTP/gRPC from our containers, which posed a challenge for our application's functionality. However, a recent production release introduced this feature, effectively resolving the limitation. This update enhanced our app's capabilities by allowing us to manage multi-port exposures seamlessly.

Azure Container App offers a consumption plan based on request volume that proved cost-effective for our application. It effectively aligns our operational costs with actual usage demands. This approach minimised overhead expenses while ensuring optimal resource utilisation, helping with project efficiency and budget management.

Azure Container App also supports stickiness for WebSocket connections and features a scale to zero capability. It optimises resource usage by scaling down to zero instances during periods of inactivity.

Embracing MLOps in Azure Machine Learning has streamlined our model training and deployment processes. MLOps tools in Azure Machine Learning ensure smoother operations and more effective deployment of AI solutions, driving overall product success and innovation.

Maximise cost efficiency by aligning resource utilisation with the specific needs of your product at each development stage. For instance, to cut expenses, we utilised Azure Functions with zip deployment instead of a dockerized approach. This allowed us to take advantage of the 'pay as you go' pricing model instead of the more costly Premium tier.

Additionally, employ a strategic migration assessment tool to evaluate your readiness for migrating an AI product. This ensures smooth operation post-migration and helps you achieve your goals. Keep stakeholders informed and involved early in the process to secure their support and facilitate a seamless transition.

Implement CI/CD tools that integrate seamlessly with your existing repositories. Given our repositories were already on GitLab, we opted for GitLab CI to facilitate rapid deployment. To address the disk space issue, we implemented a GitLab Custom Runner on an Azure VM, which provided the necessary capacity for building our AI images.

By Dockerizing all AI services and abandoning all AWS SDKs and Lambda's libraries, organisations undoubtedly expedite the migration process to Azure effectively.

Develop a data migration carefully to ensure compatibility. The command-line utility AzCopy can be used to copy unstructured files such as PDFs and images to and from a storage account. It is an invaluable tool for seamless data movement from an EC2 Jumpbox to Azure Storage Account Containers.

Containerising your AI applications with Docker and deploying them on Azure Kubernetes Service (AKS) can make your applications more scalable, portable, and manageable. It helps streamline deployment and ensures your AI solution runs smoothly across different environments.

Migrating an AI product from AWS to Azure involves challenges such as service compatibility, data transfer complexities, adapting to Azure tools, integrating existing infrastructure, and cost management through continuous monitoring and adjustment.

Switching from AWS to Azure requires adapting to Azure-specific tools, services, and infrastructure, which may be challenging initially. However, with thorough planning and understanding of Azure's offerings, the transition can be managed effectively to align with organisational goals and requirements.