From virtual machine to Kubernetes to serverless: How dacadoo saved 78% on cloud costs and automated operations

dacadoo is a global Swiss-based technology company that develops solutions for digital health engagement and health risk quantification. Their products include a software-as-a-service (SaaS)-based digital health engagement platform that uses behavioral science, AI, and gamification to help end users improve their health outcomes.

The company embarked on a journey to modernize an API to quantify health and lifestyle data plus a risk engine to calculate mortality and morbidity probabilities based on years of scientific research data.

To transform a virtual machine–based API service into a globally redundant, scalable health score and risk calculation solution dacadoo chose Amazon Web Services (AWS) technology. The service handles highly sensitive health data from a global customer base and must comply with regional regulations.

The result is a cost reduction of 78% and an infrastructure maintenance effort of less than an hour per year , allowing dacadoo to deliver and operate more AWS infrastructure without scaling its site reliability engineering (SRE) team, thanks to a high level of automation and an agile mindset.

In this post, we walk you step-by-step through dacadoo’s journey of embracing managed services, highlighting their architectural decisions as we go.

Background

The solution architecture went through a three-stage journey:

1. Incubation – Single virtual machine on premises with disaster recovery (DR) in Switzerland
2. Global and scalable – Multiple global Kubernetes clusters
3. Operational excellence – Fully serverless and geo-redundant on AWS

Stage 1: Incubation with a virtual machine

After years of scientific research and development, the service was launched, running on a single on-premises virtual machine that used hypervisor technology to provide disaster recovery (DR). However, it had no high availability (HA) capability and it required manual recovery.

The application serving the API requests and the NoSQL database were both running on the same host. Software deployment and operating system maintenance were performed manually using Secure Shell (SSH)—a typical low-automation setup that also included downtime.

The following architecture diagram shows a virtual machine encompassing the monolithic application and its database.

Challenges

A single virtual machine was quick to set up and inexpensive to operate, but it had considerable shortcomings. The health API was only available in Switzerland, infrastructure maintenance was performed manually, and software deployment was handled manually. Additionally, database backups were done using virtual machine snapshots, uptime monitoring only, and testing was conducted on the developer workstation.

Stage 2: Global and scalable with Kubernetes

At that time, dacadoo made a strategic decision to heavily invest in Kubernetes for managing containerized workloads on a global scale. As part of this technology rollout, the health score and risk service were migrated to Kubernetes.

Due to the geographically distributed customer base and low latency requirements, three Kubernetes clusters were deployed, one on each continent. The NoSQL database was hosted in proximity to the workload to reduce service latency and keep the migration effort low.

To reduce the operational maintenance, the NoSQL database was integrated as a SaaS offering, and monitoring was centralized using Datadog.

All cloud infrastructure was provisioned exclusively with Terraform, covering the Kubernetes cluster, NoSQL database , and integration with GitLab and Datadog.

dacadoo containerized the API service and used Gitlab continuous integration and continuous deployment (CI/CD) pipelines to deploy multiple environments and clusters on a global hyperscaler.

In retrospect, this was a typical replatform modernization project from virtual machine to Kubernetes, with a high level of automation and a SaaS-first approach.

The following diagram is the architecture for the container solution with managed NoSQL database.

Challenges

The service faced several challenges, including increased costs from deploying three regional Kubernetes clusters across three environments, resulting in 27 cluster nodes and additional expenses from managing NoSQL database SaaS instances for each cluster. The complexity of CI/CD pipelines for multi-environment multi-cluster deployments added to the difficulty. Significant operational effort was required to keep infrastructure and Kubernetes components up to date.

Stage 3: Operational excellence with serverless

The Kubernetes-based architecture met the requirements, but some features in the dacadoo API service backlog needed to fit better with the application architecture at the time.

This was the right moment to take a holistic view of the infrastructure and software architecture and refactor the solution according to the latest AWS technologies and best practices, the next frontier for dacadoo’s engineering team.

Solution requirements

Requirements for the solution refactoring were as follows:

• Keep the functionality of the API unmodified
• Constrain data processing to a region of choice for compliance with local data protection laws
• Avoid weekly patch cycles by exclusively using managed serverless services
• Reduce costs by choosing services with a pay-as-you-go billing model
• Delegate authentication to a dedicated service
• Use an established web framework with an extensive ecosystem

Refactoring the apps

The API service has two components: a developer portal and the health score and risk calculations API. The database is only required for API keys, algorithm parameters, quotas, and usage statistics. Health data is processed regionally by the compute layer but not persisted, opening the door for a distributed database: Amazon DynamoDB global tables is the perfect fit for the solution. Writes are distributed to all connected Regions, whereas reads are local, providing low latency for complying with dacadoo service level agreements (SLAs).

The developer portal is a web UI with API documentation and API key management features. AWS Lambda is a great fit because it scales automatically and has a pay-per-request billing model.

The health and risk API uses algorithms implemented in the C programming language for short bursting, compute-intense simulations. These calls are wrapped by a REST API using the Python FastAPI framework. These characteristics make AWS Lambda a great fit.

Serverless architecture

HTTP requests are routed to the Lambda functions using Amazon API Gateway with AWS WAF for protection from malicious requests and attacks. Static assets are served from an Amazon Simple Storage Service (Amazon S3) bucket through API Gateway. The additional features of Amazon CloudFront aren’t required, and Amazon S3 reduces the complexity.

Amazon Route 53 provides a powerful feature known as latency-based routing, which allows it to direct DNS queries to the endpoint that offers the lowest latency for the requester.

This feature provides Regional high availability for API users without data processing location requirements. Alternatively, the user can call specific Regional endpoints to make sure requests are processed in the desired Region.

API authorization is HTTP header-based and is performed in the application with data stored in Amazon DynamoDB.

The following diagram is the architecture for a geo-redundant fully serverless solution.

With a dacadoo SRE team proficient in Python, they opted for Pulumi for its advanced features such as programming language flow control constructs, powerful configuration capabilities, and multi-cloud support.

For continuous integration, GitLab CI compiles the algorithm library, tests the FastAPI applications and packages everything. The application deployment is just an update of the AWS Lambda, a simple and reliable workflow.

Summary

The solution evolved from a managed infrastructure setup, where the customer held most of the responsibility, to an AWS managed service architecture.

Infrastructure provisioning evolved from manual, error-prone processes to powerful code-driven workflows in Pulumi. The SRE needed to enhance their software engineering skills to adopt Pulumi, transitioning from configuration-based approaches to designing and maintaining an infrastructure code base using object-oriented Python. This was part of dacadoo’s investment in the SRE team and broader modernization efforts. The serverless architecture enabled a GitOps engineering culture focused on productivity.

The transformation maximized scalability and availability while reducing costs and operational effort:

Virtual machine

• Scalability: Low
• Availability: Best effort
• Infrastructure costs: Low
• Maintenance effort: High

Kubernetes

• Scalability: High
• Availability: 99.95%
• Infrastructure costs: High
• Maintenance effort: Medium

Serverless

• Scalability: Very high
• Availability: 99.999% (with failover to another AWS Region)
• Infrastructure costs: Low
• Maintenance effort: Very low

The global redundancy elevates availability to an impressive 99.999% while keeping the costs low.

Conclusion

Migrating from a virtual machine to Kubernetes and ultimately to AWS Lambda demonstrates the progression of cloud engineering toward enhanced efficiency and scalability.

Each step in this journey reduced the complexity of managing resources while increasing flexibility and automation. Transitioning dacadoo’s API service to a fully serverless, geo-redundant architecture not only advanced the platform but also upskilled engineers, maintained a lean SRE team, and kept infrastructure costs low. Get started with your own AWS serverless solution.

About the Authors