[DL.ADS.6] Use cell-based architectures for granular deployment and release
Category: OPTIONAL
A cell-based architecture segments a larger system into isolated, independently functioning replicas, or cells. These cells are smaller components of the system that contain all application logic and storage. They have their own monitoring and alerting systems, are automated for creation and update, and can be managed and scaled individually. This approach offers advantages including scalability, fault isolation, testing, and operational resilience.
A cell-based architecture is a natural fit for DevOps as it enables small, frequent changes, reduces the risk from problematic deployments, and enables rapid recovery. It allows teams to deliver incremental updates to individual cells without risking the entire system's stability.
Start by defining your cells, each of which should be a complete, independently deployable unit of your system. You should limit the maximum size of a cell and maintain this consistency across different regions or installations. You then need to establish a routing layer that redirects client requests to the appropriate cell. You can store the routing information, such as user-to-cell mapping, in a low-latency database. Every cell should have its own monitoring and alerting system.
You will need to automate the lifecycle of your cells, including initial deployment and subsequent updates. A canary cell can be helpful in initial deployment of updates and assessing their impact. Ensure that you implement a central dashboard to provide an aggregated view of the state of your cells, enabling easy system-wide monitoring. Stream changes to a central data lake for centralized querying and analysis of changes across all cells. Finally, implement an operational tool to move users between cells and create new cells as needed. This step optimizes load distribution across cells by updating the user-to-cell assignment.
Cell-based architectures are optional. While beneficial for complex systems, smaller systems might not require such architectural complexity.
Related information:
