CQRS, using Clean Architecture, multiple databases and Eventual Consistency
I also keep more detailed information in my blog, and I going to release weekly to clarify my ideas:
- CQRS Translated to Clean Architecture
- CQRS Deep dive into Commands
- CQRS Queries and Materialization
- CQRS Consensus and Consistency
- CQRS Distributed chaos, CAP Theorem
You need some of the fallowing tools:
- Docker
- Visual Studio 2017
- .Net Core 2.1
Here's the basic architecture of this microservice template:
- Respecting policy rules, with dependencies always pointing inward
- Separation of technology details from the rest of the system
- SOLID
- Single responsibility of each layer
Segregation between Commands and Queries, with isolated databases and different models
Has direct access to business rules and is responsible for only writes in the application.
Below you can find a basic interaction between components in the Command Stack:
Responsible to provide data to consumers of your application, containing a simplified and more suitable model for reading, with calculated data, aggregated values and materialized structures.
The fallowing image contains the basic interaction between components in the Query Stack:
This example contains a simplified Domain Model, with entities, aggregate roots, value objects and events, which are essential to synchronize the writing with reading database.
The project contains a well-defined IoC structure that allow you unit test almost every part of this service template, besides technology dependencies.
Inside the main layers you going to find Interfaces which are essential for the application, but with their implementations inside their own layers, what allow Mocking, Stubbing, using test doubles.
There's a simple example using Mother Object Pattern and Builder to simplify unit tests and keep it maintainable and clean.
This microservice template comes with SRP and SOC in mind. Given the own nature of CQRS, you can easily scale this application tuning each stack separately.
Having multiple data stores makes this system a Derived Data system, which means, you never lose data, you can always rebuild one store from another, for example, if you lose an event which sync data between the write and read database you can always get this data back from the write database and rebuild the read store.
Domain Model is materialized to Query Models using view materializer. Keeping this as separed component in the query stack allows fully control to mapped properties and fully testable.
Given the physical isolation of data stores, Command Stack and Query Stack must communicate to synchronize data. This is done here using a Message Broker.
Every successful handled command creates an event, which is published into a Message Broker. A synchronization background process subscribes to those events and is responsible for updating the reading database.
Everything comes with some kind of down side. The case of CQRS with multiple databases, to maintain high availability and scalability we create inconsistencies between databases.
More specifically, replicating data between two databases creates an eventual consistency, which in a specific moment in time, given the replication lag they are different, although is a temporary state and it eventually resolves itself.
Here's a list of reliable information used to bring this project to life.