Foto av Håkan Dahl

SpringOne2GX - Eventsourcing and CQRS in a microservice context

// Håkan Dahl

At SpringOne2GX I saw Michael Ploed give a good talk on the subject “Building microservices with event sourcing and CQRS”. Here I’ll do a quick rundown of the two patterns (event sourcing and CQRS) and then some reflections related to microservices.

Event sourcing

The Event sourcing pattern is, like Michael defined it, all about: “Event sourcing is an architectural pattern in which the state of the application is being determined by a sequence of events”. There are some important implications of that:

  • The full history of events are stored and can be inspected or used for playback to restore system state to a given point in time.
    • Events are typically stored in an object/document format, as opposed to being de-composed into a relational model. Note: In a traditional relational model, keeping historical records often takes quite a bit of work and being able to see what the state was at a particular point in time is often not possible, historical records are often kept on a per-table basis (typical for audit-logging).
  • Events must be immutable and may never be deleted, or the system state can’t be re-built. Delete is to be implemented as an own event, that is appended to the event store.
  • Running queries for current state against the stack of events would typically perform bad, since the application would have to rebuild/maintain the current state from all stored events. Mitigating this problem is up to our next pattern.


The Command Query Responsibility Segregation (CQRS) is basically a pattern for separating access to datastore in an application into different services for read and write operations. The major points of this being:

  • possibility to scale read-parts separately from write-parts
  • possibility to choose an appropriate datastore technique for read vs write and optimize for each case, for example: for a read-heavy database you might want to have the equivalent of a hevily de-normalized database (although it doesn’t have to be a classic SQL-databse)

If we let write-operations append to the event store and then let those events propagate asynchronously to a read optimized datastore (propagation via a messaging paradigm) we can have a solution that is scalable and have some nice features as pointed out above. Notice that the async propagation of data from write-to-read access introduced eventual consistency into the solution, but we can’t have it all, remember the CAP-theorem.

The Microservices connection

Both event-sourcing and CQRS have been around for some time so that’s not really new. What’s interesting in the microservice context is that some event-handling will most likely be needed to synchronize state between different microservices in any reasonably complex landscape and these patterns can be a useful combination. In a more traditional monolithic solution the need for synchronization would often be less, typically due to access to a large data-model all at once, sometimes paired with a codebase with loose internal boundaries for data access. If we look at the success stories listed in the beginning of the Microservices presentation, both Karma and SoundCloud use event-handling as vital parts in their architecture, although not exactly as above.

Note: the video recorded talk will show up on later on, together with all the other talks from SpringOne2GX 2015.