Domain-driven event sourcing with Akka Typed
September 24, 2019 | Technical Blog
Welcome to the Bestmile Technical Blog. These articles are written by Bestmile software engineers, developers, researchers and other members of our technical team to share learnings with peers in their fields. We hope that by publishing this blog we are contributing to knowledge sharing in each of these domains. Please note that many of these articles are highly technical and are not meant to be accessible to a general audience.
When following a domain-driven design approach, it is considered good practice to have the domain code at the core of the application, self-contained and with no dependencies on platform-specific aspects (onion architecture). Domain code is as business-centric as possible, maximizing the ratio of business logic versus “plumbing” code. This makes business logic expressive and resilient to changes.
Although at first glance related to persistence and thus infrastructure aspects, adoption of event sourcing has a deep impact on domain expression. More than a particular technology, it is a design philosophy with an emphasis on immutable data and state evolution. In event sourcing, the state of entities is the consequence of a sequence of events applied to an initial state. Events are generated by commands sent to the entity, which represent actions in the system. This has the following implications on the “shape” of domain code:
This article introduces light abstractions that allow coding domain entities without any direct reference to Akka Typed. This makes it possible to structure the application with separate domain and infrastructure packages. Within infrastructure code, domain entities and repositories are implemented using Akka Persistence Typed together with Cluster Sharding Typed. Infrastructure boilerplate is minimized thanks to a couple of generic base classes that take advantage of the recently introduced persistent entity DSL. Much of the code presented here is inspired by this excellent article by Andrzej Ludwikowski.
As mentioned above, event sourcing revolves around three main concepts which allow describing operations for any domain with an emphasis on evolution:
Let’s go ahead and capture these notions in Scala:
A command is directed to a specific entity and thus bears a target entity ID. Along with command data, it defines the reply to the command emitter. Formulating the reply needs to be quick so that the system stays responsive. Reply formulation typically consists of validation code leading to either acceptation or rejection of the command. Purely introspective commands need to be supported in this model as well, allowing for state inspection. For such read-only commands, the command typically doesn’t entail any event and the reply consists of some element of the current entity state.
Two cases can arise when receiving a command: the entity already exists, or it’s the first time we’re getting a command for that ID.
Here’s a trait for entity commands:
The type parameters are:
Along with the entityID, we define a function initializedReply: S => R and uninitializedReply: R for reply formulation when the entity already exists or doesn’t respectively.
Commands generate events via a command processor (see below). As is the case with commands, events reference a certain entity and thus bear an entity identifier. We also include a timestamp in the definition below, to represent the intrinsic chronological ordering of events.
As mentioned above, command processing is about the creation of one or several events to carry out the command. We define a dedicated single-function trait for this aspect:
The type parameters are:
When the entity hasn’t been created yet (before the first command), the command processor has no entity state to work with. To allow for a clear distinction of this initial scenario, we define a slightly different trait, which makes no mention of entity state:
Similarly, we define traits for event handling. This is a function of two parameters, the current state and the event, leading to a new version of the state:
Quite intuitively, type parameters are:
In the same way as for command processing, there is always a first event for “bootstrapping” the entity. For this initial event, there is no state parameter as the entity doesn’t yet exist. We capture this with a dedicated trait as well:
Let’s now put these definitions in practice with an example entity themed in the mobility space:
Ride This entity represents an imaginary booking for a vehicle ride. After ride creation, a vehicle from the fleet must be assigned to carry the passenger. Once assigned and in position, the vehicle picks the passenger, at which point the ride starts. When the passenger is finally dropped off, the ride completes.
We start by defining the entity state with a plain case class (we are omitting some types for simplicity):
We can now define commands relevant to this example. Let’s start by refining the EntityCommand trait into RideCommand:
We fix the
S parameters while keeping the
R open to benefit from precise reply types, this will be helpful when wiring these commands with Akka Typed.
We do the same for events:
Let’s start with the command leading to ride creation, BookRide:
As can be seen above, possible replies are:
The generated event will be:
Let’s place corresponding command processing and event application logic directly within the entity companion object as implicit definitions (we will benefit from this when wiring up infrastructure code — more on than later):
Once the booking is created, we need to select a vehicle to service the ride. For the sake of simplicity in this article, we’re not delving into details, but one can imagine a sophisticated algorithm that would select the optimal vehicle. This algorithm could be monitoring RideBooked events from the Akka event journal to launch an optimization asynchronously, and send a command to the ride entity once a vehicle has been matched (note that event projections can also be abstracted in the domain, this could be the topic of a subsequent article). Let’s name this command AssignVehicle. Here’s a simple definition for it:
And the corresponding event :
The command processor creates the event. Since we reject commands whenever the vehicle has already been assigned, this case should not arise here so we log that as an error:
The event applier simply sets the vehicle field:
The two remaining commands in our simplified description of a ride lifecycle are
CompleteRide, which describe passenger pickup and dropoff respectively. The corresponding definitions are given below:
Notice above how we could recycle errors common to multiple commands by making them extend multiple reply types.
Extension of command processor and event handler is equally trivial:
In domain-driven design, a repository is an abstraction for a collection of persistent entities. It is the entry point to access and manipulate entities of a certain type. We can capture this with a simple trait, whose implementation will make use of our commands and replies together with the Akka mappings, as we’ll see next. This makes our ride operations easy to integrate in the rest of domain code, since it won’t have to deal directly with commands. Here’s an example definition for RideRepository, defined in tagless-final style:
We have kept this example very simple and directly transposed our command “language” into a set of functions, and kept the reply types. In a real case, we would typically transpose command reply types into some other types, typically distinguishing error from success more clearly with Either.
Thanks to abstraction and separation of concerns, our commands and events logic requires minimal test setup and the distinct behavioral aspects can be covered in isolation:
Let’s take a plunge now and dive into the infrastructure layer. This is where we’ll wire our abstractions with Akka. Thanks to the expressiveness of Akka Persistence Typed API, this mapping is surprisingly short, albeit a bit complicated on the typing side.
The bulk of the code is in an abstract PersistentEntity class, which exposes a eventSourcedEntity(id: String): EventSourcedBehavior public function. This will be used by the repository implementation as the actor behavior. This class is typed like so:
Here’s the full class definition:
Notice how we left
protected def configureEntityBehavior() open for extension, this allows refining the persistence behavior in subclasses.
Defining the persistent entity behavior is now easily done by extending
PersistentEntity. This is were everything comes together:
Definitions for implicit parameters initialProcessor, processor, initialApplier, applier are picked up automatically from Ride companion object, which is why we had published them in implicit scope earlier. The only significant logic here is the additional persistent behavior configuration in configureEntityBehavior. This lets us take advantage of tweaking options, e.g. to install an event adapter, define event tags, etc.
As mentioned earlier, the repository is implemented by sending commands and decoding replies. Here’s the trait definition:
Type parameters are:
The repository trait takes care of initializing sharding with persistent entity behavior. It requires the definition of the PersistentEntity instance, and exposes a def for looking up entities for a certain ID and sending them commands:
Implementation of this trait for Ride is where instantiation of RidePersistentEntity happens. Implementation of repository methods is simply about sending the relevant command:
Let’s make this a bit more concrete by previewing how the flow for the bookRide command will unfold when calling bookRide on the repository:
Note that command processing and event application proceeds strictly sequentially: the actor processes the command, stores the events, replies to the sender, applies the event leading to a new state, then processes the next command and so on. This makes for comprehensible state machine descriptions. Rather than leading to complicated program flow, asynchronicity is harnessed by the inherent distributive nature of domain entities.
This concludes our implementation tour. Event journal and adapter configuration aside, we now have a fully functional repository for rides!
Although very simplified, this example illustrates the “good fit” of the actor model to domain-driven-design: aggregates are represented by entities with well defined sequential state transitions and a command and event “language” to represent actions and facts. We have shown an approach to describe event-sourced entities in the domain using such an abstract language, and the required infrastructure plumbing code to map these pure definitions to an Akka Persistence Typed implementation.
Supporting code for this article can be found in its entirety here. We hope this was useful and would love your feedback! Feel free to reach out for more information.
∙ ∙ ∙
Jonas Chapuis and Michal Tomanski are both senior software engineers at Bestmile. They work on the core of Bestmile’s vehicle fleet orchestration platform enabling smart mobility.