What is Event Sourcing?
Event Sourcing is an architectural design pattern where changes that occur in a domain are immutably stored as events in an append-only log.
This provides a business with richer data as each change that occurs within the domain is stored as a sequence of events which can be replayed in the order they occurred. This means you’re able to see more than just the current state of your domain - you can see what lead up to the current state.
In addition, as events also contain the context of the change – the ‘what’, ‘when’, ‘why’ and ‘who’ - an event-sourced system has a wealth of information that can be incredibly valuable to the business.
Event Sourcing has a wide variety of applications across many sectors, including finance, logistics, healthcare, retail, government, transport, video game development and many more.
An Event Sourcing Example
Let's look at an order process as an example to further explain how Event Sourcing works:
A customer places an order and an invoice is raised with the order details. The current status of the domain is 'outstanding' with the amount the customer owes, in this case '$200'.
The customer receives the invoice and pays the bill, the current status is then updated to show the outstanding balance as zero. 
In an event sourced system, the status change would be captured as an event 'OrderPaymentReceived' and stored in the append-only log in the order in which it occurred.
Another example, is an order discount. Let's say the customer placed an order, then a discount was applied. In a system that only captures the current state, you'll see the 'Outstanding' amount changed from $200 to $150 and won't know why. In an event sourced system, the change is captured in the event 'DiscountApplied', giving the context of the change - in this case a discount was applied.
This is how the event 'DiscountApplied' would look when implemented in code using C#.
Events
An event represents a fact that took place in the domain. They are the source of truth; your current state is derived from the events. They are immutable, and represent the business facts.
Events are referred to in the past tense and represent the specific business fact. For example, ‘InvoiceIssued’ shows the state of the invoice has definitely changed, rather than just come into being. It’s also a better name than `InvoiceCreated` as it’s explicitly describing the fact using the business domain language. The exact definition of an event is going to depend on the business use case, and should reflect your business data.
An event, in terms of Event Sourcing, usually contains unique metadata such as the timestamp of the event, the unique identifier of the subject, etc. The data within the event will be used in the write model to populate the state and make decisions, as well as populate read models.
It’s this implicit information in the event name ‘InvoiceIssued’, along with the metadata and the immutable nature of the event store that makes it an excellent solution for extracting more useful, in-depth insights and context for the business.
Event-native databases (Event Store)
Each change that took place in the domain is recorded in the database. Event-native databases are natively focused on storing events. Usually, they do that by having the append-only log as the central point.
Event-native databases are a different kind of database from traditional databases (graph, document, relational etc). They are specifically designed to store the history of changes, the state is represented by the append-only log of events. The events are stored in chronological order, and new events are appended to the previous event.
The events are immutable: they cannot be changed. This well-known rule of event stores is often the first defining feature of event stores and Event Sourcing that most people hear, and is absolutely true, from a certain point of view.
The events in the log can’t be changed, but their effects can be altered by later events. For example, there may be an ‘InvoiceIssued’ event appended to the log one day, only for it to be announced that the address the invoice was issued to is incorrect. A new event can be added, with the ‘InvoiceVoided’ event, then another event with the corrected address and an ‘InvoiceIssued’ event. All three events are stored, all are still immutable, but the desired outcome is achieved: an invoice has been issued to the correct address. The events are immutable, but that does not mean that log cannot be changed.
This is an example of the business context being kept; every stage the invoice has gone through has been kept, along with the dates and times of all events. Obviously, this information would need to be kept for auditing purposes, but imagine applying this level of information to any and all parts of the business: having an audit-level amount of information for every event in the system. The event store doesn’t lose information, and with the right configuration, you can have useful reports and analysis of the data whenever you need it.
Streams
Within the event store, the events referring to a particular domain or domain object are stored in a stream. Event streams are the source of truth for the domain object and contain the full history of the changes. You can retrieve state by reading all the stream events and applying them one by one in the order of appearance.
A stream should have a unique identifier representing the specific object. Each event has its own unique position within a stream. This position is usually represented by a numeric, incremental value. This number can be used to define the order of the events while retrieving the state. It can be also used to detect concurrency issues.
Event stores are built to be able to store a huge number of events efficiently. You don’t need to be afraid of creating lots of streams, however, you should watch the number of events in those streams. Streams can be short-lived with lots of events, or long-lived with fewer events. Shorter-lived streams are helpful for maintenance and makes versioning easier.
Projections
In Event Sourcing, Projections (also known as View Models or Query Models) provide a view of the underlying event-based data model. Often they represent the logic of translating the source write model into the read model. They are used in both read models and write models.
Projections in the Read Model
A common scenario in this context is taking events created in the write model (e.g. InvoiceIssued, OrderPlaced, PaymentSubmitted, OrderItemAdded, InvoicePaid, OrderConfirmed) and calculating a read model view (e.g. an order summary containing the paid invoice number, outstanding invoices items, due date status, etc.). This type of object can be stored in a different database and used for queries.
A set of events can also be a starting point for generating another set of events. For example, you can take order events, calculate the pricing summary and generate a new order payment event and place it in another stream. This type of projection is also called transformation.
Projections in the Write Model
Another form of projection is called stream aggregation. It’s a process of building the current state of the write model from the stream events. During aggregation, events are applied one by one in order of appearance. The main purpose of a stream aggregation is to rebuild the current state to validate an executed command against it.
Projections should be treated as temporary and disposable. This is one of their key benefits as they can be destroyed, reimagined and recreated at will; they should not be considered the source of truth.
There is a lot of conceptual crossover between a projection and a read model, and this can lead to some confusion. The simplest way to understand the relationship between projections and read models is that a read model is made of multiple projections. The projections are the method of populating your read model, and represent discrete parts of the whole read model. For example, a projection can be used to create invoices, and another can be used as a financial report, both of which are part of the read model.
A common misconception in explaining Event Sourcing is conflating projections with the state. In Event Sourcing, the source of truth is the events, and the state of the application is derived from these events. Facts are stored in events; projections are an interpretation of that raw data
Subscriptions
One of the most significant advantages of Event Sourcing is observability. Each action in the system triggers an event, and the event gathers business information about the fact of the action’s result. This is a simple but powerful feature, as it allows the building complex business workflows and splitting the work down into smaller chunks.
In Event-Driven Architecture, the responsibilities of services are inverted. The services are decoupled from each other due to the publish/subscribe approach. For example, a reservation service doesn't have to call the invoice service directly: it publishes an event about reservation confirmation, the invoice service subscribes to the reservation event stream, and can issue an invoice right after the notification.
Event stores (including EventStoreDB) enable that through the subscriptions functionality. It’s a similar concept to the traditional ‘Change Data Capture’ in relational databases. Each event stored in the database can trigger a notification, and subscribers can listen to those notifications and perform follow-up actions, such as:
- Run a projection to update the read model,
- Perform the next step of the business process,
- Forward the event to the queuing/streaming system to notify external services.
Usually, you can subscribe to either all events (directly from the append-only log) or partitioned/filtered events data (e.g. events from the specific stream, stream type, event type, etc.).
It’s important to remember that even though event stores can provide publish/subscribe functionality, they’re usually optimised for storage, not transport. For higher throughput needs or cross-service communication, it’s worth considering using additional specialised streaming solutions, such as Kafka or Kinesis.
Write model
The write model is responsible for handling the business logic. If you’re using CQRS, then this is the place where commands are handled. The write model has two aspects: physical and logical. The physical aspect relates to how and where data is stored, and the logical aspect reflects the business domain recreated in code structures, and can also be referred to as the Domain Model. Contrary to the traditional anaemic model, it not only contains the data but also the methods to change the data and validate it. The logic within the write model should be as close as possible to the business processes; you should be able to read the code and fully understand the business requirements from the code.
Let’s take an example: a write model for issuing invoices. It has data, e.g. amounts, vendor names, invoice numbers, and methods like “Issue Invoice”. It has rules such as “each invoice number must be unique”, “each invoice must contain a vendor name” etc. By having the design of the system derived from the domain, we can keep all the business logic and data in one place. A write model is essential, but a read model is not always needed.
A useful pattern to consider while implementing the write model is the aggregate. It is responsible for maintaining data consistency, and by using it we’re making sure that all related data will be stored in a single, atomic transaction. Aggregates are not necessary, but they are very common.
In Domain Driven Design, Eric Evans discusses the granular nature of objects, and provides a definition of an aggregate. From the Domain Driven Design “blue” book:
An aggregate is a cluster of associated objects that we treat as a unit for the purpose of data changes. Each aggregate has a root and a boundary. The boundary defines what is inside the aggregate. The root is a single, specific entity contained in the aggregate. The root is the only member of the aggregate that outside objects are allowed to hold reference to, although objects within the boundary may hold references to each other. Entities other than the root have local identity, but that identity needs to be distinguishable only within the aggregate, because no outside object can ever see it out of the context of the root entity.
This can be applied to the invoice example. An invoice is a domain object, but it is also made of several objects: the amount to be paid, the vendor name, the due date, etc. each one is part of the invoice, and the invoice joins them all together. So in the invoice example, the invoice is the aggregate. Each part of the data needed for the invoice (e.g. issuer information) is an entity. The root needs to be a single, specific entity in the aggregate, so this will be the unique invoice number.
Aggregates are the consistency guards. They take the current state, verify the business rules for the particular operation (e.g. if there is no invoice with the same number) and apply the business logic that returns the new state. The important part of this process is storing all or nothing. All aggregated data needs to be saved successfully. If one rule or operation fails then the whole state change is rejected.
In Event Sourcing, each operation made on the aggregate should result with the new event. For example, issuing the invoice should result in the InvoiceIssued event. Once the event has happened, it is recorded in the event store.
It’s easy to assume that because there are aggregates, you can have one mega aggregate. However, this will cause more problems in the long run. Smaller, more concise aggregates that focus on one aspect will make a more efficient system overall and preserve the business context where needed. As an example, consider the invoice aggregate. A VAT validation process would not need all the information contained in the invoice, so this process could be a smaller, more concise aggregate. This preserves the business context where it’s needed.
Read model
Queries in CQRS represent the intention to get data. A read model contains specific information. For example, the read model contains all the invoice information (the amount, the due date, etc) and the query represents the intention to know something about the invoice, such as whether or not it has been paid.
The read model can be, but doesn’t have to be, derived from the write model. It’s a transformation of the results of the business operation into a readable form.
As stated in the Projections section, read models are created by projections. Events appended in the event store triggers the projection logic that creates or updates the read model.
The read model is specialised for queries. It’s modelled to be a straightforward output that can be digested by various readers. It can be presented in many different ways, including a display on a monitor, an API, another dependent service or even to create the PDF of the invoice. In short, a read model is a general concept not tied to any type of storage.
However, the read model does not have to be permanent: new read models can be introduced into the system without impacting the existing ones. Read models can be deleted and recreated without losing the business logic or information, as this is stored in the write model.
Event Sourcing Architecture
Event Sourcing is an architectural design pattern that stores data in an append-only log. It is part of a wider ecosystem of design patterns that work together in various ways to allow developers to create the most effective architecture for their needs.
By deploying an event store as a design pattern within a wider architecture, it allows for the inclusion of other design patterns in the system that are the most suitable for the needs of the domain. For example, an event-sourced system works well within a CQRS architecture, however it's not necessary to use them together. Event Sourcing can be deployed alongside event driven architecture, or in conjunction with CQRS, an aggregate pattern, a command handler pattern, or one of many other patterns. Each individual pattern is a useful tool on its own, that can work in concert with other related patterns to make stronger, more specific patterns for your specific use case.
An event store can be a key element of a system, and that system can be as simple or as complex as the business domain requires it to be. It's useful to consider putting an event-sourced system in a part of the architecture that requires the preservation of context for all events, as this is where Event Sourcing is most effective.
CQRS
CQRS is a development of CQS, which is an architectural pattern, and the acronym stands for Command Query Separation. CQS is the core concept that defines two types of operations handled in a system: a command that executes a task, a query that returns information, and there should never be one function to do both of these jobs. The term was created by Bertrand Meyer in his book ‘Object-oriented Software Construction’ (1988, Prentice Hall).
CQRS is another architectural pattern acronym, standing for Command Query Responsibility Segregation. It divides a system’s actions into commands and queries. The two are similar, but not the same. Oskar Dudycz describes the difference as “CQRS can be interpreted at a higher level, more general than CQS; at an architectural level. CQS defines the general principle of behaviour. CQRS speaks more specifically about implementation”.
CQRS was defined by Greg Young, and he described it as:
Simply the creation of two objects where there was previously only one. The separation occurs based upon whether the methods are a command or a query (the same definition that is used by Meyer in Command and Query Separation, a command is any method that mutates state and a query is any method that returns a value).
Using CQRS, you should have a strict separation between the write model and the read model. Those two models should be processed by separate objects and not be conceptually linked together. Those objects are not physical storage structures but are, for example command handlers and query handlers. They’re not related to where and how the data will be stored. They’re connected to the processing behaviour. Command handlers are responsible for handling commands, mutating state or doing other side effects. Query handlers are responsible for returning the result of the requested query.
One of the common misconceptions about CQRS is that the commands and queries should be run on separate databases. This isn’t necessarily true; only that the behaviours and responsibilities for both should be separated. This can be within the code, within the structure of the database, or (if the situation calls for it), different databases.
Expanding on that concept, CQRS doesn’t even have to use a database: It could be run off an Excel spreadsheet, or anything else containing data. Oskar Dudycz described this in his article:
Nothing then stops you from using a relational database, even the same table for reading and writing. If you prefer, you can continue to use an ORM. Nothing prevents the command from adding/modifying records and query retrieving data from the same table. As long as we keep them separated in our architecture.
Event sourcing and CQRS seems like a new and interesting design patterns, but they have been around for some time. CQRS can feel like a strange new technique, especially after spending years learning with a foundation based in CRUD. CQRS was described by Greg Young as a ‘stepping stone towards Event Sourcing’. It’s important to understand event sourcing and CQRS do not rely on each other; you can have an event sourced system without CQRS and you can use CQRS without event sourcing. It’s just that the two work best together.
