DEV Community

Cover image for Developer deprogramming: Getting started in Event Sourcing
Barry O Sullivan
Barry O Sullivan

Posted on

Developer deprogramming: Getting started in Event Sourcing

There are two things I wish I knew when I started building Event Sourced Apps.

  1. Always talk to the domain expert before building or designing anything, no matter how "simple"
  2. Always start with events, don't start with a UI, no matter how "simple" it is

When we ignore the above and just forge ahead, we invariably end up with CRUD based events. Nine times out of ten these events will not mirror the actual business processes, and instead they will get in the way and make development harder and slower.

The problem with CRUD

CRUD stands for Create/Read/Update/Delete. Most Databases operate on a CRUD model, like MySQL or MongoDB. You Read a model from the DB, change a value, and then Update the record in the DB. This is all fairly straight forward.

CRUD is how we as developers have trained ourselves to think. As part of our day to day, we translate specs into tables, then model all our operations on that data as CRUD. Hell, most of our UIs are expressed as CRUD based forms, so it's no wonder we're obsessed with the concept.

The thing is, the domain experts don't speak in CRUD terms, that's our language, not theirs. Whenever we build a solution without talking to a domain expert we fill the knowledge gaps ourselves, using the language we know best. We end up modelling our events as CRUD operations, even if it in no way matches the language or the processes of the business itself. It turns out this causes massive headaches further down the line, as the following story will showcase.

An example of getting it wrong

This story is drawn from past experience, some of the details are changed, but the core message is the same.

We were a small team building an online training product. We had multiple independent services, one of which had a concept called a Course.

    title: Trust Building Tutorial

A Course is a training video on a certain topic. Conceptually it's a pretty simple object, it has a title, an image and content. These properties were simple as well, title is a non-blank string, and both image and content are URLs to files uploaded through the GUI. So we decided to make Course an aggregate, it had a simple lifecycle and was self contained. All good.

First we started with UI mockups and broke down their expected behaviour.

Add course form

We saw that a user can Create, Update and Delete a course (you can see the CRUD language bleeding in already). Thus we came up with the following commands and events.


  • CreateCourse
  • UpdateCourse
  • DeleteCourse


  • CourseCreated
  • CourseUpdated
  • CourseDeleted

What's wrong with this?

This all seems pretty straight forward, but look closer and you'll see the cracks in this design.

1. Not enough context

First off, it doesn't tell us anything interesting. A course was updated, sure, but what was actually updated? Did the title change, did the image? Any service that cares about the title changing has to keep track of that title historically in order to check if it changed. That's a lot of work for such a simple task.

2. Expensive to change

The above events are brittle. If the "Course" changes shape, eg. a description is added, then you have to upgrade every "CourseCreated" and "CourseUpdated" events and update every single service that consumes that event. That's a lot of extra work just to add a description to a course.

3. Inconsistent language

When we listened to the domain experts, we noticed that not a single person said that they had "Created a course". Instead they said they'd "Added a course". It's a subtle change, but it's clear sign that we had made up our own language, instead of using theirs. Now the dev and business teams would use different language to describe the same things, which leads to confusion and unnecessary arguments.*

Why did this happen?

We, like most developers getting started in Event Sourcing, we were still thinking in terms of CRUD. Based on the UI, we knew the type of commands and the data they needed, so we thought we had enough detail to start implementing. When it came time to actually create the events, we used our commands as a guide, converting them to past tense and calling those the events. A one to one mapping that we had completely made up without talking to a domain expert.

In short, we started with a UI, when we should have started with domain events. No wonder it was wrong. Oops.

How should the events actually look?

Well in order to answer this, we had to actually talk to a domain expert and figure out what they thought was important. While doing this, we needed to disregard what we had already modelled, as it was based on our own language, not the businesses.

So we interviewed them and asked "What do you do when administrating a course?", what's important to them. Then we flipped the question and asked them "What's important to others?", who do they tell about certain types of changes?

This was an eye opening exercise. The expert told us that whenever a course was added or changed, they told the marketing and transcoding teams. Delving deeper, we asked what they actually said to those teams, and when.

It turns out that the marketing team needed to know the current title and image for the course, or if the course was removed, they didn't care about anything else.

Conversely, the transcoding team cared about the content and nothing else.

Also, in their language, there was no distinction between creating a title/image/content and updating it, we had made that distinction up ourselves (based on our own internalisation of CRUD concepts).

This lead to the following events and commands.


  • CourseAdded
  • CourseTitleSet
  • CourseImageSet
  • CourseContentSet
  • CourseRemoved


  • AddCourse
  • ChangeCourse
  • RemoveCourse

Why is this better?

Now let's see how these events stack up compared to the old ones.

1. Lots of context

With the new events the marketing and transcoding teams/services just have to listen for the events they care about. They no longer have to sift through the events to figure out if the change is important to them, making the overall system simpler. This is an event driven example of the Tell, don't ask principle.

2. Easy to change

Taking the above hypothetical situation, we want to add a description to the course. Instead of updating the existing events, we'll model this event as a new, discreet business change. So we'll simply add a new event called CourseDescriptionSet. This change is relatively easy to make and it only affects the aggregate, no event upgraders needed. The change doesn't affect any of the external services either, they can just keep consuming the existing Course events. Thus our system is less brittle to change. This is actually the Open/Closed Principle in action.

3. Consistent language

The developer and the domain expert are speaking the same language, at the same granularity. This shared understanding is explicitly modelled in the code, making it easier to understand. This will make future interactions simpler and minimise miscommunication.


After a little bit of exploration, we managed to make the Course aggregate much more useful to our system as a whole, exposing richer functionality than the basic CRUD implementation.

In short, talk to the domain expert. It doesn't matter if you're building an event sourced app or not (though it really helps), talk to them. This will will let you move past basic CRUD implementations, so you can build useful, understandable software. Remember, a well fleshed out UI with sequence diagrams is no substitute for a real conversation with an expert. So don't assume you understand the problem, stay humble and ask questions, it will save you an incredible amount of time further down the line.

An aside on talking to Domain Experts

It's important to note that we had to delve deeper into the domain expert's answers to really understand what events were important. People intuitively condense processes when explaining them to others, sometimes omitting information to keep it simple. This is why you need to explore their answers in detail to extract hidden domain concepts. Basically, keep interrogating them until they give consistent answers (do it nicely though).

If you'd like to hear more about interviewing domain experts, please leave a comment, always happy to talk. If enough people are interested, I'll write an article about it (crowd sourced articles!).

Top comments (11)

kspeakman profile image
Kasey Speakman • Edited

I really like your example. This resonates with my experience as well.

I have observed at least 2 main reasons for data that is traditionally CRUD.

  1. For humans
    • codes/names for identification - Because humans more easily recognize words than arbitrary numbers (IDs).
    • descriptions/notes/status text for understanding context and history
  2. For software, configuration affecting run-time operation - E.g. One course requires a grade of 70 to pass. Another course, 60.

These are really two orthogonal dimensions of a given concept, but are typically packaged together for convenience.

Side tangent: I find that almost invariably, the human side of the data requires some manner of set-based constraints... the kind typically found in relational databases. Example: Humans gets really confused if there are 2 MATH101 in the course catalog. Or if PASCAL100 (a deactivated course) shows up in the catalog. So some parts of the data must be indexed for human consumption. While practicing event sourcing, I've often thought I could use a "database" product that allows me to directly update different indexes without requiring them to be attached to a table. I guess it'd be a key-value store but one that allowed me to choose the key index algorithm (hash, b+, reverse, etc) on each collection. If you know of a database product that does this, let me know. For now, I use relational tables and attach indexes to them. It's less efficient (2 copies of key data, 2 places to update, relational engine overhead), but it's already a solved problem.

kayis profile image

how do event stores handle binary data updates efficiently?

kspeakman profile image
Kasey Speakman • Edited

A good and proper DDD / ES person would further ask the question: What is the scenario? (What problem are you trying to solve?) Because there might be a better alternative.

In an event store, you don't really update data in place if you can help it. You just keep adding new events. Depending on the size of the binary, I might do the traditional thing where you save it to file storage like S3 and then add an event when the file link changes.

Side note: The product EventStore stores the event data as binary. Meaning I actually have to serialize to string and convert to bytes before calling the method to save the event. I think internally it uses ProtoBuf to store events, which I read (but have not browsed the code to verify) handles binary data pretty efficiently.

kayis profile image

The use-case was to allow versioning of files (pdf, avi, mpg, mp3, wav, doc, xml, etc.)

Thread Thread
barryosull profile image
Barry O Sullivan • Edited

As Kasey suggested, best practice there would be to store the file remotely on S3, and then have an event that references the remote file. This allows you to version binary files, as you just upload a new file and create a new a event.


# File A, Version 1
    file_id: 9f9753b8-6beb-4e36-b7ca-1f6f6bf23702

# File A, Version 2
    file_id: 9f9753b8-6beb-4e36-b7ca-1f6f6bf23702

file_id is the reference in your system, whereas reference is the remote instance of that version.

Thread Thread
kayis profile image

I see.

Doesn't this work against the "one source of truth" philosophy?

Thread Thread
barryosull profile image
Barry O Sullivan

Yeah, technically. It's just a philosophy though, so there are times when it's not the appropriate philosophy.

In this case it really comes down to practicality. Is it useful to store the full binary file in the event log? Does that give any value? If the answer is no, then there's no point in saving the file in the log, just store a reference.

Thread Thread
kspeakman profile image
Kasey Speakman • Edited

You can have different sources of truth for each "domain" or specialty. The source of truth for files is the file system... if it is gone from there, it doesn't matter what the database says. :)

It may still be important to (event-sourced) areas of the business to record that something changed and perhaps trigger a further action. You can set this up in a number of ways. The client could issue an API call after successful saving of the file (this is request-driven, probably not the way I would go). Or you could setup event notification on file operations (this is event-driven) -- S3 supports this or just use a file watcher for local apps.

At this point, this is really integration between two systems and no longer event-sourcing. Instead it is Event-Driven Architecture. Event sourcing really only applies inside individual domains, not across different systems. This is probably why you already had an inkling that event-sourcing would not solve the file management problem. By itself, it won't.

Thread Thread
norpan profile image

We have an audit requirement that it should not be possible to change data without a trace, so storing the file separately was a problem for us until we realized that we just have to store the file sha256 hash and that way we can check if the file is the right one. So we get the best of both worlds.

smart2016 profile image

Hi Barry,

I am having nightmare with a simple question, I have a crud application similar to your example, where we really Just Create update delete on a domain entity called "Ruleset",
We are simply trying to store each request or CUD into the db , each of CUD on an entity is stored as separate record in the datastore.

I can think of that the RuleSet entity has two very simple states "SUCCESS" or "FAILURE" for each of the CUD commands and so will have entry into the datastore having states as SUCCESS or FAILURE for each CUD on the entity.

Is that correct way to design the application of CUD on Ruleset Entity model ?
I mean if this cannot be the way , how else can I design the CRUD as ES?

For example in your use case of commands:

AddCourse --> what if AddCourse Fails due to some business validations, will you not store a state with the Course entity in the persistent store ,saying SUCCESS / FAILED with command as AddCourse ?

ChangeCourse --> what if ChangeCourse Fails due to some business validations, will you not store a state with the Course entity in the persistent store ,saying SUCCESS / FAILED with command as ChangeCourse ?

RemoveCourse --> what if RemoveCourse Fails due to some business validations, will you not store a state with the Course entity in the persistent store ,saying SUCCESS / FAILED with command as RemoveCourse ?

If we store things in above manner it will be helpful to retrieve and identify when commands failed and for what reason , isn't ?

and then will the SUCCESS / FAILED state not be associated with the Course entity?


kevc_g profile image
Kevin Chaves

Good article man! I find the gathering of requirement really difficult and you gave me a great example on how to do it! Keep the good job (: