Oskar Dudycz

Pragmatic about programming

How to handle multiple commands in the same transaction

2023-03-12 oskar dudyczEvent Sourcing

cover

Let’s say that we’re starting a new project. It’s a small tool for internal needs, maybe even some sort of shadow IT project. It may also be the project that won’t have massive traffic and won’t need to scale significantly. We’re more caring about the delivery time, solving the exact issue. So for our case, consistency may also be a more important factor. Yes, contrary to common belief presented in the conference talks, such projects are still run.

Take project management, for example, Jira-like software. We’d like to create a workspace for specific company departments where they can manage their projects. Our business wants to have the default project created during workspace creation. That can cut out some keystrokes and make the setup more accessible.

Is this example dumb? Maybe, as it may seem, as the typical buy-it-from-the-shelve option. Yet, the power is in the niche. I was once involved in building a product focused on project management for the construction industry. They have the specific needs to manage big CAD files and manage formal discussions on their changes. Building tools focused on solving a particular niche need may also be a correct decision, as more generic solutions won’t fulfil all the needs.

Nevertheless, the problem I’m going to explain is generic. How to handle two commands or how to change two aggregates in one transaction are one of those questions I’m asked most often.

The answer is always it depends. The classical event-driven way is to use asynchronous process together with outbox pattern. Yet, in some cases, isn’t it too much?

Let’s say that the first result of our experimentation looks like that:

[ApiController]
[Route("[controller]")]
public class WorkspaceController: ControllerBase
{
    private readonly IDocumentSession documentSession;

    public WorkspaceController(IDocumentSession documentSession) =>
        this.documentSession = documentSession;

    [HttpPost]
    public async Task<ActionResult<WorkspaceDetails>> Post(CreateWorkspaceRequest request)
    {
        var userId = GetUserId(HttpContext);
        var workspaceId = MartenIdGenerator.New();
        var cmd = new CreateWorkspace(workspaceId, request.Name, request.TaskPrefix, userId);

        var (workspaceCreatedEvent, backlogCreatedEvent) = Handle(cmd);

        documentSession.Events.Append(workspaceId, workspaceCreatedEvent);
        documentSession.Events.Append(backlogCreatedEvent.ProjectId, backlogCreatedEvent);
        await documentSession.SaveChangesAsync();

        var workspace = await documentSession.LoadAsync<WorkspaceDetails>(workspaceId);

        if (workspace is null)
            return NotFound();

        return Ok(workspace);
    }
}

It’s a simple ASP.NET controller that fulfils application layer needs. We’re using Marten to store the events defined by the business logic.

public static class WorkspaceService
{
    public static (WorkspaceCreated, ProjectCreated) Handle(CreateWorkspace command)
    {
        var slug = SlugGenerator.New(command.Name);
        var backlogId = MartenIdGenerator.New();

        return new(
            new WorkspaceCreated(
                WorkspaceId: command.WorkspaceId,
                Name: command.Name,
                TaskPrefix: command.TaskPrefix,
                Slug: slug,
                CreatedById: command.CreatedById
            ),
            new ProjectCreated(
                ProjectId: backlogId,
                Name: "Backlog",
                StartDate: null,
                EndDate: null,
                Status: ProjectStatus.Active,
                Slug: "backlog",
                WorkspaceId: command.WorkspaceId,
                CreatedById: command.CreatedById,
                IsBacklog: true
            )
        );
    }
}

If you’re not using Event Sourcing, C# or Marten, bear with me. What I will show will also apply to other logic that we need to orchestrate in a transactional way.

Getting back to our code, we’re not using aggregate or other fancy patterns, as our logic is simple enough; using function will be more than enough. We’re just taking command, generating some data and returning events that the workspace and project were created.

The code generally looks alright, but it’s already a bit smelly. We see that once it evolves with upcoming requirements, it may not withstand the test of time.

Especially those three lines look suspicious:

documentSession.Events.Append(workspaceId, workspaceCreatedEvent);
documentSession.Events.Append(backlogCreatedEvent.ProjectId, backlogCreatedEvent);
await documentSession.SaveChangesAsync();

It seems that our business logic generates two events that should be stored in different streams. We’re crossing the entity boundaries. Calling SaveChangesAsync will ensure that all of the changes will be stored or none. Consistency is guaranteed, but it seems wrong. Why?

Project and Workspace are separate things with different lifetimes, business rules, etc. If we keep the default project creation in WorkspaceService and add later on dedicated ProjectService for other operations (e.g. regular creation), then we’re not having a single source of truth.

Let’s then move the project creation to a dedicated service. Our ProjectService may look as follows:

public static class ProjectService
{
    public record CreateBacklogProject(
        Guid ProjectId,
        Guid WorkspaceId,
        Guid CreatedById
    );

    public static ProjectCreated Handle(CreateBacklogProject command) =>
        new ProjectCreated(
            ProjectId: command.ProjectId,
            Name: "Backlog",
            StartDate: null,
            EndDate: null,
            Status: ProjectStatus.Active,
            Slug: "backlog",
            WorkspaceId: command.WorkspaceId,
            CreatedById: command.CreatedById,
            IsBacklog: true
        );
}

Nothing fancy, but it’s already prepared for upcoming rules and methods specific to the project. Notice that we already embraced in the command name that backlog project may differ from a regular one. It is a small detail, but better not to generalise our business logic too early. We can always do that later when we understand the nuances of our domain better. We could also create a dedicated event for that, e.g. BacklogProjectCreated if we believe that’s different enough from the regular one.

The WorkspaceService will look now as:

public static class WorkspaceService
{
    public record CreateWorkspace(
        Guid WorkspaceId,
        string Name,
        string TaskPrefix,
        Guid CreatedById
    );
    
    public static WorkspaceCreated Handle(CreateWorkspace command) =>
        new WorkspaceCreated(
            WorkspaceId: command.WorkspaceId,
            Name: command.Name,
            TaskPrefix: command.TaskPrefix,
            Slug: SlugGenerator.New(command.Name),
            CreatedById: command.CreatedById
        );
}

Our business logic didn’t change much, but we already see explicitly that those are different operations:

var workspaceCreated = Handle(new CreateWorkspace(workspaceId, request.Name, request.TaskPrefix, userId));
var backlogCreated = Handle(new CreateBacklogProject(backlogId, workspaceId, userId));

documentSession.Events.Append(workspaceId, workspaceCreated);
documentSession.Events.Append(backlogId, backlogCreated);

await documentSession.SaveChangesAsync();

If we know already that we’ll need this pattern more often, we could also add a helper method to handle such scenarios.

public static class CommandHandling
{
    public static Task ComposeAsync(this IDocumentSession documentSession, params (Guid, object)[] events)
    {
        foreach (var (streamId, @event) in events)
        {
            documentSession.Events.Append(streamId, @event);
        }

        return documentSession.SaveChangesAsync();
    }
}

And use it as such:

await documentSession.ComposeAsync(
    (workspaceId, Handle(new CreateWorkspace(workspaceId, request.Name, request.TaskPrefix, userId))),
    (backlogId, Handle(new CreateBacklogProject(backlogId, workspaceId, userId)))
);

If we’d like to make this process explicit, we could even set up a Domain Service to orchestrate this process.

using static WorkspaceService;
using static ProjectService;

public static class WorkspaceCreationScenario
{
    public static (WorkspaceCreated, ProjectCreated) CreateWorkspace(
        Func<Guid> generateId,
        Guid userId,
        string name,
        string taskPrefix
    )
    {
        var workspaceId = generateId();
        var backlogId = generateId();

        return new(
            Handle(new CreateWorkspace(workspaceId, name, taskPrefix, userId)),
            Handle(new CreateBacklogProject(backlogId, workspaceId, userId))
        );
    }
}

We can also define an explicit code to orchestrate that with storage:

public static class WorkspaceCreationScenarioHandler
{
    public static async Task<Guid> CreateWorkspace(
        this IDocumentSession documentSession,
        Guid userId,
        string name,
        string taskPrefix
    )
    {
        var (workspaceCreated, projectCreated) =
            WorkspaceCreationScenario.CreateWorkspace(MartenIdGenerator.New, userId, name, taskPrefix);

        await documentSession.ComposeAsync(
            (workspaceCreated.WorkspaceId, workspaceCreated),
            (projectCreated.ProjectId, projectCreated)
        );

        return projectCreated.WorkspaceId;
    }
}

Then our controller method can look like this:

[HttpPost]
public async Task<ActionResult<WorkspaceDetails>> Post(CreateWorkspaceRequest request)
{
    var workspaceId = await documentSession.CreateWorkspace(
        GetUserId(HttpContext),
        request.Name,
        request.TaskPrefix
    );

    var workspace = 
        await documentSession.LoadAsync<WorkspaceDetails>(workspaceId);

    if (workspace is null)
        return NotFound();

    return Ok(workspace);
}

Didn’t we just create more code? Yes, we did; what’s the reason, then?

It could be helpful if your scenario evolves. Let’s say that besides the default project, you’d like to also:

  • create a default project manager,
  • send an email,
  • setup some storage for project files,
  • etc.

In that case, you might just need to change the parts that should be moving. It’ll only involve updating the orchestration code in Scenario by adding integration with another logic. You won’t need to touch other command handlers or API endpoint.

Of course, you need to make a sanity check. At some point, the write amplification may be too big. Modifying too many entities in the same transaction is asking youurself for concurrency conflicts and deadlocks. Asynchronous processing wasn’t invented just to make our life harder. At least, not only.

Still, we already made things explicit; we distilled the business logic from coordination and drawing boundaries. If it appears that we’re doing too much, we can replace our scenario with Saga or another asynchronous coordination style. We focused on composition instead of adding new layers, as generic does not mean simple.

You could notice that I didn’t include returning data in our scenario class. Why? I’m a huge fan of separating business logic from queries. A clear split of which classes are responsible for running business logic and querying makes code more predictable. We may decide to orchestrate that on the application layer and still return the data from the controller method, but we should keep the split layer down. I wrote longer on those considerations in Can command return a value?.

What if we do one more step and isolate the data retrieval from command handling but still give the client application a similar experience? We could use Location header for that. It tells where’s the newly created resource. The client will need to do one more call, but typically it’s not an issue as browsers, and most of the mature HTTP client libraries do it automatically after seeing 201 CREATED status.

How to do it? Simple as that:

[ApiController]
[Route("[controller]")]
public class WorkspaceController: ControllerBase
{
    [HttpPost]
    public async Task<ActionResult<WorkspaceDetails>> Post(
        [FromServices] IDocumentSession documentSession,
        CreateWorkspaceRequest request
    )
    {
        var workspaceId = await documentSession.CreateWorkspace(
            GetUserId(HttpContext),
            request.Name,
            request.TaskPrefix
        );

        return Created($"/workspace/{workspaceId}", workspaceId);
    }

    [HttpGet("{id:guid}")]
    public async Task<ActionResult<WorkspaceDetails>> GetById(
        [FromServices] IQuerySession querySession,
        Guid id
    )
    {
        var workspace = await querySession.LoadAsync<WorkspaceDetails>(id);

        return workspace is not null ? Ok(workspace) : NotFound();
    }
}

It is also a good practice not to do redundant roundtrips, especially by default, assuming the client will need this data. Most of the time, the client already knows what they submitted. Of course, some data may be generated on the backend, but will the client immediately need them? We should always try to define the UI in that way to make the user flow the most effective, but we also take other considerations. Smart UX can also cut off a lot of complexity from the technical solution.

So, how far can you go with this pattern? There’s no obvious answer. Going down that path is always a tradeoff and potential design smell. We should be careful applying that and avoid getting too attached to it. I suggest using it as a pragmatic choice, a way to experiment and deliver business value.

All models are wrong, but some of them are useful. Don’t treat your model as set in stone but as a living thing that should evolve when you learn more about your business or requirements change.

Build the code that’s explicit about intention and removable. I hope this article gives you a decent food for thought on how to achieve that.

You can also check this PR and follow commits to see the refactoring step by step.

Cheers!

Oskar

p.s. Ukraine is still under brutal Russian invasion. A lot of Ukrainian people are hurt, without shelter and need help. You can help in various ways, for instance, directly helping refugees, spreading awareness, putting pressure on your local government or companies. You can also support Ukraine by donating e.g. to Red Cross, Ukraine humanitarian organisation or donate Ambulances for Ukraine.

👋 If you found this article helpful and want to get notification about the next one, subscribe to Architecture Weekly.

✉️ Join over 5000 subscribers, get the best resources to boost your skills, and stay updated with Software Architecture trends!

Loading...
Event-Driven by Oskar Dudycz
Oskar Dudycz For over 15 years, I have been creating IT systems close to the business. I started my career when StackOverflow didn't exist yet. I am a programmer, technical leader, architect. I like to create well-thought-out systems, tools and frameworks that are used in production and make people's lives easier. I believe Event Sourcing, CQRS, and in general, Event-Driven Architectures are a good foundation by which this can be achieved.