GitLab EventStore

Background

The monolithic GitLab project is becoming larger and more domains are being defined. As a result, these domains are becoming entangled with each others due to temporal coupling.

An emblematic example is the PostReceive worker where a lot happens across multiple domains. If a new behavior reacts to a new commit being pushed, then we add code somewhere in PostReceive or its sub-components (Git::ProcessRefChangesService, for example).

This type of architecture:

  • Is a violation of the Single Responsibility Principle.
  • Increases the risk of adding code to a codebase you are not familiar with. There may be nuances you don’t know about which may introduce bugs or a performance degradation.
  • Violates domain boundaries. Inside a specific namespace (for example Git::) we suddenly see classes from other domains chiming in (like Ci:: or MergeRequests::).

What is EventStore?

Gitlab:EventStore is a basic pub-sub system built on top of the existing Sidekiq workers and observability we have today. We use this system to apply an event-driven approach when modeling a domain while keeping coupling to a minimum.

This essentially leaves the existing Sidekiq workers as-is to perform asynchronous work but inverts the dependency.

EventStore example

When a CI pipeline is created we update the head pipeline for any merge request matching the pipeline’s ref. The merge request can then display the status of the latest pipeline.

Without the EventStore

We change Ci::CreatePipelineService and add logic (like an if statement) to check if the pipeline is created. Then we schedule a worker to run some side-effects for the MergeRequests:: domain.

This style violates the Open-Closed Principle and unnecessarily add side-effects logic from other domains, increasing coupling:

graph LR subgraph ci[CI] cp[CreatePipelineService] end subgraph mr[MergeRequests] upw[UpdateHeadPipelineWorker] end subgraph no[Namespaces::Onboarding] pow[PipelinesOnboardedWorker] end cp -- perform_async --> upw cp -- perform_async --> pow

With the EventStore

Ci::CreatePipelineService publishes an event Ci::PipelineCreatedEvent and its responsibility stops here.

The MergeRequests:: domain can subscribe to this event with a worker MergeRequests::UpdateHeadPipelineWorker, so:

  • Side-effects are scheduled asynchronously and don’t impact the main business transaction that emits the domain event.
  • More side-effects can be added without modifying the main business transaction.
  • We can clearly see what domains are involved and their ownership.
  • We can identify what events occur in the system because they are explicitly declared.

With Gitlab::EventStore there is still coupling between the subscriber (Sidekiq worker) and the schema of the domain event. This level of coupling is much smaller than having the main transaction (Ci::CreatePipelineService) coupled to:

  • multiple subscribers.
  • multiple ways of invoking subscribers (including conditional invocations).
  • multiple ways of passing parameters.
graph LR subgraph ci[CI] cp[CreatePipelineService] cp -- publish --> e[PipelineCreateEvent] end subgraph mr[MergeRequests] upw[UpdateHeadPipelineWorker] end subgraph no[Namespaces::Onboarding] pow[PipelinesOnboardedWorker] end upw -. subscribe .-> e pow -. subscribe .-> e

Each subscriber, being itself a Sidekiq worker, can specify any attributes that are related to the type of work they are responsible for. For example, one subscriber could define urgency: high while another one less critical could set urgency: low.

The EventStore is only an abstraction that allows us to have Dependency Inversion. This helps separating a business transaction from side-effects (often executed in other domains).

When an event is published, the EventStore calls perform_async on each subscribed worker, passing in the event information as arguments. This essentially schedules a Sidekiq job on each subscriber’s queue.

This means that nothing else changes with regards to how subscribers work, as they are just Sidekiq workers. For example: if a worker (subscriber) fails to execute a job, the job is put back into Sidekiq to be retried.

EventStore advantages

  • Subscribers (Sidekiq workers) can be set to run quicker by changing the worker weight if the side-effect is critical.
  • Automatically enforce the fact that side-effects run asynchronously. This makes it safe for other domains to subscribe to events without affecting the performance of the main business transaction.

EventStore disadvantages

  • EventStore is built on top of Sidekiq. Although Sidekiq workers support retries and exponential backoff, there are instances when a worker exceeds a retry limit and Sidekiq jobs are lost. Also, as part of incidents, and disaster recovery, Sidekiq jobs may be dropped. Although many important GitLab features rely on the assumption of durability in Sidekiq, this may not be acceptable for some critical data integrity features. If you need to be sure the work is done eventually you may need to implement a queuing mechanism in Postgres where the jobs are picked up by Sidekiq cron workers. You can see examples of this approach in ::LooseForeignKeys::CleanupWorker and ::BatchedGitRefUpdates::ProjectCleanupWorker. Typically a partitioned table is created and you insert data which is then processed later by a cron worker and marked as processed in the database after doing some work. There are also strategies for implementing reliable queues in Redis such as that used in ::Elastic::ProcessBookkeepingService. If you are introducing new patterns for queueing in the codebase you will want to seek advice from maintainers early in the process.
  • Alternatively, consider not using EventStore if the logic needs to be processed as part of the main business transaction, and is not a side-effect.
  • Sidekiq workers aren’t limited by default but you should consider configuring a concurrency limit if there is a risk of saturating shared resources.

Define an event

An Event object represents a domain event that occurred in a bounded context. Producers can notify other bounded contexts about something that happened by publishing events, so that they can react to it. An event should be named <domain_object><action>Event, where the action is in past tense, e.g. ReviewerAddedEvent instead of AddReviewerEvent. The domain_object may be elided when it is obvious based on the bounded context, e.g. MergeRequest::ApprovedEvent instead of MergeRequest::MergeRequestApprovedEvent.

Guidance for good events

Events are a public interface, just like an API or a UI. Collaborate with your product and design counterparts to ensure new events will address the needs of subscribers. Whenever possible, new events should strive to meet the following principles:

  • Semantic: Events should describe what occurred within the bounded context, not the intended action for subscribers.
  • Specific: Events should be narrowly defined without being overly precise. This minimizes the amount of event filtering that subscribers have to perform, as well as the number of unique events to which they need to subscribe. Consider using properties to communicate additional information.
  • Scoped: Events should be scoped to their bounded context. Avoid publishing events about domain objects that are not contained by your bounded context.

Examples

Principle Good Bad
Semantic MergeRequest::ApprovedEvent MergeRequest::NotifyAuthorEvent
Specific MergeRequest::ReviewerAddedEvent • MergeRequest::ChangedEvent
• MergeRequest::CodeownerAddedAsReviewerEvent
Scoped MergeRequest::CreatedEvent Project::MergeRequestCreatedEvent

Creating the event schema

Define new event classes under app/events/<namespace>/ with a name representing something that happened in the past:

class Ci::PipelineCreatedEvent < Gitlab::EventStore::Event
  def schema
    {
      'type' => 'object',
      'required' => ['pipeline_id'],
      'properties' => {
        'pipeline_id' => { 'type' => 'integer' },
        'ref' => { 'type' => 'string' }
      }
    }
  end
end

The schema, which must be a valid JSON schema, is validated by the JSONSchemer gem. The validation happens immediately when you initialize the event object to ensure that publishers follow the contract with the subscribers.

You should use optional properties as much as possible, which require fewer rollouts for schema changes. However, required properties could be used for unique identifiers of the event’s subject. For example:

  • pipeline_id can be a required property for a Ci::PipelineCreatedEvent.
  • project_id can be a required property for a Projects::ProjectDeletedEvent.

Publish only properties that are needed by the subscribers without tailoring the payload to specific subscribers. The payload should fully represent the event and not contain loosely related properties. For example:

Ci::PipelineCreatedEvent.new(data: {
  pipeline_id: pipeline.id,
  # unless all subscribers need merge request IDs,
  # this is data that can be fetched by the subscriber.
  merge_request_ids: pipeline.all_merge_requests.pluck(:id)
})

Publishing events with more properties provides the subscribers with the data they need in the first place. Otherwise subscribers have to fetch the additional data from the database. However, this can lead to continuous changes to the schema and possibly adding properties that may not represent the single source of truth. It’s best to use this technique as a performance optimization. For example: when an event has many subscribers that all fetch the same data again from the database.

Update the schema

Changes to the schema require multiple rollouts. While the new version is being deployed:

  • Existing publishers can publish events using the old version.
  • Existing subscribers can consume events using the old version.
  • Events get persisted in the Sidekiq queue as job arguments, so we could have 2 versions of the schema during deployments.

As changing the schema ultimately impacts the Sidekiq arguments, refer to our Sidekiq style guide with regards to multiple rollouts.

Add properties

  1. Rollout 1:
    • Add new properties as optional (not required).
    • Update the subscriber so it can consume events with and without the new properties.
  2. Rollout 2:
    • Change the publisher to provide the new property
  3. Rollout 3: (if the property should be required):
    • Change the schema and the subscriber code to always expect it.

Remove properties

  1. Rollout 1:
    • If the property is required, make it optional.
    • Update the subscriber so it does not always expect the property.
  2. Rollout 2:
    • Remove the property from the event publishing.
    • Remove the code from the subscriber that processes the property.

Other changes

For other changes, like renaming a property, use the same steps:

  1. Remove the old property
  2. Add the new property

Publish an event

To publish the event from the previous example:

Gitlab::EventStore.publish(
  Ci::PipelineCreatedEvent.new(data: { pipeline_id: pipeline.id })
)

Events should be dispatched from the relevant Service class whenever possible. Some exceptions exist where we may allow models to publish events, like in state machine transitions. For example, instead of scheduling Ci::BuildFinishedWorker, which runs a collection of side effects, we could publish a Ci::BuildFinishedEvent and let other domains react asynchronously.

ActiveRecord callbacks are too low-level to represent a domain event. They represent more database record changes. There might be cases where it would make sense, but we should consider those exceptions.

Create a subscriber

A subscriber is a Sidekiq worker that includes the Gitlab::EventStore::Subscriber module. This module takes care of the perform method and provides a better abstraction to handle the event safely via the handle_event method. For example:

module MergeRequests
  class UpdateHeadPipelineWorker
    include Gitlab::EventStore::Subscriber

    def handle_event(event)
      Ci::Pipeline.find_by_id(event.data[:pipeline_id]).try do |pipeline|
        # ...
      end
    end
  end
end

Register the subscriber to the event

To subscribe the worker to a specific event in lib/gitlab/event_store.rb, add a line like this to the Gitlab::EventStore.configure! method:

note
New workers should be introduced with a feature flag in order to ensure compatibility with canary deployments.
module Gitlab
  module EventStore
    def self.configure!(store)
      # ...

      store.subscribe ::Sbom::ProcessTransferEventsWorker, to: ::Projects::ProjectTransferedEvent,
        if: ->(event) do
          actor = ::Project.actor_from_id(event.data[:project_id])
          Feature.enabled?(:sync_project_archival_status_to_sbom_occurrences, actor)
        end

      # ...
    end
  end
end

A worker that is only defined in the EE codebase can subscribe to an event in the same way by declaring the subscription in ee/lib/ee/gitlab/event_store.rb.

Subscriptions are stored in memory when the Rails app is loaded and they are immediately frozen. It’s not possible to modify subscriptions at runtime.

Conditional dispatch of events

A subscription can specify a condition when to accept an event:

store.subscribe ::MergeRequests::UpdateHeadPipelineWorker,
  to: ::Ci::PipelineCreatedEvent,
  if: -> (event) { event.data[:merge_request_id].present? }

This tells the event store to dispatch Ci::PipelineCreatedEvents to the subscriber if the condition is met.

This technique can avoid scheduling Sidekiq jobs if the subscriber is interested in a small subset of events.

caution
When using conditional dispatch it must contain only cheap conditions because they are executed synchronously every time the given event is published.

For complex conditions it’s best to subscribe to all the events and then handle the logic in the handle_event method of the subscriber worker.

Delayed dispatching of events

A subscription can specify a delay when to receive an event:

store.subscribe ::MergeRequests::UpdateHeadPipelineWorker,
  to: ::Ci::PipelineCreatedEvent,
  delay: 1.minute

The delay parameter switches the dispatching of the event to use perform_in method on the subscriber Sidekiq worker, instead of perform_async.

This technique is useful when publishing many events and leverage the Sidekiq deduplication.

Publishing group of events

In some scenarios we publish multiple events of same type in a single business transaction. This puts additional load to Sidekiq by invoking a job for each event. In such cases, we can publish a group of events by calling Gitlab::EventStore.publish_group. This method accepts an array of events of similar type. By default the subscriber worker receives a group of max 10 events, but this can be configured by defining group_size parameter while creating the subscription. The number of published events are dispatched to the subscriber in batches based on the configured group_size. If the number of groups exceeds 100, we schedule each group with a delay of 10 seconds, to reduce the load on Sidekiq.

store.subscribe ::Security::RefreshProjectPoliciesWorker,
  to: ::ProjectAuthorizations::AuthorizationsChangedEvent,
  delay: 1.minute,
  group_size: 25

The handle_event method in the subscriber worker is called for each of the events in the group.

Testing

Testing the publisher

The publisher’s responsibility is to ensure that the event is published correctly.

To test that an event has been published correctly, we can use the RSpec matcher :publish_event:

it 'publishes a ProjectDeleted event with project id and namespace id' do
  expected_data = { project_id: project.id, namespace_id: project.namespace_id }

  # The matcher verifies that when the block is called, the block publishes the expected event and data.
  expect { destroy_project(project, user, {}) }
    .to publish_event(Projects::ProjectDeletedEvent)
    .with(expected_data)
end

It is also possible to compose matchers inside the :publish_event matcher. This could be useful when we want to assert that an event is created with a certain kind of value, but we do not know the value in advance. An example of this is when publishing an event after creating a new record.

it 'publishes a ProjectCreatedEvent with project id and namespace id' do
  # The project ID will only be generated when the `create_project`
  # is called in the expect block.
  expected_data = { project_id: kind_of(Numeric), namespace_id: group_id }

  expect { create_project(user, name: 'Project', path: 'project', namespace_id: group_id) }
    .to publish_event(Projects::ProjectCreatedEvent)
    .with(expected_data)
end

When you publish multiple events, you can also check for non-published events.

it 'publishes a ProjectCreatedEvent with project id and namespace id' do
  # The project ID is generated when `create_project`
  # is called in the `expect` block.
  expected_data = { project_id: kind_of(Numeric), namespace_id: group_id }

  expect { create_project(user, name: 'Project', path: 'project', namespace_id: group_id) }
    .to publish_event(Projects::ProjectCreatedEvent)
    .with(expected_data)
    .and not_publish_event(Projects::ProjectDeletedEvent)
end

Testing the subscriber

The subscriber must ensure that a published event can be consumed correctly. For this purpose we have added helpers and shared examples to standardize the way we test subscribers:

RSpec.describe MergeRequests::UpdateHeadPipelineWorker do
  let(:pipeline_created_event) { Ci::PipelineCreatedEvent.new(data: ({ pipeline_id: pipeline.id })) }

  # This shared example ensures that an event is published and correctly processed by
  # the current subscriber (`described_class`). It also ensures that the worker is idempotent.
  it_behaves_like 'subscribes to event' do
    let(:event) { pipeline_created_event }
  end

  # This shared example ensures that an published event is ignored. This might be useful for
  # conditional dispatch testing.
  it_behaves_like 'ignores the published event' do
    let(:event) { pipeline_created_event }
  end

  it 'does something' do
    # This helper directly executes `perform` ensuring that `handle_event` is called correctly.
    consume_event(subscriber: described_class, event: pipeline_created_event)

    # run expectations
  end
end

Best practices

  • Maintain CE & EE separation and compatibility:
    • Define the event class and publish the event in the same code where the event always occurs (CE or EE).
      • If the event occurs as a result of a CE feature, the event class must both be defined and published in CE. Likewise if the event occurs as a result of an EE feature, the event class must both be defined and published in EE.
    • Define subscribers that depends on the event in the same code where the dependent feature exists (CE or EE).
      • You can have an event published in CE (for example Projects::ProjectCreatedEvent) and a subscriber that depends on this event defined in EE (for example Security::SyncSecurityPolicyWorker).
  • Define the event class and publish the event within the same bounded context (top-level Ruby namespace).
    • A given bounded context should only publish events related to its own context.
  • Evaluate signal/noise ratio when subscribing to an event. How many events do you process vs ignore within the subscriber? Consider using conditional dispatch if you are interested in only a small subset of events. Balance between executing synchronous checks with conditional dispatch or schedule potentially redundant workers.