Skip to content

Latest commit

 

History

History
178 lines (115 loc) · 13.7 KB

README.md

File metadata and controls

178 lines (115 loc) · 13.7 KB
Raw

Beef.Events

NuGet version

To support the goals of an Event-driven Architecture Beef enables the key capabilities - the publishing and subscribing of events (messages) to and from an event-stream (or equivalent).

Layers


Azure Event Hubs

Beef is largely agnostic to the underlying event/messaging infrastructure (event-stream) and for the most part this must be implemented by the developer. This is by design, to decouple Beef and enable multiple capabilities to be supported as required.

However, further capabilities have also been enabled within Beef, leveraging the agnostic capabilities described above, to directly support Azure Event Hubs as an event-stream.


Event Data

The Beef EventData provides a standardised, and flexible, message contract that is used to both publish and subscribe to events in a consistent manner. The key properties as follows:

Property Description
Subject The subject (noun) that uses dot-notation to uniquely describe an entity instance.
Action The action (verb) that describes what was performed on the subject.
Key The corresponding entity key (could be single value or an array of values).
Value The optional value that contains the underlying message data.

The code-generation of Beef will automatically infer the Subject and Action for an entity and the operation being performed; this can be overridden. The Subject defaults to <domain>.<entity>.<uniquekey>; there is a code-gen configuration attribute EventSubjectRoot that prepends the specified value to the Subject. The Action defaults to the Operation name; there is a code-gen configuration attribute EventActionFormat that enables automatic reformatting to past tense where required (Create would be renamed to Created for example).

For example, an invoice creation event could result in the following:

  • Subject: Billing.Invoice.7634626f-edee-e911-bd3b-bc8385e26041
  • Action: Create
  • Key: 7634626f-edee-e911-bd3b-bc8385e26041

Publishing

The Event provides the standardised event processing/publishing. The underlying PublishAsync set of methods provide the means to publish the event(s). By default, the events are not processed until a publisher has been registered; the Register method will allow one or more publishers to be registered.

To publish to Azure Event Hubs the EventHubPublisher can be used.

The publishing of events is integrated into the API processing pipeline; this is enabled within the Service orchestration layer to ensure consistency of approach. All Create, Update and Delete operations raise events automatically; others will need to be issued directly by the developer.


Subscribing

Within the Beef context each Domain would subscribe (listen) to events and process accordingly. The following components are required:

  • Subscribers - one or more subscribers;
  • Host - a host to receive events and route to a subscriber.

Subscribers

To start with a developer would create one or more subcribers; one per Subject and Action combination. A subscriber is created by inheriting from EventSubscriber or EventSubscriber<T> (defines the Event.Value Type to automatically deserialize) specifying the Subject template (supports wildcards) and optional Action(s); finally implementing the ReceiveAsync logic. The ReceiveAsync must return a Result to describer the processing outcome; being one of: Success, DataNotFound (also inferred from a NotFoundException), InvalidData (also inferred from a ValidationException or BusinessException).

There are the following properties that can be set that will change the runtime logic:

Property Description
UnhandledExceptionHandling Provides the unhandled Exception option as either Stop (stops and bubbles up the Exception allowing the host process to determine the appropriate action) or Continue (skips and continues effectively swallowing the Exception). Defaults to Stop.
RunAsUser Provides the run as user option as either Originating (originating user being EventData.Username) or System (EventSubscriberHost.SystemUsername). Defaults to Originating.
InvalidEventDataHandling Overrides the default ResultHandling as specified for the owning EventSubscriberHost.
DataNotFoundHandling Overrides the default ResultHandling as specified for the owning EventSubscriberHost.
InvalidDataHandling Overrides the default ResultHandling as specified for the owning EventSubscriberHost.

See example below:

public class PowerSourceChangeSubscriber : EventSubscriber<string>
{
    // Subscribe to all events with a Subject starting with 'Demo.Robot.' as per wildcard
    // and Action 'PowerSourceChange'.
    public PowerSourceChangeSubscriber() : base("Demo.Robot.*", "PowerSourceChange") { }

    public override async Task<Result> ReceiveAsync(EventData<string> @event)
    {
        var val = @event.Value;
        // Processing logic...
        return Result.Success();
    }
}

Subscriber Host

The EventSubscriberHost provides the base capabilities for the host. The host is responsible for receiving (ReceiveAsync) each event and converting to an EventData instance. The host checks whether there is a subscriber and will invoke where found; otherwise, the event will be skipped (ignored).

The EventSubscriberHost infers the subscribers automatically by reflecting on the Assembly that instatiates the host. The subscribers can alternatively be specified by using the EventSubscriberHostArgs.

There are the following properties that can be set that will change the runtime logic:

Property Description
AreMultipleMessagesSupported Indicates whether multiple messages can be processed; default is false.
InvalidEventDataHandling Determines the behaviour (ResultHandling) where the EventData is unable to be converted or has no Subject specified. Defaults to Stop.
DataNotFoundHandling Determines the behaviour (ResultHandling) where the corresponding data is not found. Defaults to Stop.
InvalidDataHandling Determines the behaviour (ResultHandling) where the specified data is invalid. Defaults to Stop.

Warning: The ResultHandling.Stop implies that the data is currently unable to be processed; and should be attempted again (it could be transient) - this however could result in a poison message that could stop further processing until resolved. Consider the ResultHandling.ContinueWithAudit, ResultHandling.ContinueWithLogging or ResultHandling.ContinueSilent as an alternative.

The ExecutionContext creation can be overridden by overridding the CreateExecutionContext method where neccessary.

To subscribe to Azure Event Hubs the EventHubSubscriberHost can be used.


Azure Function

The process host, for example an Azure Function, will create/instantiate the subscriber host. To set up the likes of connection strings, etc. a startup-style component will be required.

An example using a Resilient Event Hub Trigger is as follows:

public static class EventSubscriber
{
    [FunctionName("EventSubscriber")]
    public static async Task Run([ResilientEventHubTrigger] EventHubs.EventData @event, ILogger log)
    {
        await EventHubSubscriberHost.Create(log).ReceiveAsync(@event);
    }
}

Resilient Event Hub Trigger

A new custom Azure Function Trigger, ResilientEventHubTrigger, has been created to support greater resiliency when processing (reading from) an Azure Event Hub Consumer Group. This should be used in scenarios where each message must be processed, unless explicitly skipped.

Unfortunately, probably by design, the out-of-the-box EventHubTrigger does not have any resiliency built-in. In that, if the underlying Function being executed fails (throws an Exception) this is essentially logged and swallowed, moving onto the next message in sequence (for the Consumer Group + Partition Id).

This behaviour is tolerable where message loss is acceptable, for example processing sensor data. Whilst this loss is not ideal it should not result in underlying data inconsistencies, etc.

For the Event-Driven approach between the likes of Domains and/or Applications it is assumed that each message must be processed; either explicitly ignored (skipped), or acted on; i.e. message loss is not acceptable.


Logic

The underpinnings of the ResilientEventHubTrigger logic is:

  • All unhandled Function exceptions will be caught for a message; and then re-processed (retry) until successfully processed; or alternatively, skipped (explictly).
  • The retry logic leverages Polly and uses a base 2 exponential back-off strategy. Each retry timespan is calculated as 2^count to calculate the sleep seconds. For example the 3rd retry would be 2^3, being 8 seconds. There is also a jitter to add a random number of milliseconds to minimise bunching. A maximum timespan can also be specified, defaults to 15 minutes - so that it eventually falls back to a standard frequency upon which to retry.
  • At times the exception/failure could be transient, and as such, the message is not considered Poison until it has been retried a number of times. By default, the message will be logged as Poison after 6 attempts (2^6, being 64 seconds). Initially, warnings will be logged; then they will be treated as errors.
  • The Poison message will be written to an Azure Storage Table (EventHubPoisonMessage) for the Event Hub + Consumer Group + Partition Id, that includes the likes of: Exception, EventData.Body (as a string), Function Name, etc. There is a property SkipMessage that determines whether the message is to be skipped (defaults to false).
  • Whilst there is a Poison message with a SkipMessage of false it will continue to retry - even if the Function is restarted; i.e. the message will continue to be retried.
  • To skip the Poison message the SkipMessage must be set explicitly to true. Each time a Poison message is re-tried it looks for (re-reads) the SkipMessage value and acts accordingly. Once skipped the Poison messages is removed (deleted) from the storage table; a copy is written to a corresponding skipped Azure Storage Table (EventHubPoisonMessageSkipped) as an audit.
  • Where the ResultHandling of an EventSubscriber is ResultHandling.ContinueWithAudit the message will be written to the skipped Azure Storage Table (EventHubPoisonMessageSkipped) as the audit. The SkipMessage value will be false to indicate that this was not explicitly (externally) set due to being Poison.

The PoisonMessagePersistence provides the Poison message Azure storage persistence functionality described. This behaviour can be overridden using the IPoisonMessagePersistence where required.


Configuration options

The following additional configuration options exist:

Option Description
MaxRetryMinutes The maximum retry TimeSpan. Defaults to 15 minutes. Maximum allowed value is 24 hours.
LogPoisonMessageAfterRetryCount Determines whether a possible Poison message has been encountered and should be logged (persisted) after the specified retry count (must be between 1 and 10, defaults to 6).

Constraints

The following run-time constraints exist:

  • Given the required resiliency retry requirements within this trigger it is unable to be executed on an Azure consumption plan; i.e. must be always-on.
  • The invoked function can only accept a single EventData parameter; i.e. arrays are not supported.