architecture.md

Design goals

The main goal of event integrations is to easily enable adding new integrations over time without the need for a lot of custom work to expose events to a new integration. The ability of fan-out offered by AMQP (either in RabbitMQ or in Azure Service Bus) gives us a way to attach any number of new integrations to the existing event system without needing to add special handling. By adding a new listener to the existing pipeline, it gains an independent stream of events without the need for additional broadcast code.

We want to enable robust handling of failures and retries. By utilizing the two-tier approach (described below), we build in support at the service level for retries. When we add new integrations, they can focus solely on the integration-specific logic and reporting status, with all the process of retries and delays managed by the messaging system.

Another goal is to not only support this functionality in the cloud version, but offer it as well to self-hosted instances. RabbitMQ provides a lightweight way for self-hosted instances to tie into the event system using the same robust architecture for integrations without the need for Azure Service Bus.

Finally, we want to offer organization admins flexibility and control over what events are significant, where to send events, and the data to be included in the message. The configuration architecture allows Organizations to customize details of a specific integration; see Integrations and integration configurations below for more details on the configuration piece.

Architecture

The entry point for the event integrations is the IEventWriteService. By configuring the EventIntegrationEventWriteService as the EventWriteService, all events sent to the service are broadcast on the RabbitMQ or Azure Service Bus message exchange. To abstract away the specifics of publishing to a specific AMQP provider, an IEventIntegrationPublisher is injected into EventIntegrationEventWriteService to handle the publishing of events to the RabbitMQ or Azure Service Bus service.

Two-tier exchange

When EventIntegrationEventWriteService publishes, it posts to the first tier of our two-tier approach to handling messages. Each tier is represented in the AMQP stack by a separate exchange (in RabbitMQ terminology) or topic (in Azure Service Bus).

See the EventsProcessor / ASB detailed diagram for mote info.

Event tier

In the first tier, events are broadcast in a fan-out to a series of listeners. The message body is a JSON representation of an individual EventMessage or an array of EventMessage. Handlers at this level are responsible for handling each event or array of events. There are currently two handlers at this level:

• EventRepositoryHandler
  • The EventRepositoryHandler is responsible for long term storage of events. It receives all events and stores them via an injected IEventRepository into the database.
  • This mirrors the behavior of when event integrations are turned off - cloud stores to Azure Tables and self-hosted is stored to the database.
• EventIntegrationHandler
  • The EventIntegrationHandler is a generic class that is customized to each integration (via the configuration details of the integration) and is responsible for determining if there's a configuration for this event / organization / integration, fetching that configuration, and parsing the details of the event into a template string.
  • The EventIntegrationHandler uses the injected IOrganizationIntegrationConfigurationRepository to pull the specific set of configuration and template based on the event type, organization, and integration type. This configuration is what determines if an integration should be sent, what details are necessary for sending it, and the actual message to send.
  • The output of EventIntegrationHandler is a new IntegrationMessage, with the details of this the configuration necessary to interact with the integration and the message to send (with all the event details incorporated), published to the integration level of the message bus.

Integration tier

At the integration level, messages are JSON representations of IIntegrationMessage - specifically they will be concrete types of the generic IntegrationMessage<T> where <T> is the configuration details of the specific integration for which they've been sent. These messages represent the details required for sending a specific event to a specific integration, including handling retries and delays.

Handlers at the integration level are tied directly to the integration (e.g. SlackIntegrationHandler, WebhookIntegrationHandler). These handlers take in IntegrationMessage<T> and output IntegrationHandlerResult, which tells the listener the outcome of the integration (e.g. success / fail, if it can be retried and any minimum delay that should occur). This makes them easy to unit test in isolation without any of the concerns of AMQP or messaging.

The listeners at this level are responsible for firing off the handler when a new message comes in and then taking the correct action based on the result. Successful results simply acknowledge the message and resolve. Failures will either be sent to the dead letter queue (DLQ) or re-published for retry after the correct amount of delay.

Retries

One of the goals of introducing the integration level is to simplify and enable the process of multiple retries for a specific event integration. For instance, if a service is temporarily down, we don't want one of our handlers blocking the rest of the queue while it waits to retry. In addition, we don't want to retry all integrations for a specific event if only one integration fails nor do we want to re-lookup the configuration details. By splitting out the IntegrationMessage<T> with the configuration, message, and details around retries, we can process each event / integration individually and retry easily.

When the IntegrationHandlerResult.Success is set to false (indicating that the integration attempt failed) the Retryable flag tells the listener whether this failure is temporary or final. If the Retryable is false, then the message is immediately sent to the DLQ. If it is true, the listener uses the ApplyRetry(DateTime) method in IntegrationMessage which handles both incrementing the RetryCount and updating the DelayUntilDate using the provided DateTime, but also adding exponential backoff (based on RetryCount) and jitter. The listener compares the RetryCount in the IntegrationMessage to see if it's over the MaxRetries defined in Global Settings. If it is over the MaxRetries, the message is sent to the DLQ. Otherwise, it is scheduled for retry.

flowchart TD
A[Success == false] --> B{Retryable?}
    B -- No --> C[Send to Dead Letter Queue DLQ]
    B -- Yes --> D[Check RetryCount vs MaxRetries]
    D -->|RetryCount >= MaxRetries| E[Send to Dead Letter Queue DLQ]
    D -->|RetryCount < MaxRetries| F[Schedule for Retry]

Loading

Azure Service Bus supports scheduling messages as part of its core functionality. Retries are scheduled to a specific time and then ASB holds the message and publishes it at the correct time.

RabbitMQ retry options

For RabbitMQ (which will be used by self-host only), we have two different options. The useDelayPlugin flag in GlobalSettings.RabbitMqSettings determines which one is used. If it is set to true, we use the delay plugin. It defaults to false which indicates we should use retry queues with a timing check.

1. Delay plugin

   • Delay plugin GitHub repo
   • This plugin enables a delayed message exchange in RabbitMQ that supports delaying a message for an amount of time specified in a special header.
   • This allows us to forego using any retry queues and rely instead on the delay exchange. When a message is marked with the header it gets published to the exchange and the exchange handles all the functionality of holding it until the appropriate time (similar to ASB's built-in support).
   • The plugin must be setup and enabled before turning this option on (which is why it defaults to off).

2. Retry queues + timing check

   • If the delay plugin setting is off, we push the message to a retry queue which has a fixed amount of time before it gets re-published back to the main queue.
   • When a message comes off the queue, we check to see if the DelayUntilDate has already passed.
     • If it has passed, we then handle the integration normally and retry the request.
     • If it is still in the future, we put the message back on the retry queue for an additional wait.
   • While this does use extra processing, it gives us better support for honoring the delays even if the delay plugin isn't enabled. Since this solution is only intended for self-host, it should be a pretty minimal impact with short delays and a small number of retries.

Listener / Handler pattern

To make it easy to support multiple AMQP services (RabbitMQ and Azure Service Bus), the act of listening to the stream of messages is decoupled from the act of responding to a message.

Listeners

• Listeners handle the details of the communication platform (i.e. RabbitMQ and Azure Service Bus).
• There is one listener for each platform (RabbitMQ / ASB) for each of the two levels - i.e. one event listener and one integration listener.
• Perform all the aspects of setup / teardown, subscription, message acknowledgement, etc. for the messaging platform, but do not directly process any events themselves. Instead, they delegate to the handler with which they are configured.
• Multiple instances can be configured to run independently, each with its own handler and subscription / queue.

Handlers

• One handler per queue / subscription (e.g. per integration at the integration level).
• Completely isolated from and know nothing of the messaging platform in use. This allows them to be freely reused across different communication platforms.
• Perform all aspects of handling an event.
• Allows them to be highly testable as they are isolated and decoupled from the more complicated aspects of messaging.

This combination allows for a configuration inside of ServiceCollectionExtensions.cs that pairs instances of the listener service for the currently running messaging platform with any number of handlers. It also allows for quick development of new handlers as they are focused only on the task of handling a specific event.

Publishers and Services

Listeners (and EventIntegrationHandler) interact with the messaging system via the IEventPublisher interface, which is backed by a RabbitMQ and ASB specific service. By placing most of the messaging platform details in the service layer, we are able to handle common things like configuring the connection, binding or creating a specific queue, etc. in one place. The IRabbitMqService and IAzureServiceBusService implement the IEventPublisher interface and therefore can also handle directly all the message publishing functionality.

Integrations and integration configurations

Organizations can configure integration configurations to send events to different endpoints -- each handler maps to a specific integration and checks for the configuration when it receives an event. Currently, there are integrations / handlers for Slack, webhooks, and HTTP Event Collector (HEC).

OrganizationIntegration

• The top-level object that enables a specific integration for the organization.
• Includes any properties that apply to the entire integration across all events.
  • For Slack, it consists of the token: { "Token": "xoxb-token-from-slack" }.
  • For webhooks, it is optional. Webhooks can either be configured at this level or the configuration level, but the configuration level takes precedence. However, even though it is optional, an organization must have a webhook OrganizationIntegration (even will a null Configuration) to enable configuration via OrganizationIntegrationConfiguration.
  • For HEC, it consists of the scheme, token, and URI:

    {
      "Scheme": "Bearer",
      "Token": "Auth-token-from-HEC-service",
      "Uri": "https://example.com/api"
    }

OrganizationIntegrationConfiguration

• This contains the configurations specific to each EventType for the integration.
• Configuration contains the event-specific configuration.
  • For Slack, this would contain what channel to send the message to: { "channelId": "C123456" }
  • For webhooks, this is the URL the request should be sent to: { "url": "https://api.example.com" }
    • Optionally this also can include a Scheme and Token if this webhook needs Authentication.
    • As stated above, all of this information can be specified here or at the OrganizationIntegration level, but any properties declared here will take precedence over the ones above.
  • For HEC, this must be null. HEC is configured only at the OrganizationIntegration level.
• Template contains a template string that is expected to be filled in with the contents of the actual event.
  • The tokens in the string are wrapped in # characters. For instance, the UserId would be #UserId#.
  • The IntegrationTemplateProcessor does the actual work of replacing these tokens with introspected values from the provided EventMessage.
  • The template does not enforce any structure — it could be a freeform text message to send via Slack, or a JSON body to send via webhook; it is simply stored and used as a string for the most flexibility.

OrganizationIntegrationConfigurationDetails

• This is the combination of both the OrganizationIntegration and OrganizationIntegrationConfiguration into a single object. The combined contents tell the integration's handler all the details needed to send to an external service.
• OrganizationIntegrationConfiguration takes precedence over OrganizationIntegration - any keys present in both will receive the value declared in OrganizationIntegrationConfiguration.
• An array of OrganizationIntegrationConfigurationDetails is what the EventIntegrationHandler fetches from the database to determine what to publish at the integration level.

Filtering

In addition to the ability to configure integrations mentioned above, organization admins can also add Filters stored in the OrganizationIntegrationConfiguration. Filters are completely optional and as simple or complex as organization admins want to make them. These are stored in the database as JSON and serialized into an IntegrationFilterGroup. This is then passed to the IntegrationFilterService, which evaluates it to a bool. If it's true, the integration proceeds as above. If it's false, we ignore this event and do not route it to the integration level.

IntegrationFilterGroup

Logical AND / OR grouping of a number of rules and other subgroups.

Property	Description
AndOperator	Indicates whether all (true) or any (false) of the Rules and Groups must be true. This applies to both the inner group and the list of rules; for instance, if this group contained Rule1 and Rule2 as well as Group1 and Group2: true: Rule1 && Rule2 && Group1 && Group2 false: Rule1 || Rule2 || Group1 || Group2
Rules	A list of IntegrationFilterRule. Can be null or empty, in which case it will return true.
Groups	A list of nested IntegrationFilterGroup. Can be null or empty, in which case it will return true.

IntegrationFilterRule

The core of the filtering framework to determine if the data in this specific EventMessage matches the data for which the filter is searching.

Property	Description
Property	The property on EventMessage to evaluate (e.g., CollectionId).
Operation	The comparison to perform between the property and Value. Supported operations: • Equals: Guid equals Value • NotEquals: logical inverse of Equals • In: Guid is in Value list • NotIn: logical inverse of In
Value	The comparison value. Type depends on Operation: • Equals, NotEquals: Guid • In, NotIn: list of Guid