Detailed Component Design

Ability to send notification synchronously:
- clients can call sendNotification API in notification service to send notification synchronously, targeting specific cohort of userId and devices.
- notification service will call third party tools (AWS pinpoint) or iOS/Android endpoint to send the notification.
- Upon sending, there will be an error code coming back, notification service will store notification id and error code in tracking DB.
AsyncScheduleService will contain the following components
- Trigger function: step function that gets triggered every time window (5 to 10 mins), get scheduled notification in the next time window, and put them into SQS queue
- SQS queue: processing queue for notification to be sent
- NotificationProcesser: Lambda that read from SQS queue, and calls notification services to send the notification
Personalization: clients call call set configuration to store personalization configuration that will alter be used for different set of notification. When notification specifics a configuration, notification services will call configuration DB to get relevant configuration.
Ability to track how many notifications are delivered, to evaluation effectiveness: this can be done by calling getDeliverySuccessRate API in notification service, which will call tracking DB.

Scaling bottleneck

Large amount of notification: it's possible that there's a large amount of notification scheduled that failed to be sent within a certain time window. In this case, we can add functionality to the trigger function to move notification failed to send within a time window to a lower priority queue. When the Notification processor has bandwidth after processing messages from higher priority queue it will go to the lower priority queue.
Limit: we can also limit the notification scheduling that's higher than capacity and return a 400s.

Edge cases

Notification are not evenly distributed. when a notification is scheduled among a longer period of time, within the schlepping API we can break time into smaller time windows, so we don't have a large amount of notification send within a short period of time.
Past notification: past notification are likely not useful. We can create a async clean up function in the shceulder to clean up stale data in schedule DB and tracking DB