Running a projection doesn't have to mean that all we can do is to update read models used by APIs. It is possible to handle more complex use cases that involve triggering side effects on other systems, and still keep the code concise and resilient.
Event Sourced systems implemented in tandem with Command Query Responsibility Segregation can handle very high volume of events and still deliver great experience for the users.
Because the Write Stack (command handling) is clearly separated from the Read Stack (query handling) we can prioritise which read models get updated first and which can be updated a bit later in case of a spike in traffic.
In a typical CQRS/ES system events delivered to projections have At-least-once delivery guarantee and because of that it is usually necessary to implement deduplication in order to achieve idempotency.
Although some of the message queue / streaming platforms claim that they can deliver exactly one semantics the answer to the deduplication problem problem is nuanced and we can’t get away without solving it.