Moving to an external secret manager absolutely introduces a new failure mode, and you have to design for it. The pipeline will queue, but the behavior depends on the timeout you set for the lookup function. If the manager times out, the event gets marked as a failure and can be sent to a dead-letter queue, but your processing throughput will crater.
The real gotcha isn't just the queue filling up, it's the cascading latency. One slow lookup in a pipeline with a high event rate can block processing for every event after it in that pipeline, because the function is synchronous by default. You have to wrap those lookups in aggressive caching and use worker groups sized specifically for the latency you can tolerate from the external service.
Our team learned the hard way that you need to treat the secret manager as a tier-1 dependency with its own circuit breakers, not just another config setting.