Dead letter queues: patterns for handling failed messages gracefully

we've accumulated about 50k messages in our dead letter queues across various services, and honestly, nobody's really looking at them. it's become a dumping ground for messages that failed for various reasons, and it's hard to distinguish transient failures from actual bugs without digging deep into logs. we need a better strategy for handling failed messages gracefully. we're thinking of implementing exponential backoff for retries, classifying failure types (e.g., permanent vs. transient), and building a proper monitoring dashboard for dlqs with clear alerting. the big question is, how do people manage bulk replay of messages once issues are resolved? do you just re-push them to the original queue, or do you have a separate process for selective reprocessing? what's worked for others to make dlqs actionable rather than just a black hole?