Implementing Stripe-like Idempotency Keys is an interesting idea and we should consider incorporating this strategy directly in PgDCP.
Instead of building our own idempotency strategy, though, it might make sense to use OpenTelemetry taxonomy of spans, baggage, etc. to track multi-stage workflows as described in the article above. I'm wondering if span IDs might be more powerful than idempotency keys.