When is sharding preferable to horizontal scaling for read replicas?

Hey everyone, I'm working on a service with a rapidly growing dataset and increasing read load. We've primarily relied on horizontal scaling with read replicas so far, which has worked well for our analytics and reporting needs. However, I'm starting to wonder if we're approaching the limits of what read replicas can efficiently handle without introducing significant operational complexity or data staleness. When do you typically start considering database sharding as a more effective scaling strategy compared to simply adding more read replicas? Are there specific thresholds (e.g., number of replicas, query latency, data size) that act as a good indicator? What are the key trade-offs in terms of complexity, consistency, and maintenance for each approach?