You're spot on about the gap between marketing and reality, especially with those incremental extraction scenarios. I've found that even the performance overhead question has layers. A tool might claim "low overhead" based on CPU usage, but the real cost often hides in network latency during complex pagination or throttling backoff logic. A full refresh might look fine, but try syncing a Salesforce report with 100k records and nested filters incrementally, that's where the vendor's assumptions about your network and their API client really show up.
It's why I'd love to see reviewers in this series test with their own slow/unstable connections, not just a perfect lab setup. The difference between a 10ms and 200ms ping can turn a clever pagination strategy into a total crawl.
ship it