My vague understanding of pg_upgrade is that it just updates the metadata tables; any data changes that need to be made (which is very rare) are usually then done online and incrementally during updates... it should be a very fast operation.
That's correct, but it's not really incremental. It's all or nothing at the time you run it, even though it is a relatively small downtime because it twiddles catalog (corner case: statistics need to be regathered, of which we have a gross hack in place to copy to the best of our ability from the previous version).
All in all, logical replication will have to happen for a number of reasons:
* More OS independence
* Architecture independence
* Cross-version
* Intentional schema drift
* Alternative indexing strategies on a replica
* Avoiding both query cancellation and bloating hot-standby feedback
What Postgres has now is a more physical type of replication, based on the recovery logs, which think in terms of pages and btree/heap inserts et al rather than in terms of tuples and so forth. The main advantage of this approach is that it is very exact (to the point of being inflexible) and quite fast, and can be made even faster.
I also think logical replication will allow for a much better backup tool than pg_dump and better composition with other systems. Also, easier experimentation with distributed databases.
The release notes are available at: http://www.postgresql.org/docs/9.2/static/release-9-2.html