How do teams with tight downtime budgets choose between pg_upgrade in link mode and a logical-replication switchover for major PostgreSQL upgrades?

Open question

Both methods are documented, but the trade-off is operational. pg_upgrade --link needs a maintenance window whose length is unknown before the first rehearsal and makes the old cluster unusable once the new one starts. A logical-replication standby needs double storage, a primary key or replica identity on every table, sequences and DDL handled by hand, and a cut-over that must catch the last transactions and repoint clients. For databases from tens of gigabytes to several terabytes, with write rates that keep a subscriber busy, which method did teams pick, how long did the window or the catch-up actually take, and which step failed in rehearsal or in production? Of particular interest is whether the second copy and the manual sequence and DDL handling of the replication path were judged worth it once a rehearsed pg_upgrade window was known.

What a useful answer contains

Database size and write rate, the versions involved, the method and mode, whether standbys were upgraded with the documented rsync procedure or rebuilt, the measured downtime and the number of rehearsals, what was overlooked (statistics, extension versions, collation versions, sequences, connection strings), and whether the team would make the same choice again. Reports from managed services should state what the provider did on the team's behalf and what remained the team's job.