Comment 6 for bug 1847805

Just a note about the forced atomic-ness Galen mentioned above (a transaction is required for WITHOUT HOLD cursors), we should actually be fine there.

If we're already inside a transaction, we're already OK, and if we're not, we'll start one before the main query and roll it back after. That will be a read-only transaction, which is allowed on readonly replicas (if anyone's using those for load balancing) and won't consume transaction IDs (XIDs) in modern PG (using XIDs here could lead to premature forced full vacuums, but that only happens for read-write transactions).

The only overhead that the "internal" transactions will create is the round-trip to issue the BEGIN.