I wrote: > BTW, fd0b9dceb is in v15, so are you sure this doesn't fail in 15?
Ah-hah: simple test cases only fail since b7ae03953. Before that, the default situation was that pg_publication_rel.prattrs was null and that would be passed on as the transmitted value. b7ae03953 decided it'd be cool if pg_get_publication_tables() expanded that to show all the actually-transmitted columns, and as of that point we can get an overflow in int2vectorin given the default case where all columns are published. To break it in v15, you need a test case that publishes more than 100 columns, but not all of them (else the optimization in fetch_remote_table_info's query prevents the issue). regards, tom lane