Hannes Dorbath <light@theendofthetunnel.de> writes:
> I have the bad feeling that I need to correct this into "It happens when
> autovacuum is active on the table".
Ah-hah, I realize how to explain that too, now. If you start the CREATE
INDEX CONCURRENTLY while an autovacuum is in progress on that table, the
autovacuum gets booted off so that C.I.C. can get the lock. And, most
likely, it leaves the syncscan pointer positioned someplace past block
zero. So the part that is a bit hard to reproduce is to launch the
C.I.C. while an autovacuum is in progress.
regards, tom lane