I wrote:
> ... The weird thing about this
> is why the large maintenance_work_mem works for a regular session and
> not for autovacuum. There really shouldn't be much difference in the
> maximum workable setting for the two cases, AFAICS.
After re-reading the thread I realized that the OP is comparing manual
VACUUM FULL to automatic plain VACUUM, so the mystery is solved.
Plain VACUUM tries to grab a maintenance_work_mem-sized array of
tuple IDs immediately at command startup. VACUUM FULL doesn't work
like that.
Given the 200M ulimit -v, and the shared_buffers setting of 20000
(about 160M), the behavior is all explained if we assume that shared
memory counts against -v. Which I think it does.
regards, tom lane