"Tomeh, Husam" <htomeh(at)firstam(dot)com> writes:
> =# vacuum analyze code;
> ERROR: out of memory
> DETAIL: Failed on request of size 1073741820.
That looks a whole lot like a corrupt-data issue. The apparent
dependency on maintenance_work_mem is probably illusory --- I suspect
some of your trials are selecting the corrupted row to use in the
ANALYZE stats, and others are randomly selecting other rows.
If you are able to pg_dump the table in question then this theory is
wrong, but I'd suggest trying that first.
regards, tom lane