Artiom Makarov <artiom(dot)makarov(at)gmail(dot)com> writes:
> When temp tables with "on commit delete rows" exists, I can see a
> strange delay at any begin and commit.
A delay at commit is hardly surprising, because each such temp table
requires filesystem operations at commit (basically an "ftruncate").
I don't recall any operations at transaction start for such tables,
but there may be some.
regards, tom lane