On Mon, Aug 19, 2019 at 02:12:51PM +0900, Michael Paquier wrote:
> Thanks. I have somewhat not catched what Andres was suggesting here.
> So attached are two patches:
> - 0001 should take care of the compilation failure, by moving
> FD_SETSIZE into scripts_parallel.c.
I have been able to work on this one more, and applied a fix that
should address the compilation failure. While on it, I have reduced
the maximum number of parallel slots allowed to give some room for
pre-existing fds as pgbench did before moving to its poll()
implementation.
> - 0002 makes vacuumdb and reindexdb fail when trying to assign a
> socket with an unsupported range. Should this bit be backpatched? We
> are doing that for vacuumdb for some time now, and the error is
> confusing so I would prefer fixing it on older branches as well.
For this one, I am still not completely sure which way we would want
to go. The issue exists down to 9.6 for vacuumdb, so could a fix like
the one I previously proposed be acceptable for a back-patch as this
is not worth the complexity? Should we try to move to a poll()-based
implementation like pgbench on HEAD, which most likely would result in
having pgbench also use parallel slots with a bit of refactoring, no?
--
Michael