Re: Network access during build
Thomas Goirand <zigo@debian.org> writes:
> While I do agree that a package *must* be able to build without Internet
> access (for example, the test suite should never mandate access to a
> working DNS, or a query to a google search, both of which are real world
> cases...), I'm not sure about the severity: serious.
I will go farther: I am quite certain that severity: serious is simply
wrong for things like this.
I'm sure this is not the only package that attempts to test DNS functions
by looking up some well-known name. The information leak of looking up a
well-known DNS name is minimal to nonexistent. (What conclusions is
someone really going to draw from a query for www.google.com or some
similar host?) Those test suites should ideally be made robust against
that DNS query not working, but I don't even see a point in patching out
attempting to run the test provided that the test tolerates the lack of
network access to a DNS server. In other words, as long as the test is
okay with DNS not being available or not having access to public DNS, I
don't think just attempting the query is a bug of any kind. If the
current Policy wording says that it is, well, that's a bug in Policy, IMO.
Now, that said, assuming that "fail" is not a valid host in the local
domain isn't a good assumption and makes the build fragile. My packages
that perform a similar test use the DNS name addrinfo-test.invalid to
force a failure, which is guaranteed by IANA reservations to not exist.
So I think there's legitimately a bug here, but I think it's a bug of
normal severity, not some sort of emergency that would otherwise require
removing the package from Debian.
Full disclosure: several of my packages in the archive have similar tests.
Those tests are part of the upstream test suite for the getaddrinfo and
getnameinfo replacement functions for OSes that are too old to have them.
They check the results of the replacement getaddrinfo function against the
results of gethostbyname, and similarly with getnameinfo and
gethostbyaddr, and tolerate environments with no DNS by skipping the
tests. I intentionally run those tests on all builds, even ones that have
getaddrinfo and getnameinfo, because otherwise they would never run for me
(I have no ancient hosts around) and I wouldn't know if the portability
code bit rotted for some reason.
While I could put in some sort of elaborate workaround to avoid running
those tests in a Debian package build environment, I see no point in doing
so, don't really like the additional complexity, see no particular reason
why Debian should require this, and would probably just close any bugs
asking for that. (That said, I'm open to being convinced by good
arguments.)
> I don't think it is a so big issue if a package is doing some network
> operation, but doesn't fail building if there's no Internet
> connectivity.
Exactly.
> The only problem (as Christian mentioned) would be a privacy concern in
> some cases. In such a case, the severity would be "important", but not
> "serious" (ie: probably not serious enough to be an RC bug), and it'd be
> nice if the subject of the bug was reflecting the privacy concern rather
> than the "no network during build" policy thing (though I can imagine
> it'd be harder to file the bug).
"normal" is the correct severity, IMO. Even "important" strikes me as
significant severity inflation. And it would need a real justification as
to why this is a privacy concern, since typically a DNS request that's
part of a standard test suite of a free software package would not be one.
--
Russ Allbery (rra@debian.org) <http://www.eyrie.org/~eagle/>
Reply to: