Differences between koji and mock rawhide environments?

[Josh Boyer]

On Wed, Nov 9, 2011 at 2:05 PM, Panu Matilainen
<pmatilai at laiskiainen.org> wrote:
> On 11/09/2011 08:49 PM, Jim Meyering wrote:
>> Richard Shaw wrote:
>>> On Wed, Nov 9, 2011 at 11:18 AM, Tom Lane<tgl at redhat.com>  wrote:
>>>> postgresql is currently failing to rebuild in rawhide:
>>>> http://koji.fedoraproject.org/koji/taskinfo?taskID=3499379
>>>> This seems quite repeatable, in koji, but the package builds fine on my
>>>> workstation under mock's fedora-rawhide-x86_64 and fedora-rawhide-i386
>>>> environments.  Shouldn't this have exactly the same results?  Can anyone
>>>> suggest how I might reproduce the failure someplace where I can debug it?
>>>
>>> I'm having a similar problem with OpenImageIO. I can build against TBB
>>> in a mock environment but it fails in koji.
>>
>> I had a similar problem with iwhd.
>> All tests passed locally(F16), but "make check" in koji would always fail.
>> I ended up leaving %check disabled in the spec file.
>>
>> Need to recheck...
>
> I've seen some similar incidents too:
>
> At some point rpm test-suite was failing because of DNS issues on (some
> of) the builders, causing unexpected "Could not canonicalize hostname"
> spew in test output.
>
> Sqlite has had a couple of rather entertaining cases: one had to do with
> the filesystem and how many syncs sqlite tests were expecting to count
> on it, this was different the builder vs fedora-of-the-time era. Another
> more recent one was that sqlite test-suite tries to test failure when
> opening file descriptors, and in order to do this it opens ~2000 fds.
> This starts failing on "normal" systems but on the builders the max
> number of fds is much higher, causing the test to fail because there was
> no failure.
>
> The only way to figure these out (that I know of) is to have some logic
> in %check to detect test-suite failure and dump out what failure logs
> you might have into koji's build.log so you have *some* clue where to
> look. Eg in rpm I have this (relying on "make check" return code wont
> work thanks to something in autofoo masking it)
>
> %if %{with check}
> %check
> make check
> [ "$(ls -A tests/rpmtests.dir)" ] && cat tests/rpmtests.log
> %endif

While I hesitate to point it out because of how rare it actually does
impact things, it's also a good idea to keep in mind that mock
duplicates the userspace package set, but it is still running on the
host kernel.  The koji builders are all RHEL 6.x, which is obviously a
different kernel version than a f15/f16 desktop.

As I said, it's fairly rare for that to be a reason for failure, but
it does happen.

josh