Re: [PATCH 4/4] t4300 (rebase): don't unnecessarily set GIT_TRACE

From: Junio C Hamano <gitster@pobox.com>
Date: May 10, 2013, 21:06 UTC
Message-ID: <7vk3n6zgr3.fsf@alter.siamese.dyndns.org>
In-Reply-To: <20130510191639.GB3478@elie>

Jonathan Nieder <jrnieder@gmail.com> writes:

Show 7 quoted lines

> Ramkumar Ramachandra wrote:
>
>>                           Can we do better by not printing the -v
>> output of the passing tests though?
>
> Not for my use.  The output from comprable tests before is often
> useful for comparison.

Perhaps.

But the output from passing "-v" before the test that breaks is not very useful for two reasons.

 - Often they are identical between passing and failing runs because
   the resulting states in their trash directories are truly the
   same, in which case the output is just being too noisy.

 - Some other times, the output from passing and failing runs look
   identical but only because earlier tests are sloppy and not
   showing differences that happens to matter as precondition to the
   failing test, in which case the output is too sketchy to be
   useful.

In the latter case, where a breakage needs deeper inspection, I haven't found a way better than to stop the test _before_ the broken one (by temporarily inserting "exit" just before it), run it once again so that the trash directory just before the failing test is left behind, and go there to manually inspect the states in the trash directories of passing and failing ones to compare where in the earlier tests that passed left a subtle differences that were not detected but still caused to make a difference to the broken test.

I am not saying "it is not very useful, so remove it" but "it is not very useful, and I wish somebody clever thinks of a way to make it more useful", though ;-)