On Mon, Jan 26, 2015 at 7:57 AM, Julian Foad <julianfoad_at_btopenworld.com>
> Hi Bert! Thanks for airing your concerns.
> Bert Huijben wrote:
> > I see added value in these tests, but can we please make this behavior
> > before enabling for everybody all the time?
> Certainly! That's one of the three TODO tasks I listed.
> > I don't see why every test in the testsuite needs a double dump and
> > comparison in every testrun (on every test invocation)
> Every test potentially generates a different repo. Every RA layer
> potentially gives different behaviour with 'svnrdump' (issue #4551 includes
> an example). Each FS type potentially behaves differently.
> The problem of excessive duplication of coverage in our testing regime is
> not a new concern here.
> > (And then the patch appears to ignore the fact that we have tests that
> > multiple repositories)
> Ignore? No, just not implemented yet. In the patch's log message it says:
> Ideas for improvement:
> - Improve the logic for finding repositories created by a test: detect
> when a test created a repository even if the sandbox is not marked as
> 'built'; detect when a test created additional repositories.
> - Implement the same cross-checking for the C tests.
> > I can't see why the coverage is better this way, than running this just
> in a
> > single configuration...
> Obviously the coverage is "better" in the sense of "more", so maybe you
> mean "better" in the sense of amount of coverage in proportion to the time
> > except by slowing developers down (and thereby reducing
> > the number of new bugs... just by reducing their productivity)
> This extra test coverage will be optional. Don't enable it if you don't
> want to.
> Trying to unpick what you really mean, I feel you are unhappy that the
> current set of tests that you run frequently (before each commit, perhaps)
> is too slow for your liking, and you think this addition will make it
> slower without a proportional increase in coverage. You are right about the
> last part -- this extra testing doubtless doesn't add as much coverage, in
> proportion to its run time, as adding a regression test targeted to a
> specific bug.
> So maybe the point you are trying to make here is that this kind of
> "blanket" testing is not as "efficient", in the sense of coverage over
> execution time, as specifically targeted tests. Is that right?
> Of course in another sense it is very efficient, in that it can detect a
> large class of bugs with very little human effort.
> > With the same reasoning: better coverage is better, we can just as well
> > the flag on which filesystem we test, and always run BDB and FSFS.
> > Or skip the check which RA layer, and run them all.
> What's your point? Of course we don't want to run all the possible test
> permutations a hundred times a day during our own development work flow.
> And of course we DO want to run all the possible tests sometimes, before
> shipping software.
> You seem to be thinking that there is exactly one set of tests, and that
> everybody has to run the same set of tests every time for every purpose.
> As developers, each of us chooses what subset of all possible tests to
> run, and how often, depending on our work patterns, our machine speed, the
> likelihood that the change we're working on will be detected by
> certain tests, the importance of getting the change right first time, and
> other considerations.
> > We separated these tests over multiple configurations for a reason, and
> I think
> > this should behave the same way
> What way do you mean?
> By being so negative about it, you are sending out a signal that
> additional testing is unwelcome. It is very easy to spread a negative
> feeling. I feel that any time I commit or even propose to implement some
> extra testing, you'll likely argue against it. Why should developers bother
> trying to make the software well tested if that's the attitude?
> I'm sure that's not what you mean, but that's how it comes across. Please
> can we resolve this argument so we're clear on how extra testing can both
> be welcomed and its run time kept under control?
FWIW, I read Bert's reply as "Can we please make this behavior optional"
and everything that came after that was just his pre-emptive argument
against anyone who does not think it should be possible to turn these tests
on/off. Since you are in agreement on Bert's point, I do not think there
is any disagreement.
Received on 2015-01-26 17:03:16 CET