Re: What's cooking in git.git (Aug 2021, #02; Tue, 3)

On Wed, Aug 04, 2021 at 11:28:25PM +0200, SZEDER Gábor wrote:

On Wed, Aug 04, 2021 at 09:53:02PM +0200, Ævar Arnfjörð Bjarmason wrote:

quoted

On Wed, Aug 04 2021, Junio C Hamano wrote:

quoted

Ævar Arnfjörð Bjarmason [off-list ref] writes:

quoted

quoted

* ab/test-columns (2021-08-02) 3 commits
 - test-lib.sh: use GIT_TEST_COLUMNS over COLUMNS
 - test-lib-functions.sh: add a test_with_columns function
 - test-lib-functions.sh: rename test_must_fail_acceptable()

We're going to need this or another solution to the v2.33.0-rc0
regression in c49a177beca (test-lib.sh: set COLUMNS=80 for --verbose
repeatability, 2021-06-29) before the final v2.33.0.

Just a question.  Is that true?  Wouldn't a system that needs these
on top of c49a177beca already break the tests without c49a177beca?

IOW, is c49a177beca truly a "regression", or is it merely "a half
solution that solves for most but not all platforms"?

Yes, because with c49a177beca your tests only break if you use the
--verbose option, i.e. if your stderr is connected to a terminal. I.e.:

    ./t0500-progress-display.sh --verbose

So in practice it mostly affects git developers who run with --verbose,
but probably nobody doing a build in the wild.

With c49a177beca they break on e.g.:

    /bin/bash ./t0500-progress-display.sh

If your bash is recent enough, so "make test" if you're on a system with
a recent bash whose /bin/sh is /bin/bash.

This is because post-c49a177beca we don't "unset" COLUMNS anymore, which
bash takes as license to update it.

So we really do need that series in before the release to avoid that
common annoyance, a revert of c49a177beca is also possible, i.e. it
would still leave things broken under --verbose, but that breakage is
rare and existed before v2.33.

I think given the triviality of the fixes and that the cause is
well-understood it makes sense to go forward in this case.

On one hand, there is feedback to be addressed:

  https://public-inbox.org/git/20210802171429.GC23408@szeder.dev/

OTOH, setting the checkwinsize is the truly trivial, minimal, reliable

Oh, this should have been s/setting/turning off/

and uncontroversial fix for this issue, and IMO that should go into
the next release.

This fix in this patch series is not trivial: it introduces yet
another GIT_TEST variable and a helper function that developers will
have to remember to use in the future.  Worse, this means that despite
aiming for future proofing I can't consider this approach future
proof, because it's easy to forget about such a rarely used test
helper function, and if anyone introduces yet another test setting
COLUMNS, then that will be prone to similar failures when run with
Bash.

I don't think that in this case we should aim for future proofing when
the cost is the additional cognitive load of yet another helper
function.  I would instead prefer to go with the really trivial fix
for now and wait whether this issue pops up again with other shell or
terminal, hoping that this issue is a "one-hit-wonder" [1] and it
won't happen ever again.


[1] "one-hit-wonder" as in:
    https://en.wikipedia.org/wiki/Bloom_filter#Cache_filtering