Re: [PATCH] t1092: use GIT_PROGRESS_DELAY for consistent results

From: "Ævar Arnfjörð Bjarmason" <avarab@gmail.com>
To: Taylor Blau <me@ttaylorr.com>
Cc: Derrick Stolee <stolee@gmail.com>,
	Jonathan Nieder <jrnieder@gmail.com>,
	Derrick Stolee via GitGitGadget <gitgitgadget@gmail.com>,
	git@vger.kernel.org, gitster@pobox.com,
	Derrick Stolee <derrickstolee@github.com>,
	Derrick Stolee <dstolee@microsoft.com>
Subject: Re: [PATCH] t1092: use GIT_PROGRESS_DELAY for consistent results
Date: Tue, 25 May 2021 00:57:52 +0200	[thread overview]
Message-ID: <87fsybohy5.fsf@evledraar.gmail.com> (raw)
In-Reply-To: <YKwd2e5VxVmU6zqj@nand.local>

On Mon, May 24 2021, Taylor Blau wrote:

> On Mon, May 24, 2021 at 04:38:18PM -0400, Derrick Stolee wrote:
>> On 5/24/21 4:28 PM, Jonathan Nieder wrote:
>> > Hm, I think this test strategy is going to be fundamentally flaky
>> > regardless: Git doesn't intend to guarantee any kind of stability in
>> > the exact stderr output it writes.
>>
>> There are no expectations that stderr is stable across
>> versions of Git. These tests don't add friction to developers
>> making new features or changing the error messages that appear
>> over stderr. It's just that these tests should catch any
>> unintended inconsistency across these modes.
>
> I agree with Stolee that these tests are valuable for asserting that
> output is the consistent whether or not you are using the sparse index.
>
> I find setting GIT_PROGRESS_DELAY to a large number to a be a little
> ugly, but there isn't an apparent better way to accomplish the same
> thing. Of course, it would be nice to have an environment variable to
> specify where progress meters are written to, or a global option to
> disable progress meters altogether.
>
> But I don't think this isolated instance should push in the direction of
> adding support for either of the above, regardless of how easy it might
> be.

I don't see why we wouldn't just tweak GIT_PROGRESS_DELAY to support -1
or something for "inf".

It was added as a one-off (it seems for testing, but made public, so not
in the GIT_TEST_* namespace) in 44a4693bfce (progress: create
GIT_PROGRESS_DELAY, 2019-11-25).

The progress.c API will already nicely deal with this case if something
in start_progress_delay() is made to return NULL if we pass a flag down
to it.

> What would perhaps make more sense is to silence the progress meters
> from the commands themselves. AFAICT the only command called by
> run_on_sparse() which generates a progress meter is 'git checkout',
> 'git merge', and 'git submodule', all of which support '--no-progress'.
> Might it be worth passing that option instead of setting
> GIT_PROGRESS_DELAY to a large value?
>
> (For what it's worth, I have no strong opinion either way, so I would be
> happy to attach my Reviewed-by to even the current version of this patch).
>
> Thanks,
> Taylor