On Tue, Feb 22, 2022 at 03:24:31PM +0100, Ævar Arnfjörð Bjarmason wrote:
>
> On Tue, Feb 22 2022, Shubham Mishra wrote:
>
> > pipes doesn't care about error codes and ignore them thus we should not use them in tests.
>
> Aside from what Derrick Stolee mentioned in his feedback, all of which I
> agree with.
>
> I think these changes are good, but it's not the case that we try to
> avoid using pipes at all in our tests.
>
> It's often a hassle, and just not worth it, e.g.:
>
> oid=$(echo foo | git hash-object --stdin -w) &&
>
> Sure, we can make that:
>
> echo foo >in &&
> oid=$(git hash-object --stdin -w <in) &&
>
> But in the general case it's not worth worrying about.
Agreed, and I would add that we don't necessarily need to worry about
non-Git commands on the left-hand side of a pipe. So something like:
find ... | sort >actual
isn't a problem for us, because our test suite assumes that something
like find will not fail. So leaving instances of those alone is OK,
but...
> What we *do* try to avoid, and what's actually important is to never
> invoke "git" or other programs we invoke on the LHS of a pipe, or to
> otherwise do so in a way that hides potential errors.
>
> That's not isolated to just pipes, but e.g. calling it within helper
> functions that don't &&-chain, but pipes are probably the most common
> offender.
>
> The reason we do that is because in hacking git we may make it error,
> segfault etc. If it's on the LHS of a pipe that failure becomes
> indistinguishable from success.
>
> And if the test is really checking e.g. "when I run git like this, it
> produces no output" printing nothing with an exit of 0 will become the
> same as a segafault for the purposes of test.
...yes, we do care about Git failures. So something like:
git ls-files | grep "want"
would be no-good, since any failures running 'git ls-files' would be
quashed by the pipe.
Thanks,
Taylor
