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.
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.
And that's bad.
But just invoking things on the LHS of a pipe? Sometimes it's a good
thing not do do that, because we'll be able to see a failure more
incrementally, and with intermediate files.
But it's generally not a problem, our test suite assumes that the OS is
basically sane. We're not going to call every invocation of "sed",
"grep" etc. with a wrapper like "test_must_fail grep" instead of "!
grep".
The same goes for our own helper utility functions such as "q_to_nul"
etc, as long as (and this is critical) that they're implemented by
shelling out to "sed", "grep", "perl" or whatever, and not to "git" or
"test-tool" etc. Then we need to start being as paranoid about them as
"git" on the LHS of pipes.
