Re: [PATCH v3 2/3] test-lib-functions: optionally keep HOME, TERM and SHELL in 'test_pause'

From: Junio C Hamano <gitster@pobox.com>
Date: Sep 1, 2021, 20:26 UTC
Message-ID: <xmqq7dg0oxdm.fsf@gitster.g>
In-Reply-To: <328b5d6e76f598590d24b35ec23b5fd854c6cf05.1630503102.git.gitgitgadget@gmail.com>

"Philippe Blain via GitGitGadget" <gitgitgadget@gmail.com> writes:

Show 15 quoted lines

> +# Usage: test_pause [options]
> +#   -t
> +#	Use your original TERM instead of test-lib.sh's "dumb".
> +#	This usually restores color output in the invoked shell.
> +#	WARNING: this can break test reproducibility.
> +#   -s
> +#	Invoke $SHELL instead of $TEST_SHELL_PATH
> +#	WARNING: this can break test reproducibility.
> +#   -h
> +#	Use your original HOME instead of test-lib.sh's "$TRASH_DIRECTORY".
> +#	This allows you to use your regular shell environment and Git aliases.
> +#	WARNING: this can break test reproducibility.
> +#	CAUTION: this can overwrite files in your HOME.
> +#   -a
> +#	Shortcut for -t -s -h

As this is not end-user facing but facing our developer, I do not deeply care, but I find the warnings in this help text problematic.

Because a new process instance of $PAUSE_SHELL is run, the options you add when inserting test_pause does not affect what happens in the tests after you exit the $PAUSE_SHELL [*], right?

Of course, you can modify the repository or the working tree files used in the test in the $PAUSE_SHELL, and that can break "test reproducibility"---if you run "git ls-files -s" and take its output in a temporary file, a step later in the test that runs "git status" will see an extra untracked file, for example, and such a change may (or may not) unnecessarily break the tests.

But it is not anything new introduced by these options. It is inherent to test_pause itself.

If we want to add a warning to the help text here, I think it should be written in such a way that it is clear that the warning applies to any use of the test_pause helper, not just to the form with the options.

Thanks.

[Footnote]

* If we had an alternative implementation of test_pause that does
  not use a separate $PAUSE_SHELL process, but instead like
  inserting a read-eval-print loop, that would be far more powerful
  and useful debugging aid.  You can not just stop the execution of
  the test, and observe the files in the test repository and the
  environment variables---you can also access shell variables and
  functions.

  Such a test_pause from another world would deserve a "if you touch
  anything, the damage is permanent" warning even more than the
  current one, because you can modify even a shell variable to
  affect the execution of the test after you leave the paused state.