René's URL Explorer Experiment

Title: Make clear every test's status in every CI run by EliahKagan · Pull Request #1679 · gitpython-developers/GitPython · GitHub

Open Graph Title: Make clear every test's status in every CI run by EliahKagan · Pull Request #1679 · gitpython-developers/GitPython

X Title: Make clear every test's status in every CI run by EliahKagan · Pull Request #1679 · gitpython-developers/GitPython

Description: This makes changes so tests give more informative output, both locally and on CI but especially on CI. It comprises several interrelated changes that should help lead up to, but do not include, adding CI jobs for Windows. My motivation for doing this was to make it easier to add those jobs, and to get the most out of those jobs once they are added, but it seems to me that the benefit of these changes is actually largely independent of those goals, and can be reviewed independently of them (and thus before they are done). This description is divided into descriptions of the problems I believe the changes solve, and how they solve them. pytest-sugar output didn't display correctly on CI The pytest-sugar plugin is working very well for local testing, but on CI it has been producing output that is voluminous, yet fairly low in information. This happens because instead of outputting a row of check marks as it does locally (at least when run in most terminals), updating the line creates a new one: test/test_actor.py ✓ 0% test/test_actor.py ✓✓ 0% ▏ test/test_actor.py ✓✓✓ 1% ▏ test/test_actor.py ✓✓✓✓ 1% ▏ Those are four passing tests in test_actor.py, reported with one line per test, yet those lines give no information about the individual tests they represent, not even those tests' names. This happens because, from pytest-sugar's perspective, it is updating a single line in a terminal. The cause of the problem is the progress bar animated on the right side. Passing -v or -vv improves things somewhat, in that it shows specific information about the tests (their names and, where applicable, which in this project is almost everywhere, the test class they appear in). However, the progress bar of that style is still drawn, and extra newlines are still often shown, with the resulting excess blank lines making the output hard to read. I was not able to find a way to get pytest-sugar not to show this progress bar, or to show it in a different style. Fortunately, that turned out not to be necessary. pytest-sugar provides two useful features: its pretty output, and showing each failure immediately. The former is not achieved on CI, for this project, currently. However, the latter is, and may be considered important. But there is another pytest plugin that separately provides just that feature: pytest-instafail. When enabled, pytest-instafail reports each failure with full details immediately. (It does not stop running the tests early or cancel any tests.) So I added pytest-instafail as a development dependency (in the test extra) but did not set things up for it to be used automatically. And I configured pytest to use pytest-sugar by default, but invoked pytest on CI so that it runs with pytest-sugar turned off and pytest-instafail turned on. This plays well with how GitHub Actions handles output, at any level of verbosity. But I have also passed -vv on CI, to solve... CI did not show which tests had which statuses CI is the easiest and also most reliable way to run tests in this project, and probably the most common. Often I found myself wanting to know what tests were actually running, versus being skipped. I often made assumptions about this that were mistaken, and others that were correct but hard to be confident about; see #1657 (comment). Furthermore, it is useful to be able to notice--even when one is not looking for it, so long as it's not too distracting--what tests are passing, skipped, etc. The failing tests are reported in detail at each failure and listed at the end, but tests with other statuses are not listed with those statuses--or, really, at all--without -v or -vv. So I have passed -vv so that, for non-failing tests, each test shows a line indicating the file, the class where applicable, the name of the test case (with @ddt-parameterized tests appearing as separate test cases but listed together, and named in terms of their arguments), and the test's status. This distinguishes tests with pass, skip, xfail, and xpass statuses. It also provides reassurance that particular tests really are running, and allows one to search for tests by name while viewing CI output. The custom pytest configuration was rigid and hid failures Implicit options for pytest are configured in pyproject.toml. It is not immediately obvious, to someone running pytest on the project or inspecting CI output, what these are set to. One was --force-sugar, which appears to have been necessary to get pytest-sugar to run on CI even in spite of its (in this case correct) guess that it does not have a suitable output device when doing so. This made it very difficult to run pytest without pytest-sugar, as is needed on CI (detailed above) and as may also be useful to do locally in some cases. So I removed that. The plugin is still loaded automatically, but it can now be turned off with -p no:sugar (as well as automatically if it detects an unsuitable output device, but I am not relying on that in the CI workflows). Another was --maxfail=10. I have removed this as well. It made pytest stop running tests once the tenth failure occurred, so no more than ten failures would ever be reported. It was possible to notice this by looking carefully at test output, either by noticing the line that warns about it (amongst lots of other output) or by seeing that the total number of tests that are run was lower than the number the runner found, by more than the number of skipped tests. On CI, it is useful to see the output of all tests. Sometimes it makes sense to weaken that, but stopping in the middle of a test run because of a failure in that run (where one may want to know what else does, and what doesn't, also fail), is not, in my view, the best way to do it. Instead, fail-fast could be enabled on the CI test matrix, or some available combinations of platforms and Python versions could even be omitted (for example, when Windows CI jobs are added, fewer than six versions could be tested, depending on how fast they run). The bigger issue, though, is that I thought I was getting only ten test failures in a number of circumstances! Fortunately, I never relied on that, and even on my native Windows system where I was unable to get all tests passing, I ran the tests that seemed most germane to the changes I was making. Since it is more involved--on any operating system--to get all tests passing locally with GitPython than with most codebases, showing only ten failures seems like a stumbling block that is best removed. Finally, removing --maxfail=10 makes it much easier to investigate what fails due to removing @skipIf and related annotations, turning off HIDE_WINDOWS_KNOWN_ERRORS, and the like, because one can see everything that fails. I have retained --disable-warnings, though that should be revisited in the future, since it could be hiding something of value. Delving into that would have been in keeping with the theme of this PR, but I have not done so, in part to limit the scope of the changes (as turning that off would likely lead to the addition of a number of more fine-grained suppressions). The other arguments pertain to code coverage and are fine; code coverage reports are often wanted, especially for full test runs on CI, and they can be turned off easily by passing --no-cov to pytest. It wasn't clear when is_win was true is_win is true only on native Windows systems, but this is not obvious, nor is it in all cases obvious what constitutes a native Windows system. In a couple cases, a test was skipped when is_win and a Cygwin check both hold. Those never happen together--those tests were running on all systems and doing fine (and I removed those @skipIf annotations). is_win, which is provided by GitPython, is true when os.name == "nt". On Cygwin, os.name == "posix". Cygwin can be distinguished from other Unix-like platforms by checking that sys.platform == "cygwin", which also holds on systems like MSYS2 that are derived from Cygwin. However, it does not hold on MinGW builds of Python, which are a native build that is provided with MSYS 2, and which is correctly detected as native Windows with os.name and thus GitPython's is_win. (The "Git Bash" environment also provides MinGW builds, though it does not include Python.) Given this situation, I am unsure if it's really a good idea to have is_win in the project, rather than testing the above conditions directly. However, we do have it, and I believe users are supposed to be able to use it directly (so removing it would be a breaking change). I have certainly not endeavored to remove it here, nor to deprecate it. However, I have extended the part of both of the CI test workflows that output version information such as git and python versions to also output the values of all of the above, so they can be readily checked. Together with being able to see the status of each test by name (by either perusing or searching), I think this should avoid the kind of situation I faced in #1636 and #1650 where I didn't know what platforms my tests were running (or supposed to run) on and couldn't tell by examining CI output. Expected failures were handled by skipping I have not completely addressed this, but I have made significant headway, and the areas where it remains so are either marked or are already known to be especially tricky cases. When the actions a test performs should not be attempted at all, the test should be skipped. However, in testing frameworks with limited support for marking tests as expected to fail, skipping is often used to express this as well. This is best avoided when possible, because it requires that both initiative and manual action be taken to check whether tests that were not working before are still not working, as well as to check that how they are failing remains the same, at least in the general sense of what exception is raised. I replaced most, but not all, @skipIf and similar annotations, in places where they represent expected failure of a test that ought to pass, with @pytest.mark.xfail. Both unittest and pytest provide facilities for expressing expected failure, but unittest.expectedFailure is very limited, not allowing a condition, reason, or expected exception to be passed. In contrast, pytest.mark.xfail supports such arguments (among others), and I have always passed them. The raises argument takes an exception type (or tuple of them) that is expected to be raised, and test failures due to other exceptions will be reported as regular failures. Because tests marked xfail still run (except in limited cases where one prevents this, which I have not done), examining pytest output (when -v or -vv have been passed) reveals, for each test, whether it really did fail as expected (xfail status), or if it unexpectedly passed (xpass status). It is possible to have the xpass status treated as a failure, but I have not done that. An xfailing or xpassing test does not produce full test output with a traceback, so we are still only getting that highly verbose inline failure information (due to pytest-sugar locally and pytest-instafail on CI) if we get an unexpected failure, i.e., one that fails the run. My goal with xfail is that, taken together with other changes in this PR, it should make it unnecessary to take any special action to check whether a test is still unable to pass--or to check whether the system or other condition for its failure, or the way it fails, has remained the same. Instead, one simply takes a look at the xfail output, a key part of the approach articulated there thus being continuously automated. Going along with that, it should be little more obtrusive than the effect of a passing test, so one is not distracted by it. I believe this PR largely achieves both goals, in a way that is worthwhile even though imperfect. But the imperfections are worth noting: I don't believe ddt has its own xfail feature. @pytest.mark.parametrize is made to work together with @pytest.mark.xfail, so generated tests with some arguments can be marked xfail while others generated from the same function/method are not. But @pytest.mark.parametrize cannot be used effectively on a method in a class that inherits directly or indirectly from unittest.TestCase, even if (as in this project) it is only necessary to support the pytest test runner. Most GitPython test classes derive indirectly from TestCase through a custom TestBase class that provides the import rorepo fixture, and I did not want migrate that as part of the work on this PR. But the effect is that there are some xpassing functions already, because I could only mark the whole group (i.e,, defined function) as xfail. The situation is clearly documented with comments and the reason keyword argument. It affects two groups of cygpath tests. test_includes_order contains a few assertions, the last of which is expected to fail, which is handled by placing it in a try block, catching the AssertionError, and reraising it as a SkipTest exception. It is possible to do something analogous to signal that an expected failure has occurred. But without splitting up or otherwise reorganizing the test case, this would not carry the benefit of an xpass status if the assertion unexpectedly starts working. It might still be reasonable to do that, but because I don't want to create the appearance of observability where we don't have it, I have not done so at this point. My hope is that either the test can be reorganized or the underlying problem can be fixed. I have clearly commented the situation. The git module (rather than its unit tests) has three places where unittest.SkipTest is raised from inside it to skip a test (#790). As in test_includes_order, these are situations where, if nothing goes wrong, no indication is given. So for the same reason as there--but also that it would be much harder because I am using xfail from pytest but the git module shouldn't have testing framework dependencies, per #526 and #527--I have not changed that here. Note that, in spite of the connection to the related #525, most occurrences of HIDE_WINDOWS_KNOWN_ERRORS do not straddle git and test in this way, and I have changed then from @skipIf and similar to @pytest.mark.xfail. I have not applied xfail to HIDE_WINDOWS_FREEZE_ERRORS, but that is because I don't think they are expected failures, so I do not class this as an imperfection. (We don't want those tests to run automatically on native Windows systems, but @pytest.mark.xfail supports run=False for such cases, though that forgoes being able to observe an xpass. But I believe these tests usually work on all systems, but freeze occasionally on native Windows. I also think this should be addressed as its own change rather than here; I suspect there is actually a connection to #1676, as I've noted there, and maybe they can be fixed together.) Some expected failures were out of date (or originally erroneous) This is to be expected, and a benefit of the above changes (of xfail and the other changes above it that support seeing and using that output) was to be able to identify these cases and fix the skip and xfail specifications, and most of all to verify that they are fixed. To be clear, I didn't actually fix any bugs in the code under test. The changes to the test code should be evident in the diff (as, if not, then the test code is not as clear as I intend as a result of the changes), but a summary may help in distinguishing these from nearby changes, so in case it helps, here are the main (kinds of) changes of this kind: Removing skips/xfails for tests that had been broken but are now fully working. Fixing a wrong platform (some Cygwin tests were run only on native Windows). Removing skips for an impossible platform (the is_win and Cygwin case mentioned above). Identifying more specific failure conditions (in particular, test_commit_msg_hook_success on native Windows works with some bash interpreters and not others, possibly arising later than #1399). Removing the Cygwin xfail on test_blocking_lock_file, which turned out just to need its time window extended on Cygwin the same as for native Windows. Very old git versions would skip tests needing newer features I think that was the right choice at the time, but now those versions are even older. I've changed the logic to raise RuntimeError to error out those tests as being unable to run if one attempts to run them with extremely old versions of git that don't support them, rather than skipping. Intuitively, people don't expect a bug to be caught if it only affects one platform and they test on another. In contrast, skipping these tests is likely to go unnoticed: just as one does not get a passing test run if one doesn't have any git installed, I think it makes sense for the few specific tests that require a non-ancient git to error out without it, now that new enough versions are so widely available. Some operating systems, such as CentOS 7, that maintain downstream versions of extremely old software for an extremely long time, might have such an old git version that nonetheless has downstream security fixes that make it acceptably safe to use. But I suspect that anyone developing improvements to GitPython on such a system would also have installed a newer git to test with, even if not aware of any specific incompatibilities. The CI test workflows still appeared more different than they are For debugging CI (and also for future efforts in extending it, including in adding native Windows jobs), it is useful to be able to readily identify the fundamental similarities and differences between the two CI test workflows. I had made some progress on this before, but I've taken care of some stuff I missed. Most notably, commands provided by Cygwin are now run with relative paths in the Cygwin workflow (and found), and set -x in the pytest step no longer breaks the tests. There is probably more that could reasonably be done in this area.

Open Graph Description: This makes changes so tests give more informative output, both locally and on CI but especially on CI. It comprises several interrelated changes that should help lead up to, but do not include, add...

X Description: This makes changes so tests give more informative output, both locally and on CI but especially on CI. It comprises several interrelated changes that should help lead up to, but do not include, add...

Opengraph URL: https://github.com/gitpython-developers/GitPython/pull/1679

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width