git-test is a command-line script for running automated tests against commits in a Git repository. It is especially targeted at developers who like their tests to pass on every commit in a branch, not just the branch tip.
The best way to use git test is to keep a window open in a second linked worktree of your repository, and as often as you like run
git test run master..mybranch
git test will test the commits in the specified range, reporting any failures. The pass/fail results of running tests are also recorded permanently in your repository as Git "notes" (see git-notes(1)).
If a commit in the range has already been tested, then by default git test reports the old results rather than testing it again. This means that you can run the above command over and over as you work, and git test won't repeat tests whose results it already knows. (Of course there are options to allow you to request explicitly that commits be retested.)
The test results are recorded by the tree that was tested, not the commit, so old test results remain valid even across some kinds of commit rewriting:
- If commits are rewritten to change their log messages, authorship, dates, etc., the test results remain valid.
- If consecutive commits are squashed, the results remain valid.
- If a commit is split into two, only the first (partial) commit needs to be tested.
- If some commits deep in a branch are reordered, the test results for commits built on top of the reordered commits often remain valid.
Of course this means that your tests should not depend on things besides the files in the tree. For example, whether your test passes/fails should not depend on the current branch name or commit message.
First define the test that you would like to run; for example,
git test add "make -j8 && make test"
The string that you specify can be an arbitrary command; it is run with sh -c. Its exit code should be 0 if the test passes, or nonzero if it fails. The test definition is stored in your Git config.
By default, git test run tests HEAD:
git test run
(If the working copy is dirty, the test is run anyway but the results are not recorded.)
You can test a range of Git commits with a single command:
git test run commit1..commit2
The test is run against each commit in the range, in order from old to new. If a commit fails the test, git test reports the error and stops with the broken commit checked out. You can also specify individual commits to test:
git test run commit1 commit2 commit3
or test an arbitrary set of commits supplied via standard input:
git rev-list feature1 feature2 ^master | git test run --stdin
You can adjust the verbosity of the output using the --verbosity/-v or --quiet/-q options. Either of these options can be specified multiple times.
You can define multiple tests in a single repository (e.g., cheap vs. expensive tests). Their results are kept separate. By default, the test called default is run, but you can specify a different test to add/run using the --test=<name>/-t <name> option:
git test add "make test"
git test run commit1..commit2
git test add --test=build "make"
git test run --test=build commit1..commit2
If you have flaky tests that occasionally fail for bogus reasons, you might want to re-run the test against a commit even though git test has already recorded a result for that commit. To do so, run git test run with the --force/-f or --retest options.
If you want to forget particular old test results without retesting, run git test run with the --forget option.
If you want to permanently forget all stored results for a particular test (e.g., if something in your environment has changed), run
git test forget-results [--test=<name>]
Normally, git test run stops at the first broken commit that it finds. If you'd prefer for it to continue even after a failure, use the --keep-going/-k option.
To permanently remove a test definition and all of its stored results, run
git test remove [--test=<name>]
General help about git test can be obtained by running
git test help
Help about a particular subcommand can be obtained via either
git test help run
or
git test run --help
git test exports some environment variables which can be used by test
commands. See this file for details.
git test works really well together with git worktree. Keep a second worktree and use it for testing your current branch continuously as you work:
git worktree add --detach ../test HEAD
cd ../test
git test run master..mybranch
The last command can be re-run any time; it only does significant work when something changes on your branch. Plus, with this setup you can continue to work in your main working tree while the tests run.
Because linked worktrees share branches and the git configuration with the main repository, test definitions and test results are visible across all worktrees. So you could even run multiple tests at the same time in multiple linked worktrees.
Requirements:
- A recent Git command-line client
- A Python interpreter.
git testhas been tested with Python versions 2.7 and 3.4. It will probably work with any Python3 version starting with 3.2 (it requiresargparse).
Just put bin/git-test somewhere in your $PATH, adjusting its first line if necessary to invoke the desired Python interpreter properly in your environment.
Some other features that would be nice:
-
Be more consistent about restoring
HEAD.git test runcurrently checks out the branch that you started on when it is finished, but only if all of the tests passed. We need some kind ofgit test resetcommand analogous togit bisect reset. -
git test bisect: rungit bisect runagainst a range of commits, using a configured test as the command thatbisectuses to decide whether a commit is good/bad. -
git test prune: delete notes for obsolete trees. -
Continuous testing mode, where
git testwatches the repository for changes and re-runs itself automatically whenever the commits it is watching change. -
Dependencies between tests; for example:
-
Provide a way to say "if my
fulltest passes, that implies that thebuildtest would also pass". -
Provide a way to run the
buildtest (and record thebuildtest's results) as the first step of thefulltest.
-
-
Allow trees to be marked
skip, if they shouldn't be tested (e.g., due to a known breakage). Perhaps allow the test script to emit a special return code to ask that the commit be markedskip(probably following the convention ofgit bisect run). -
Remember return codes and give them back out if the old result is reused.
-
Add a
git test fix <range>, which starts an interactive rebase, changing the command for the first broken commit from "pick" to "edit". -
Support tests that depend on the commit, not the tree, that they are run against.
git test is released under the GPLv2+ license. Pull requests are welcome at the project's GitHub page, https://github.com/mhagger/git-test
git test has pretty good automated tests, but it undoubtedly still has bugs and rough edges. Use it at your own risk.
Please note that when you tell git test run to test specified commits, it checks those commits out in your working directory. If the tests fail, it leaves the failing commit checked out in a detached HEAD state. This is intentional, so that you can examine the cause of the failure. But it means that if you had changes on your original HEAD that weren't part of any branch, they will now be unreachable.
If you don't know what a detached HEAD state is, please read up on it. Additionally, it is recommended that you run git test in a separate worktree, which is more convenient anyway (see above for instructions). Note that the git worktree command was added in Git release 2.5, so make sure you are using that version of Git or (preferably) newer.
The above considerations don't apply to running git test against HEAD or your current working tree. In other words,
git test run
and
git test run HEAD
don't change the commit that is checked out, and they won't change your working copy to a detached HEAD state.