Code coverage with Bazel

Report an issue View source Nightly · 8.0 · 7.4 · 7.3 · 7.2 · 7.1 · 7.0 · 6.5

Bazel features a coverage sub-command to produce code coverage reports on repositories that can be tested with bazel coverage. Due to the idiosyncrasies of the various language ecosystems, it is not always trivial to make this work for a given project.

This page documents the general process for creating and viewing coverage reports, and also features some language-specific notes for languages whose configuration is well-known. It is best read by first reading the general section, and then reading about the requirements for a specific language. Note also the remote execution section, which requires some additional considerations.

While a lot of customization is possible, this document focuses on producing and consuming lcov reports, which is currently the most well-supported route.

Creating a coverage report

Preparation

The basic workflow for creating coverage reports requires the following:

  • A basic repository with test targets
  • A toolchain with the language-specific code coverage tools installed
  • A correct "instrumentation" configuration

The former two are language-specific and mostly straightforward, however the latter can be more difficult for complex projects.

"Instrumentation" in this case refers to the coverage tools that are used for a specific target. Bazel allows turning this on for a specific subset of files using the --instrumentation_filter flag, which specifies a filter for targets that are tested with the instrumentation enabled. To enable instrumentation for tests, the --instrument_test_targets flag is required.

By default, bazel tries to match the target package(s), and prints the relevant filter as an INFO message.

Running coverage

To produce a coverage report, use bazel coverage --combined_report=lcov [target]. This runs the tests for the target, generating coverage reports in the lcov format for each file.

Once finished, bazel runs an action that collects all the produced coverage files, and merges them into one, which is then finally created under $(bazel info output_path)/_coverage/_coverage_report.dat.

Coverage reports are also produced if tests fail, though note that this does not extend to the failed tests - only passing tests are reported.

Viewing coverage

The coverage report is only output in the non-human-readable lcov format. From this, we can use the genhtml utility (part of the lcov project) to produce a report that can be viewed in a web browser:

genhtml --branch-coverage --output genhtml "$(bazel info output_path)/_coverage/_coverage_report.dat"

Note that genhtml reads the source code as well, to annotate missing coverage in these files. For this to work, it is expected that genhtml is executed in the root of the bazel project.

To view the result, simply open the index.html file produced in the genhtml directory in any web browser.

For further help and information around the genhtml tool, or the lcov coverage format, see the lcov project.

Remote execution

Running with remote test execution currently has a few caveats:

  • The report combination action cannot yet run remotely. This is because Bazel does not consider the coverage output files as part of its graph (see this issue), and can therefore not correctly treat them as inputs to the combination action. To work around this, use --strategy=CoverageReport=local.
    • Note: It may be necessary to specify something like --strategy=CoverageReport=local,remote instead, if Bazel is set up to try local,remote, due to how Bazel resolves strategies.
  • --remote_download_minimal and similar flags can also not be used as a consequence of the former.
  • Bazel will currently fail to create coverage information if tests have been cached previously. To work around this, --nocache_test_results can be set specifically for coverage runs, although this of course incurs a heavy cost in terms of test times.
  • --experimental_split_coverage_postprocessing and --experimental_fetch_all_coverage_outputs
    • Usually coverage is run as part of the test action, and so by default, we don't get all coverage back as outputs of the remote execution by default. These flags override the default and obtain the coverage data. See this issue for more details.

Language-specific configuration

Java

Java should work out-of-the-box with the default configuration. The bazel toolchains contain everything necessary for remote execution, as well, including JUnit.

Python

See the rules_python coverage docs for additional steps needed to enable coverage support in Python.