These changes are listed in decreasing version number order. Note this can be different from a strict chronological order when there are two branches in development at the same time, such as 4.5.x and 5.0.
- Performance: fixed a slow-down with dynamic contexts that appeared in 6.4.3, closing issue 1538. Hopefully this doesn't break the Cython change that fixed issue 972. Thanks to Mathieu Kniewallner for the deep investigative work and comprehensive issue report.
- Added: the debug output file can now be specified with
[run] debug_filein the configuration file. Closes issue 1319. - Typing: all product and test code has type annotations.
- Fix: On Python 3.7, a file with type annotations but no
from __future__ import annotationswould be missing statements in the coverage report. This is now fixed, closing issue 1524.
- Performance: an internal cache of file names was accidentally disabled, resulting in sometimes drastic reductions in performance. This is now fixed, closing issue 1527. Thanks to Ivan Ciuvalschii for the reproducible test case.
- Fix: when using pytest-cov or pytest-xdist, or perhaps both, the combining
step could fail with
assert row is not Noneusing 7.0.2. This was due to a race condition that has always been possible and is still possible. In 7.0.1 and before, the error was silently swallowed by the combining code. Now it will produce a message "Couldn't combine data file" and ignore the data file as it used to do before 7.0.2. Closes issue 1522.
- Fix: when using the
[run] relative_files = Truesetting, a relative[paths]pattern was still being made absolute. This is now fixed, closing issue 1519. - Fix: if Python doesn't provide tomllib, then TOML configuration files can
only be read if coverage.py is installed with the
[toml]extra. Coverage.py will raise an error if TOML support is not installed when it sees your settings are in a .toml file. But it didn't understand that[tools.coverage]was a valid section header, so the error wasn't reported if you used that header, and settings were silently ignored. This is now fixed, closing issue 1516. - Fix: adjusted how decorators are traced on PyPy 7.3.10, fixing issue 1515.
- Fix: the
coverage lcovreport did not properly implement the--fail-under=MINoption. This has been fixed. - Refactor: added many type annotations, including a number of refactorings. This should not affect outward behavior, but they were a bit invasive in some places, so keep your eyes peeled for oddities.
- Refactor: removed the vestigial and long untested support for Jython and IronPython.
- When checking if a file mapping resolved to a file that exists, we weren't considering files in .whl files. This is now fixed, closing issue 1511.
- File pattern rules were too strict, forbidding plus signs and curly braces in directory and file names. This is now fixed, closing issue 1513.
- Unusual Unicode or control characters in source files could prevent reporting. This is now fixed, closing issue 1512.
- The PyPy wheel now installs on PyPy 3.7, 3.8, and 3.9, closing issue 1510.
Nothing new beyond 7.0.0b1.
A number of changes have been made to file path handling, including pattern
matching and path remapping with the [paths] setting (see
:ref:`config_paths`). These changes might affect you, and require you to
update your settings.
(This release includes the changes from 6.6.0b1, since 6.6.0 was never released.)
- Changes to file pattern matching, which might require updating your
configuration:
- Previously,
*would incorrectly match directory separators, making precise matching difficult. This is now fixed, closing issue 1407. - Now
**matches any number of nested directories, including none.
- Previously,
- Improvements to combining data files when using the
:ref:`config_run_relative_files` setting, which might require updating your
configuration:
- During
coverage combine, relative file paths are implicitly combined without needing a[paths]configuration setting. This also fixed issue 991. - A
[paths]setting like*/foowill now matchfoo/bar.pyso that relative file paths can be combined more easily. - The :ref:`config_run_relative_files` setting is properly interpreted in more places, fixing issue 1280.
- During
- When remapping file paths with
[paths], a path will be remapped only if the resulting path exists. The documentation has long said the prefix had to exist, but it was never enforced. This fixes issue 608, improves issue 649, and closes issue 757. - Reporting operations now implicitly use the
[paths]setting to remap file paths within a single data file. Combining multiple files still requires thecoverage combinestep, but this simplifies some single-file situations. Closes issue 1212 and issue 713. - The
coverage reportcommand now has a--format=option. The original style is now--format=text, and is the default.- Using
--format=markdownwill write the table in Markdown format, thanks to Steve Oswald, closing issue 1418. - Using
--format=totalwill write a single total number to the output. This can be useful for making badges or writing status updates.
- Using
- Combining data files with
coverage combinenow hashes the data files to skip files that add no new information. This can reduce the time needed. Many details affect the speed-up, but for coverage.py's own test suite, combining is about 40% faster. Closes issue 1483. - When searching for completely un-executed files, coverage.py uses the
presence of
__init__.pyfiles to determine which directories have source that could have been imported. However, implicit namespace packages don't require__init__.py. A new setting[report] include_namespace_packagestells coverage.py to consider these directories during reporting. Thanks to Felix Horvat for the contribution. Closes issue 1383 and issue 1024. - Fixed environment variable expansion in pyproject.toml files. It was overly broad, causing errors outside of coverage.py settings, as described in issue 1481 and issue 1345. This is now fixed, but in rare cases will require changing your pyproject.toml to quote non-string values that use environment substitution.
- An empty file has a coverage total of 100%, but used to fail with
--fail-under. This has been fixed, closing issue 1470. - The text report table no longer writes out two separator lines if there are no files listed in the table. One is plenty.
- Fixed a mis-measurement of a strange use of wildcard alternatives in match/case statements, closing issue 1421.
- Fixed internal logic that prevented coverage.py from running on implementations other than CPython or PyPy (issue 1474).
- The deprecated
[run] notesetting has been completely removed.
(Note: 6.6.0 final was never released. These changes are part of 7.0.0b1.)
- Changes to file pattern matching, which might require updating your
configuration:
- Previously,
*would incorrectly match directory separators, making precise matching difficult. This is now fixed, closing issue 1407. - Now
**matches any number of nested directories, including none.
- Previously,
- Improvements to combining data files when using the
:ref:`config_run_relative_files` setting:
- During
coverage combine, relative file paths are implicitly combined without needing a[paths]configuration setting. This also fixed issue 991. - A
[paths]setting like*/foowill now matchfoo/bar.pyso that relative file paths can be combined more easily. - The setting is properly interpreted in more places, fixing issue 1280.
- During
- Fixed environment variable expansion in pyproject.toml files. It was overly broad, causing errors outside of coverage.py settings, as described in issue 1481 and issue 1345. This is now fixed, but in rare cases will require changing your pyproject.toml to quote non-string values that use environment substitution.
- Fixed internal logic that prevented coverage.py from running on implementations other than CPython or PyPy (issue 1474).
- The JSON report now includes details of which branches were taken, and which are missing for each file. Thanks, Christoph Blessing. Closes issue 1425.
- Starting with coverage.py 6.2,
classstatements were marked as a branch. This wasn't right, and has been reverted, fixing issue 1449. Note this will very slightly reduce your coverage total if you are measuring branch coverage. - Packaging is now compliant with PEP 517, closing issue 1395.
- A new debug option
--debug=pathmapshows details of the remapping of paths that happens during combine due to the[paths]setting. - Fix an internal problem with caching of invalid Python parsing. Found by OSS-Fuzz, fixing their bug 50381.
- Wheels are now provided for Python 3.11.
- Fix a failure when combining data files if the file names contained glob-like patterns. Thanks, Michael Krebs and Benjamin Schubert.
- Fix a messaging failure when combining Windows data files on a different drive than the current directory, closing issue 1428. Thanks, Lorenzo Micò.
- Fix path calculations when running in the root directory, as you might do in a Docker container. Thanks Arthur Rio.
- Filtering in the HTML report wouldn't work when reloading the index page. This is now fixed. Thanks, Marc Legendre.
- Fix a problem with Cython code measurement, closing issue 972. Thanks, Matus Valo.
- Updated for a small change in Python 3.11.0 beta 4: modules now start with a
line with line number 0, which is ignored. This line cannot be executed, so
coverage totals were thrown off. This line is now ignored by coverage.py,
but this also means that truly empty modules (like
__init__.py) have no lines in them, rather than one phantom line. Fixes issue 1419. - Internal debugging data added to sys.modules is now an actual module, to avoid confusing code that examines everything in sys.modules. Thanks, Yilei Yang.
- Greatly improved performance on PyPy, and other environments that need the pure Python trace function. Thanks, Carl Friedrich Bolz-Tereick (pull 1381 and pull 1388). Slightly improved performance when using the C trace function, as most environments do. Closes issue 1339.
- The conditions for using tomllib from the standard library have been made more precise, so that 3.11 alphas will continue to work. Closes issue 1390.
- A new setting, :ref:`config_run_sigterm`, controls whether a SIGTERM signal
handler is used. In 6.3, the signal handler was always installed, to capture
data at unusual process ends. Unfortunately, this introduced other problems
(see issue 1310). Now the signal handler is only used if you opt-in by
setting
[run] sigterm = true. - Small changes to the HTML report:
- Added links to next and previous file, and more keyboard shortcuts:
[and]for next file and previous file;ufor up to the index; and?to open/close the help panel. Thanks, J. M. F. Tsang. - The time stamp and version are displayed at the top of the report. Thanks, Ammar Askar. Closes issue 1351.
- Added links to next and previous file, and more keyboard shortcuts:
- A new debug option
debug=sqldataadds more detail todebug=sql, logging all the data being written to the database. - Previously, running
coverage report(or any of the reporting commands) in an empty directory would create a .coverage data file. Now they do not, fixing issue 1328. - On Python 3.11, the
[toml]extra no longer installs tomli, instead using tomllib from the standard library. Thanks Shantanu. - In-memory CoverageData objects now properly update(), closing issue 1323.
- Fix: Coverage.py now builds successfully on CPython 3.11 (3.11.0b1) again. Closes issue 1367. Some results for generators may have changed.
- Fix: adapt to pypy3.9's decorator tracing behavior. It now traces function decorators like CPython 3.8: both the @-line and the def-line are traced. Fixes issue 1326.
- Debug: added
pybehaveto the list of :ref:`coverage debug <cmd_debug>` and :ref:`cmd_run_debug` options. - Fix: show an intelligible error message if
--concurrency=multiprocessingis used without a configuration file. Closes issue 1320.
- Fix: deadlocks could occur when terminating processes. Some of these deadlocks (described in issue 1310) are now fixed.
- Fix: a signal handler was being set from multiple threads, causing an error: "ValueError: signal only works in main thread". This is now fixed, closing issue 1312.
- Fix:
--precisionon the command-line was being ignored while considering--fail-under. This is now fixed, thanks to Marcelo Trylesinski. - Fix: releases no longer provide 3.11.0-alpha wheels. Coverage.py uses CPython internal fields which are moving during the alpha phase. Fixes issue 1316.
- Feature: Added the
lcovcommand to generate reports in LCOV format. Thanks, Bradley Burns. Closes issues 587 and 626. - Feature: the coverage data file can now be specified on the command line with
the
--data-fileoption in any command that reads or writes data. This is in addition to the existingCOVERAGE_FILEenvironment variable. Closes issue 624. Thanks, Nikita Bloshchanevich. - Feature: coverage measurement data will now be written when a SIGTERM signal is received by the process. This includes :meth:`Process.terminate <python:multiprocessing.Process.terminate>`, and other ways to terminate a process. Currently this is only on Linux and Mac; Windows is not supported. Fixes issue 1307.
- Dropped support for Python 3.6, which reached end-of-life on 2021-12-23.
- Updated Python 3.11 support to 3.11.0a4, fixing issue 1294.
- Fix: the coverage data file is now created in a more robust way, to avoid problems when multiple processes are trying to write data at once. Fixes issues 1303 and 883.
- Fix: a .gitignore file will only be written into the HTML report output directory if the directory is empty. This should prevent certain unfortunate accidents of writing the file where it is not wanted.
- Releases now have MacOS arm64 wheels for Apple Silicon, fixing issue 1288.
- Feature: Now the
--concurrencysetting can now have a list of values, so that threads and another lightweight threading package can be measured together, such as--concurrency=gevent,thread. Closes issue 1012 and issue 1082. - Fix: A module specified as the
sourcesetting is imported during startup, before the user program imports it. This could cause problems if the rest of the program isn't ready yet. For example, issue 1203 describes a Django setting that is accessed before settings have been configured. Now the early import is wrapped in a try/except so errors then don't stop execution. - Fix: A colon in a decorator expression would cause an exclusion to end too early, preventing the exclusion of the decorated function. This is now fixed.
- Fix: The HTML report now will not overwrite a .gitignore file that already exists in the HTML output directory (follow-on for issue 1244).
- API: The exceptions raised by Coverage.py have been specialized, to provide finer-grained catching of exceptions by third-party code.
- API: Using
suffix=Falsewhen constructing a Coverage object with multiprocessing wouldn't suppress the data file suffix (issue 989). This is now fixed. - Debug: The
coverage debug datacommand will now sniff out combinable data files, and report on all of them. - Debug: The
coverage debugcommand used to accept a number of topics at a time, and show all of them, though this was never documented. This no longer works, to allow for command-line options in the future.
- Python 3.11 is supported (tested with 3.11.0a2). One still-open issue has to do with exits through with-statements.
- Fix: When remapping file paths through the
[paths]setting while combining, the[run] relative_filessetting was ignored, resulting in absolute paths for remapped file names (issue 1147). This is now fixed. - Fix: Complex conditionals over excluded lines could have incorrectly reported a missing branch (issue 1271). This is now fixed.
- Fix: More exceptions are now handled when trying to parse source files for
reporting. Problems that used to terminate coverage.py can now be handled
with
[report] ignore_errors. This helps with plugins failing to read files (django_coverage_plugin issue 78). - Fix: Removed another vestige of jQuery from the source tarball (issue 840).
- Fix: Added a default value for a new-to-6.x argument of an internal class. This unsupported class is being used by coveralls (issue 1273). Although I'd rather not "fix" unsupported interfaces, it's actually nicer with a default value.
- Fix: The sticky header on the HTML report didn't work unless you had branch coverage enabled. This is now fixed: the sticky header works for everyone. (Do people still use coverage without branch measurement!? j/k)
- Fix: When using explicitly declared namespace packages, the "already imported a file that will be measured" warning would be issued (issue 888). This is now fixed.
- Deprecated: The
annotatecommand and theCoverage.annotatefunction will be removed in a future version, unless people let me know that they are using it. Instead, thehtmlcommand gives better-looking (and more accurate) output, and thereport -mcommand will tell you line numbers of missing lines. Please get in touch if you have a reason to useannotateover those better options: [email protected]. - Feature: Coverage now sets an environment variable,
COVERAGE_RUNwhen running your code with thecoverage runcommand. The value is not important, and may change in the future. Closes issue 553. - Feature: The HTML report pages for Python source files now have a sticky header so the file name and controls are always visible.
- Feature: The
xmlandjsoncommands now describe what they wrote where. - Feature: The
html,combine,xml, andjsoncommands all accept a-q/--quietoption to suppress the messages they write to stdout about what they are doing (issue 1254). - Feature: The
htmlcommand writes a.gitignorefile into the HTML output directory, to prevent the report from being committed to git. If you want to commit it, you will need to delete that file. Closes issue 1244. - Feature: Added support for PyPy 3.8.
- Fix: More generated code is now excluded from measurement. Code such as attrs boilerplate, or doctest code, was being measured though the synthetic line numbers meant they were never reported. Once Cython was involved though, the generated .so files were parsed as Python, raising syntax errors, as reported in issue 1160. This is now fixed.
- Fix: When sorting human-readable names, numeric components are sorted correctly: file10.py will appear after file9.py. This applies to file names, module names, environment variables, and test contexts.
- Performance: Branch coverage measurement is faster, though you might only notice on code that is executed many times, such as long-running loops.
- Build: jQuery is no longer used or vendored (issue 840 and issue 1118). Huge thanks to Nils Kattenbeck (septatrix) for the conversion to vanilla JavaScript in pull request 1248.
- Namespace packages being measured weren't properly handled by the new code that ignores third-party packages. If the namespace package was installed, it was ignored as a third-party package. That problem (issue 1231) is now fixed.
- Packages named as "source packages" (with
source, orsource_pkgs, or pytest-cov's--cov) might have been only partially measured. Their top-level statements could be marked as un-executed, because they were imported by coverage.py before measurement began (issue 1232). This is now fixed, but the package will be imported twice, once by coverage.py, then again by your test suite. This could cause problems if importing the package has side effects. - The :meth:`.CoverageData.contexts_by_lineno` method was documented to return a dict, but was returning a defaultdict. Now it returns a plain dict. It also no longer returns negative numbered keys.
- In 6.0, the coverage.py exceptions moved from coverage.misc to coverage.exceptions. These exceptions are not part of the public supported API, CoverageException is. But a number of other third-party packages were importing the exceptions from coverage.misc, so they are now available from there again (issue 1226).
- Changed an internal detail of how tomli is imported, so that tomli can use coverage.py for their own test suite (issue 1228).
- Defend against an obscure possibility under code obfuscation, where a function can have an argument called "self", but no local named "self" (pull request 1210). Thanks, Ben Carlsson.
- The
coverage htmlcommand now prints a message indicating where the HTML report was written. Fixes issue 1195. - The
coverage combinecommand now prints messages indicating each data file being combined. Fixes issue 1105. - The HTML report now includes a sentence about skipped files due to
skip_coveredorskip_emptysettings. Fixes issue 1163. - Unrecognized options in the configuration file are no longer errors. They are now warnings, to ease the use of coverage across versions. Fixes issue 1035.
- Fix handling of exceptions through context managers in Python 3.10. A missing exception is no longer considered a missing branch from the with statement. Fixes issue 1205.
- Fix another rarer instance of "Error binding parameter 0 - probably unsupported type." (issue 1010).
- Creating a directory for the coverage data file now is safer against conflicts when two coverage runs happen simultaneously (pull 1220). Thanks, Clément Pit-Claudel.
- Dropped support for Python 2.7, PyPy 2, and Python 3.5.
- Added support for the Python 3.10
match/casesyntax. - Data collection is now thread-safe. There may have been rare instances of exceptions raised in multi-threaded programs.
- Plugins (like the Django coverage plugin) were generating "Already imported a file that will be measured" warnings about Django itself. These have been fixed, closing issue 1150.
- Warnings generated by coverage.py are now real Python warnings.
- Using
--fail-under=100with coverage near 100% could result in the self-contradictory messagetotal of 100 is less than fail-under=100. This bug (issue 1168) is now fixed. - The
COVERAGE_DEBUG_FILEenvironment variable now acceptsstdoutandstderrto write to those destinations. - TOML parsing now uses the tomli library.
- Some minor changes to usually invisible details of the HTML report:
- Use a modern hash algorithm when fingerprinting, for high-security environments (issue 1189). When generating the HTML report, we save the hash of the data, to avoid regenerating an unchanged HTML page. We used to use MD5 to generate the hash, and now use SHA-3-256. This was never a security concern, but security scanners would notice the MD5 algorithm and raise a false alarm.
- Change how report file names are generated, to avoid leading underscores (issue 1167), to avoid rare file name collisions (issue 584), and to avoid file names becoming too long (issue 580).
Note: 5.6 final was never released. These changes are part of 6.0.
- Third-party packages are now ignored in coverage reporting. This solves a few problems:
- The HTML report uses j/k to move up and down among the highlighted chunks of code. They used to highlight the current chunk, but 5.0 broke that behavior. Now the highlighting is working again.
- The JSON report now includes
percent_covered_display, a string with the total percentage, rounded to the same number of decimal places as the other reports' totals.
coverage combinehas a new option,--keepto keep the original data files after combining them. The default is still to delete the files after they have been combined. This was requested in issue 1108 and implemented in pull request 1110. Thanks, Éric Larivière.- When reporting missing branches in
coverage report, branches aren't reported that jump to missing lines. This adds to the long-standing behavior of not reporting branches from missing lines. Now branches are only reported if both the source and destination lines are executed. Closes both issue 1065 and issue 955. - Minor improvements to the HTML report:
- The state of the line visibility selector buttons is saved in local storage so you don't have to fiddle with them so often, fixing issue 1123.
- It has a little more room for line numbers so that 4-digit numbers work well, fixing issue 1124.
- Improved the error message when combining line and branch data, so that users will be more likely to understand what's happening, closing issue 803.
- The text report produced by
coverage reportnow always outputs a TOTAL line, even if only one Python file is reported. This makes regex parsing of the output easier. Thanks, Judson Neer. This had been requested a number of times (issue 1086, issue 922, issue 732). - The
skip_coveredandskip_emptysettings in the configuration file can now be specified in the[html]section, so that text reports and HTML reports can use separate settings. The HTML report will still use the[report]settings if there isn't a value in the[html]section. Closes issue 1090. - Combining files on Windows across drives now works properly, fixing issue 577. Thanks, Valentin Lab.
- Fix an obscure warning from deep in the _decimal module, as reported in issue 1084.
- Update to support Python 3.10 alphas in progress, including PEP 626: Precise line numbers for debugging and other tools.
- When using
--sourceon a large source tree, v5.x was slower than previous versions. This performance regression is now fixed, closing issue 1037. - Mysterious SQLite errors can happen on PyPy, as reported in issue 1010. An immediate retry seems to fix the problem, although it is an unsatisfying solution.
- The HTML report now saves the sort order in a more widely supported way, fixing issue 986. Thanks, Sebastián Ramírez (pull request 1066).
- The HTML report pages now have a :ref:`Sleepy Snake <sleepy>` favicon.
- Wheels are now provided for manylinux2010, and for PyPy3 (pp36 and pp37).
- Continuous integration has moved from Travis and AppVeyor to GitHub Actions.
- The
sourcesetting has always been interpreted as either a file path or a module, depending on which existed. If both interpretations were valid, it was assumed to be a file path. The newsource_pkgssetting can be used to name a package to disambiguate this case. Thanks, Thomas Grainger. Fixes issue 268. - If a plugin was disabled due to an exception, we used to still try to record its information, causing an exception, as reported in issue 1011. This is now fixed.
The complete history is available in the coverage.py docs.