Kyua 0.12: Parallel execution of tests
Released on November 22nd, 2015.
This is a huge release and marks a major milestone for Kyua as it finally
implements a long-standing feature request: the ability to execute test
cases in parallel. This is a big deal because test cases are rarely
CPU-bound: running them in parallel yields much faster execution times for
large test suites, allowing faster iteration of changes during development.
As an example: the FreeBSD test suite as of this date contains 3285 test
cases. With sequential execution, a full test suite run takes around 12
minutes to complete, whereas on a 4-core machine with a high level of
parallelism it takes a little over 1 minute.
Implementing parallel execution required rewriting most of Kyua's core and
partly explains explains why there has not been a new release for over a
year. The current implementation is purely subprocess-based, which works
but has some limitations and has resulted in a core that is really complex
and difficult to understand. Future versions will investigate the use of
threads instead for a simplified programming model and additional
parallelization possibilities.
- Issue #2: Implemented support to execute test cases in parallel when
invokingkyua test
. Parallel execution is only enabled when the new
parallelism
configuration variable is set to a value greater than1
.
The default behavior is still to run tests sequentially because some test
suites contain test cases with side-effects that might fail when run in
parallel. To resolve this, the new metadata propertyis_exclusive
can
be set totrue
on a test basis to indicate that the test must be run on
its own. - Known regression: Running
kyua debug
on a TAP-based test program does
not currently report the output in real time. The output will only be
displayed once the test program completes. This is a shortcoming of
the new parallel execution engine and will be resolved. - Removed the external C-based testers code in favor of the new built-in
implementations. The new approach feels significantly faster than the
previous one. - Fixed the handling of relative paths in the
fs.*
functions available
inKyuafile
s. All paths are now resolved relative to the location of
the callerKyuafile
.Kyuafile.top
has been updated with these
changes and you should update custom copies of this file with the new
version. - Changed temporary directory creation to always grant search
permissions on temporary directories. This is to prevent potential
problems when running Kyua as root and executing test cases that require
dropping privileges (as they may later be unable to use absolute paths
that point inside their work directory). - The cleanup of work directories does not longer attempt to deal with
mount points. If a test case mounts a file system and forgets to unmount
it, the mount point will be left behind. It is now the responsibility of
the test case to clean after itself. The reasons for this change are
simplicity and clarity: there are many more things that a test case can
do that have side-effects on the system and Kyua cannot protect against
them all, so it is better to just have the test undo anything it might
have done. - Improved
kyua report --verbose
to properly handle environment
variables with continuation lines in them, and fixed the integration
tests for this command to avoid false negatives. - Changed the configuration file format to accept the definition of
unknown variables without declaring them local. The syntax version
number remains at 2. This is to allow configuration files for newer Kyua
versions to work on older Kyua versions, as there is no reason to forbid
this. - Fixed stacktrace gathering with FreeBSD's ancient version of GDB.
GDB 6.1.1 (circa 2004) does not have the-ex
flag so we need to
generate a temporary GDB script and feed it to GDB with-x
instead. - Issue #136: Fixed the XML escaping in the JUnit output so that
non-printable characters are properly handled when they appear in the
process's stdout or stderr. - Issue #141: Improved reporting of errors triggered by sqlite3. In
particular, all error messages are now tagged with their corresponding
database filename and, if they are API-level errors, the name of the
sqlite3 function that caused them. - Issue #144: Improved documentation on the support for custom properties
in the test metadata. - Converted the
INSTALL
,NEWS
, andREADME
distribution documents to
Markdown for better formatting online.