Either unschedule all requested tests, or none if it's not possible -
if some of the requested tests have already been processed by the time
the request arrives. It may happen if the worker runs tests faster than
the controller receives and processes status updates. But in this case
maybe it's just better to let the worker keep running.
This is a prerequisite for group/scope support in worksteal scheduler -
so they won't be broken up incorrectly.
This change could break schedulers that use "steal" command. However:
1) worksteal scheduler doesn't need any adjustments.
2) I'm not aware of any external schedulers relying on this command yet.
So I think it's better to keep the protocol simple, not complicate it for
imaginary compatibility with some unknown and likely non-existent
schedulers.
Co-authored-by: Bruno Oliveira <nicoddemus@gmail.com>
Cache execnet gateway info during WorkerController setup for backward
compatibility, in order to avoid a later main_thread_only deadlock
error triggered when pytest-cov calls rinfo after the main thread is
already busy. See pytest-dev/execnet#274 for corresponding test case.
Fixes: 20e3ac774e ("Use execnet main_thread_only execmodel (#1027)")
Use the execnet main_thread_only execmodel so that code which expects
to run in the main thread will just work. This execmodel has been
merged to the execnet master branch via pytest-dev/execnet#243, so this
patch should not be merged until there is a released version of execnet
supporting the main_thread_only execmodel.
Closes#620
The setuptools implementation of editable installs will insert a
placeholder entry into sys.path as part of its magic to register its
custom import mechanism.
These are not real filesystem paths and as such should not be
rewritten to absolute paths.
Fixes issue with license file format to allow security scanners such as Black Duck to correctly identify the license type of this project as MIT.
Co-authored-by: Bruno Oliveira <bruno@soliv.dev>
Currently, a reason to terminate early (e.g. test failure with --exitfail
option set) causes DSession to immediately raise an Interrupt exception.
Subsequent reports generated by the workers, during the shutdown phase, are
discarded.
One consequence is that, for the failing test, teardown and testfinish are
ignored, which prevents corresponding hooks pytest_runtest_logreport and
pytest_runtest_logfinish being executed for the failing test (this problem
covered in #54). The reporting of tests executing in other workers is also left in an
indeterminate state. This can affect other plugin code.
This is a relatively simple fix, which appears to have minimal and, I think,
acceptable impact on text execution behaviour.
The observable differences are differences in what is reported about a test
run. For example, when running, for example with the '-x/--exitfail' option, it
is possible that more than a single test failure is reported. This is because
more than one test did fail before the test run was completely stopped. The
reporting is absolutely correct; and complete. Prior to this change, only a
single failure would have been reported, but because of incomplete and arguably
incorrect reporting.
Hopefully this fixes#907, as seems this is the only
change in #901 which is somehow related.
---------
Co-authored-by: Ronny Pfannschmidt <opensource@ronnypfannschmidt.de>
Maximum number of tests scheduled in one step.
Setting it to 1 will force pytest to send tests to workers one by one -
might be useful for a small number of slow tests.
Larger numbers will allow the scheduler to submit consecutive chunks of tests
to workers - allows reusing fixtures.
Unlimited if not set.
Fixes#855Fixes#255
* Get `-nauto` default from `PYTEST_XDIST_AUTO_NUM_WORKERS`
A few additional tests for existing functionality are added too.
And the ``-n logical`` option is documented.
Fixes: https://github.com/pytest-dev/pytest-xdist/issues/792
* Restore old initial batch distribution logic in LoadScheduling
pytest orders tests for optimal sequential execution - i. e. avoiding
unnecessary setup and teardown of fixtures. So executing tests in consecutive
chunks is important for optimal performance.
Commit 09d79ace35 optimized test distribution for
the corner case, when the number of tests is less than 2 * number of nodes.
At the same time, it made initial test distribution worse for all other cases.
If some tests use some fixture, and these tests fit into the initial batch,
the fixture will be created min(n_tests, n_workers) times, no matter how many
other tests there are. With the old algorithm (before
09d79ace35), if there are enough tests not using
the fixture, the fixture was created only once.
So restore the old behavior for typical cases where the number of tests is
much greater than the number of workers (or, strictly speaking, when there
are at least 2 tests for every node).
In my test suite, where fixtures create Docker containers, this change reduces
total run time by 10-15%.
This is a partial revert of commit 09d79ace35
Co-authored-by: Bruno Oliveira <nicoddemus@gmail.com>
