resmoke.py's FlushThread attempts to reference imported members during Python interpreter shutdown

This seemed to fail most often on Windows (Python 2.7.3) when the Python interpreter was relatively short-lived (~30 seconds). To achieve that, the failpoints and audit tasks were run repeatedly using the following patch: diff --git a/etc/evergreen.yml b/etc/evergreen.yml index 1631460..d08cae8 100644 --- a/etc/evergreen.yml +++ b/etc/evergreen.yml @@ -286,7 +286,15 @@ functions: extra_args="$extra_args --continueOnFailure" fi - ${path_prefix} ${asan_symbolizer} ${enable_lsan} ${rlp_environment} ${python|python} buildscripts/resmoke.py ${resmoke_args} $extra_args ${test_flags} --log=buildlogger --reportFile=report.json + ret=0 + for i in {0..100}; do + ${path_prefix} ${asan_symbolizer} ${enable_lsan} ${rlp_environment} ${python|python} buildscripts/resmoke.py ${resmoke_args} $extra_args ${test_flags} --log=buildlogger --reportFile=report.json + ret=$? + if [ $ret -ne 0 ]; then + break + fi + done + exit $ret "run jstestfuzz": - command: shell.exec The following patch appears to be sufficient to reproduce on a Windows spawn host. diff --git a/buildscripts/resmokelib/logging/flush.py b/buildscripts/resmokelib/logging/flush.py index e49d3d9..0661f60 100644 --- a/buildscripts/resmokelib/logging/flush.py +++ b/buildscripts/resmokelib/logging/flush.py @@ -62,6 +62,7 @@ class _FlushThread(threading.Thread): Flushes and closes all handlers of 'logger'. """ - for handler in logger.handlers: - handler.flush() - handler.close() + # Continually import a module and instantiate a member of it. + while True: + import collections + assert collections.defaultdict() == collections.defaultdict()