PythonAlgorithm::__del__ method is not called

[Martyn Gigg]

This scripts demonstrates that the Python object __del__ method is never called when it should be, indicating that the objects are not being deleted.

from mantid.kernel import *
from mantid.api import *

OBJECT_COUNT = 0

class ObjectTracker(Algorithm):

    def __init__(self):
        global OBJECT_COUNT
        OBJECT_COUNT += 1
        Algorithm.__init__(self)

    def __del__(self):
        global OBJECT_COUNT
        OBJECT_COUNT -= 1
        print "Called __del__"

    def PyInit(self):
        pass

    def PyExec(self):
        print "PyExec"

####################################################################                                                                          
AlgorithmFactory.subscribe(ObjectTracker)
####################################################################                                                                          

OBJECT_COUNT = 0
for i in range(10):
    alg = AlgorithmManager.createUnmanaged("ObjectTracker")
    print "Global object count=%d" % OBJECT_COUNT
    alg.initialize()
    alg.setLogging(False)
    alg.execute()
    del alg

print "Final count",OBJECT_COUNT

The final count should be zero but it will be 10.

[Martyn Gigg]

Simplify shared_ptr extraction from Python object

Also avoid using the raw Python C API in favour of the boost python wrappers to ensure that the reference counting is dealt with properly Refs #9513

Changeset: c0c99159c8b164ad2acf60f46ffdc9881228221f

[Martyn Gigg]

Fix PythonAlgorithmTraitsTest

The object returned from the factory now has been downcast to the correct type in Python meaning in the tests the methods are bypassing the base class in some cases. Refs #9513

Changeset: f1c325329f0c8043bcc9644a0fc5113a672fd5a4

[Martyn Gigg]

Agreed with Ross that this is okay in 3.3

[Martyn Gigg]

Wrap the boost::python deleter with one that locks the GIL first.

For objects that might be deleted when the GIL is not already held already, e.g. live data or a sub algorithm of another C++ algorithm, this avoids a segfault on later versions of Python 2.7. Refs #9513

Changeset: fd72806fa12fd838bc984d8b0a9100cf27029cae

[Martyn Gigg]

Test script that demonstrates a segfault that happened before the commit in comment #8 was added.

from mantid.kernel import *
from mantid.api import *
from mantid.simpleapi import StartLiveData
import time

class PassThrough(Algorithm):

    def PyInit(self):
        self.declareProperty(MatrixWorkspaceProperty("InputWorkspace", "", Direction.Input))
        self.declareProperty(MatrixWorkspaceProperty("OutputWorkspace", "", Direction.Output))

    def PyExec(self):
        inws = self.getProperty("InputWorkspace").value

        self.setProperty("OutputWorkspace", inws)

####################################################################
AlgorithmFactory.subscribe(PassThrough)
####################################################################

StartLiveData(UpdateEvery=10, Instrument='FakeEventDataListener',
                      StartTime='1990-01-01T00:00:00', AccumulationMethod='Append',
                      PostProcessingAlgorithm='PassThrough', PostProcessingProperties='InputWorkspace=live',
                      AccumulationWorkspace='accum', OutputWorkspace='live')

time.sleep(3)
AlgorithmManager.cancelAll()

[Martyn Gigg]

Branch: bugfix/9513_pyobj_del_not_called

Tester: The script in the original comment should produce a 0 final count and the script in comment:9 should not crash on fedora 20.

[Ross Miller]

Merge remote-tracking branch 'origin/bugfix/9513_pyobj_del_not_called'

Full changeset: ​3513a74ded2813e63c962c80c37e1bf545a65918

[Stuart Campbell]

This ticket has been transferred to github issue ​10356