Memory Leak: when python algorithms fail

[Nick Draper]

If a python algorithm fails it appears that it is held in memory together with any shared pointer it holds.

Test case:

from MantidFramework import *
from mantidsimple import *

class SimplePowderDiffraction2(PythonAlgorithm):
    def category(self):
        return "Examples"

    def name(self):
        return "SimplePowderDiffraction"
    def version(self):
        return 2

    def PyInit(self):
        # Input file
        self.declareFileProperty("Filename", "", FileAction.Load)
	#cal file
        self.declareFileProperty("CalFile", "", FileAction.Load)
        # Output workspace to put the transmission histo into
        self.declareWorkspaceProperty("OutputWorkspace", "", Direction.Output)
        
    def PyExec(self):
        file = self.getProperty("Filename")
        calFile = self.getProperty("CalFile")
        
	tempWorkspace = "__tmp_SimplePowderDiffraction"
        
	#Load and Load RAW do now work with executeSubAlgorithms
	Load(file,tempWorkspace)
        tmpWs = mtd[tempWorkspace]
	
	a = self.executeSubAlg(AlignDetectors, tmpWs, tmpWs, CalibrationFile=calFile)
	tmpWs = a._getWorkspaceProperty("OutputWorkspace")
	mtd.sendLogMessage(str(type(tmpWs)))
	mtd.sendLogMessage(Error)
	a = self.executeSubAlg(DiffractionFocussing, tmpWs,"b", GroupingFileName=calFile)
	wspace = a._getWorkspaceProperty("OutputWorkspace")

        self.setProperty("OutputWorkspace", wspace)
	
	#need to delete the temp workspace names
	mtd.deleteWorkspace(tempWorkspace)
        
#############################################################################################

mtd.registerPyAlgorithm(SimplePowderDiffraction2())

[Martyn Gigg]

In [13962]:

Fixes #3407. Ensure that the ref counts of the traceback, exception and value objects are decremented properly when a Python exception is translated into a C++ one. The traceback holds references to the local variables that were present when the exception was raised so it needs to be cleaned up for these objects to have their destructors called.

[Martyn Gigg]

Note to tester: When the error is raised it will look as if there are no workspaces present but in fact the workspace used is hidden with this name '__tmp_SimplePowderDiffraction'. After the bugfix the memory should be freed when DeleteWorkspace('__tmp_SimplePowderDiffraction') is called. Before the bugfix calling delete had no effect.

[Nick Draper]

still leaks on the first run, subsequent runs do not leak.

Error in execution of algorithm SimplePowderDiffraction:
Python error: global name 'Error' is not defined
  at line 33 in '<input>'

reopen to next iteration

[Nick Draper]

testing was on windows 7

[Nick Draper]

Moved to iteration 33 at iteration 32 code freeze

[Nick Draper]

Moved at end of release 2.1

[Nick Draper]

Moved at the end of release 2.2

[Martyn Gigg]

Tested this again after fixing #6779 and it seems like the leak as gone away.

Tester: Take note of comment 2 and use the algorithm in the main description to test.

The algorithm loads a workspace into a hidden workspace name and then an error occurs. The workspace is still in memory and clicking 'File->Clear All Memory' should now remove the workspace from memory properly.

[Peter Peterson]

This defect no longer exists. Very likely due to #6779 which have been on master for the better part of a week.

[Stuart Campbell]

This ticket has been transferred to github issue ​4254