Python cpp_implementation segfaults during test

[yingzong]

Followed instructions here for OSX:
https://github.com/google/protobuf/tree/master/python

With latest version v3.2.0, got a segfault on this line of the test python setup.py test --cpp_implementation about once every two to four tries. It doesn't always happen, but when it happens it is always on this line.

testMergeFrom (google.protobuf.internal.message_test.Proto3Test) ... Segmentation fault: 11

trace from lldb:

frame #0: 0x0000000104ec52c1 _message.so`google::protobuf::python::cmessage::GetFieldDescriptor(google::protobuf::python::CMessage*, _object*) [inlined] google::protobuf::Message::GetDescriptor(this=0x00007f9132e86ad0) const + 3 at message.h:322
   319 	
   320 	  // Get a Descriptor for this message's type.  This describes what
   321 	  // fields the message contains, the types of those fields, etc.
-> 322 	  const Descriptor* GetDescriptor() const { return GetMetadata().descriptor; }
   323 	
   324 	  // Get the Reflection interface for this Message, which can be used to
   325 	  // read and modify the fields of the Message dynamically (in other words,
(lldb) up
frame #1: 0x0000000104ec52be _message.so`google::protobuf::python::cmessage::GetFieldDescriptor(self=<unavailable>, name=<unavailable>) + 14 at message.cc:986
   983 	// handling is not consistent.
   984 	static const FieldDescriptor* GetFieldDescriptor(
   985 	    CMessage* self, PyObject* name) {
-> 986 	  const Descriptor *message_descriptor = self->message->GetDescriptor();
   987 	  char* field_name;
   988 	  Py_ssize_t size;
   989 	  if (PyString_AsStringAndSize(name, &field_name, &size) < 0) {
(lldb) up
frame #2: 0x0000000104ec5381 _message.so`google::protobuf::python::cmessage::GetAttr(self=0x000000010548ac80, name=0x000000010291db48) + 65 at message.cc:2695
   2692	    return value;
   2693	  }
   2694	
-> 2695	  const FieldDescriptor* field_descriptor = GetFieldDescriptor(self, name);
   2696	  if (field_descriptor == NULL) {
   2697	    return CMessage_Type.tp_base->tp_getattro(
   2698	        reinterpret_cast<PyObject*>(self), name);

[wilstoff]

I am also getting this exact segfault when running tests on our custom objects. I'm running v3.0.2, and we run our python tests with nosetests. Any idea what is going on?

[anandolee]

Can you give more trace information and which line of testMergeFrom() fail? I am unable to reproduce the segfault. Does all other tests always pass if comments out testMergeFrom()?

[anandolee]

I am closing this for clean up and unable to reproduce. Feel free to reopen if anyone can provide more detail of the failure