Can custom TensorFlow user_ops be served through TensorFlow Serving - tensorflow-serving

Question

I have created custom TensorFlow operators in C++ similar to the examples in tensorflow/user/ops/ and they are working fine when used in TensorFlow sessions.
When saving a SavedModel using the operator, the resulting saved_model does contain the operators (at least a cursory inspection of a text protocol buffer of such a model shows that). Trying to serve this with a tensorflow_model_server of course fails at first, since the operator is unknown.
So I proceeded to extend the tensorflow_model_server with an option to specify the user_ops libraries to be loaded beforehand. The relevant code snippet inserted into "main.cc" of the tensorflow_model_server is:
if (librarypath.size() > 0) {
// Load the library.
TF_Status* status = TF_NewStatus();
TF_LoadLibrary(librarypath.c_str(), status);
if (!TF_GetCode(status) == TF_OK) {
string status_msg(TF_Message(status));
std::cout << "Problem loading user_op library " << librarypath << ": " << TF_Message(status);
return -1; }
TF_DeleteStatus(status);
}
Unfortunately, this does not quite work as expected, I get
Problem loading user_op library /usr/lib64/multipolygon_op.so: /usr/lib64 /multipolygon_op.so: undefined symbol: _ZTIN10tensorflow8OpKernelE
This somehow refers to _pywrap_tensorflow_internal.so symbols. Do I need to build the user op library differently or am I just out of luck ?

Answer 1

Ok, after a trying out a number of different venues the answer turns out to be relatively simple:
The tensorflow_model_server is being linked in such a way that is does not provide its own symbols to newly loaded shared libraries. Thus adding "-rdynamic" to the linker options to change that makes everything fall into place:
bazel build --linkopt=-rdynamic //tensorflow_serving/model_servers:tensorflow_model_server