[BUG] get_backend is not thread-safe about einops

[arogozhnikov]

Hmmm, good point, does wrapping _backends.items() into list solves the problem?

[tran-khoa]

I don't think so. Assume two threads T1 and T2 using the same backend cannot find the backend in the dict.
If T1 then imports the respective backend, T2 may encounter
if BackendSubclass.framework_name not in _backends: therefore not finding the backend that has already been imported by T1.

A simple solution would be to introduce a lock, something like

def get_backend(tensor) -> 'AbstractBackend':
    """
    Takes a correct backend (e.g. numpy backend if tensor is numpy.ndarray) for a tensor.
    If needed, imports package and creates backend
    """
    for framework_name, backend in _backends.items():
        if backend.is_appropriate_type(tensor):
            return backend

    with lock:
        # Try to find backend again
        for framework_name, backend in _backends.items():
            if backend.is_appropriate_type(tensor):
                return backend

        # Find backend subclasses recursively
        backend_subclasses = []
        backends = AbstractBackend.__subclasses__()
        while backends:
            backend = backends.pop()
            backends += backend.__subclasses__()
            backend_subclasses.append(backend)

        for BackendSubclass in backend_subclasses:
            if _debug_importing:
                print('Testing for subclass of ', BackendSubclass)
            if BackendSubclass.framework_name not in _backends:
                # check that module was already imported. Otherwise it can't be imported
                if BackendSubclass.framework_name in sys.modules:
                    if _debug_importing:
                        print('Imported backend for ', BackendSubclass.framework_name)
                    backend = BackendSubclass()
                    _backends[backend.framework_name] = backend
                    if backend.is_appropriate_type(tensor):
                        return backend

    raise RuntimeError('Tensor type unknown to einops {}'.format(type(tensor)))

[arogozhnikov]

If T1 then imports the respective backend, T2 may encounter
if BackendSubclass.framework_name not in _backends: therefore not finding the backend that has already been imported by T1.

True, but not an issue: rest of function is idempontent, and no problem if backend is created twice.