[BUG]: Failed to pickle objects using protocol 0 from Python 3.9

[fbriol]

Problem description

Serialization of objects using protocol 0 (original "human-readable" protocol) does not work anymore from Python 3.9.

Reproducible example code

core.cpp

#include <pybind11/pybind11.h>

class Foo {
 public:
  explicit Foo() {}

  inline auto getstate() const { return pybind11::tuple(); }

  inline static Foo setstate(pybind11::tuple state) {
    if (state.size() != 0) throw std::runtime_error("Invalid state!");
    return Foo();
  }
};

auto init_module(pybind11::module& m) {
  pybind11::class_<Foo>(m, "Foo")
      .def(pybind11::init<>())
      .def(pybind11::pickle(
          [](const Foo& f) { return f.getstate(); },
          [](pybind11::tuple state) { return Foo::setstate(state); }));
}

PYBIND11_MODULE(core, m) { init_module(m); }

compilation

clang++ -O3 -Wall -shared -std=c++17 -fPIC $(python3 -m pybind11 --includes) core.cpp -o core$(python3-config --extension-suffix)

foo.py

import pickle

import core


def main():
    instance = core.Foo()
    # It always works.
    result = pickle.dumps(instance, protocol=-1)
    # Starting with python 3.9, the protocol=0 throws the error:
    # terminate called after throwing an instance of 'std::runtime_error'
    #   what():  instance allocation failed: new instance has no pybind11-registered base types
    # Aborted (core dumped)
    result = pickle.dumps(instance, protocol=0)


if __name__ == '__main__':
    main()

[henryiii]

Pickle protocol 0 and 1 have never been supported by pybind11. This is why the docs had a warning never to leave the pickle as the "default" on Python 2, since it defaulted to 0 (or was it 1?). This is true both before and after Python 3.9.

See

pybind11/tests/test_pickling.py

Line 16 in 6493f49

data = pickle.dumps(p, 2) # Must use pickle protocol >= 2

or https://pybind11.readthedocs.io/en/stable/advanced/classes.html#pickling-support (can't link to the exact line, it's a bit down in the section).

[fbriol]

I understand. But it's not me who decides to serialize the class using protocol 0, but Dask. When the class has a __call__ method, Dask use the protocol 0. I don't know why, but that's how it is. So Dask uses my object with protocol 0 and the program crashes.
pybind11 may not support the zero protocol, but it will be nice if the program doesn't crash. It took me 8 hours to find the origin of the problem. I think it would be super helpful for users if the program doesn't crash, or displays a clear error message, to fix the code.

Thanks a lot for your help.

[henryiii]

Maybe you could investigate why dask uses protocol 0 - maybe they can use something higher? And why it uses it if you have a __call__? There might be another way to avoid this, by defining a __reduce__ for example.

I'm not really aware of how or why protocol 0-1 are not supported, other than the suspicious feature that they are text based rather than binary.