Public Cython APIs#

As of Apr 2020, the following modules in SciPy expose functionality via a public cdef Cython API declarations:

  • scipy.linalg.cython_blas

  • scipy.linalg.cython_lapack

  • scipy.optimize.cython_optimize

  • scipy.special.cython_special

This uses Cython’s declaration sharing features, where shared cdef items are declared in *.pxd files that are distributed together with the corresponding DLL/SO files in binary SciPy installations.

Application Binary Interface#

Using these features in SciPy however requires SciPy contributors to take additional care with regard to maintaining Application Binary Interface (ABI) stability. This is similar to developing libraries in C, and different from how backward compatibility works in pure Python.

The main difference to Python originates from the fact that the declarations in the header .pxd files are used when code written by users is compiled, but they must also match with what is available in SciPy when the user code is imported.

User code may be compiled with one version of SciPy, and the compiled binary (which uses the binary interface declared in the .pxd files) can be used with a different SciPy version installed on the system. If the interfaces are not compatible, either an exception is raised or runtime memory corruption and crash ensue.

At import time, Cython checks that signatures of functions in the installed SciPy SO/DLL file match the one in the .pxd file used by the user during compilation, and raises a Python exception if there is a mismatch. If the SciPy code is structured correctly (see below), this check is performed only for functions that are actually imported in the user code.

We rely on this feature to provide a runtime safety check, which makes it easier for the users to detect incompatible SciPy versions via Python exceptions, instead of hard-to-trace runtime crashes.

ABI stability aim#

SciPy aims to maintain ABI stability in Cython code, in the following sense:

Binaries produced by compiling user source with one version of SciPy, are compatible with any other SciPy version with which the source code can be compiled.

Trying to use an incompatible version of SciPy at runtime will result in a Python exception at user module import time.

Trying to use an incompatible version of SciPy at compile time will result in a Cython error.

This means that users can use any compatible version of SciPy to compile binaries without having to pay attention to ABI, i.e.,

ABI compatibility = API compatibility

Cython API backward/forward compatibility will be handled with a similar deprecation/removal policy as for the Python API, see Deprecations.

Implementing ABI stability in SciPy#

The following rules in development of Cython APIs in SciPy are necessary to maintain the ABI stability aim above:

  • Adding new cdef declarations (functions, structs, types, etc.) is allowed.

  • Removing cdef declarations is allowed, but should follow general deprecation/removal policy.

  • cdef declarations of functions may be changed.

    However, changes result in a backward incompatible API change which breaks any code using the changed signature, and should follow general deprecation/removal policy.

  • cdef declarations of anything else (e.g. struct, enum, and types) are final. Once a declaration is exposed in the public Cython API in a released SciPy version, it must not be changed.

    If changes are necessary, they need to be carried out by adding new declarations with different names, and removing old ones.

  • cdef classes are not allowed in the public APIs (TBD: backward compatibility of cdef classes needs more research, but must not be allowed when we are not sure)

  • For each public API module (as in scipy.linalg.cython_blas), use a single interface .pxd declaration file.

    The public interface declaration file should not contain cimport statements. If it does, Cython’s signature check will check all of the cimported functions, not only the ones that are used by user code, so that changing one of them breaks the whole API.

  • If data structures are necessary, prefer opaque structs in the public API. The interface declarations should not contain any declarations of struct members. Allocation, freeing, and attribute access of data structures should be done with functions.

Deprecating public Cython APIs#

To deprecate a public Cython API function, for example:

# scipy/something/foo.pxd
cdef public int somefunc()

# scipy/something/foo.pyx
cdef public int somefunc():
    return 42

you can add use the scipy._lib.deprecation.deprecate_cython_api function to do the deprecations at the end of the corresponding .pyx file:

# scipy/something/foo.pyx
cdef public int somefunc():
    return 42

from scipy._lib.deprecation import deprecate_cython_api
import as mod
deprecate_cython_api(mod, "somefunc", new_name="scipy.something.newfunc",
                     message="Deprecated in Scipy 1.5.0")
del deprecate_cython_api, mod

After this, Cython modules that cimport somefunc, will emit a DeprecationWarning at import time.

There is no way to deprecate Cython data structures and types. They can be however removed after all functions using them in the API are removed, having gone through the deprecation cycle.

Whole Cython modules can be deprecated similarly as Python modules, by emitting a DeprecationWarning on the top-level.