[Python-Dev] PyCapsule_Import semantics, relative imports, module names etc.

[Nick Coghlan]

On 27 July 2015 at 01:21, Larry Hastings <larry at hastings.org> wrote:
>
> PyCapsule_Import() is a simple helper function, a slightly-updated analogue
> to PyCObject_Import().  It's not particularly sophisticated, and I'm not
> surprised it's bewildered by complicated scenarios like relative imports in
> subpackages.  For now all I can recommend is that you not try and torture
> PyCapsule_Import().  And, as always... patches welcome.

In this case, there are actually a lot of limitations related to the
fact that extension modules generally have far more limited
information about where they live in the package hierarchy than normal
Python modules do. PEP 489 addressed quite a few of those with
multi-phase initialisation.

> On Sat, Jul 25, 2015 at 1:41 AM, John Dennis <jdennis at redhat.com> wrote:
>>
>> While porting several existing CPython extension modules that form a
>> package to be 2.7 and 3.x compatible the existing PyObject_* API was
>> replaced with PyCapsule_*. This introduced some issues the existing CPython
>> docs are silent on. I'd like clarification on a few issues and wish to raise
>> some questions.
>>
>> 1. Should an extension module name as provided in PyModule_Create (Py3) or
>> Py_InitModule3 (Py2) be fully package qualified or just the module name? I
>> believe it's just the module name (see item 5 below) Yes/No?

Fully qualified is generally better (if you know the ultimate
location), but it's mainly for introspection support, so most things
will work fine even if you set the name to something like "<this is
mostly ignored>".

>> 2. PyCapsule_Import does not adhere to the general import semantics. The
>> module name must be fully qualified, relative imports are not supported.

Correct, as it has no knowledge of the current module name to anchor a
relative import.

>> 3. PyCapsule_Import requires the package (e.g. __init__.py) to import
>> *all* of it's submodules which utilize the PyCapsule mechanism preventing
>> lazy on demand loading. This is because PyCapsule_Import only imports the
>> top level module (e.g. the package). From there it iterates over each of the
>> module names in the module path. However the parent module (e.g. globals)
>> will not contain an attribute for the submodule unless it's already been
>> loaded. If the submodule has not been loaded into the parent
>> PyCapsule_Import throws an error instead of trying to load the submodule.
>> The only apparent solution is for the package to load every possible
>> submodule whether required or not just to avoid a loading error. The
>> inability to load modules on demand seems like a design flaw and change in
>> semantics from the prior use of PyImport_ImportModule in combination with
>> PyObject. [One of the nice features with normal import loading is setting
>> the submodule name in the parent, the fact this step is omitted is what
>> causes PyCapsule_Import to fail unless all submodules are unconditionally
>> loaded). Shouldn't PyCapsule_Import utilize PyImport_ImportModule?

This sounds like it may be a bug in PyCapsule_Import, but I don't know
the capsule API very well myself (I've never had a reason to use it -
all the extension modules I've worked with personally have been
self-contained).

>> 4. Relative imports seem much more useful for cooperating submodules in a
>> package as opposed to fully qualified package names. Being able to import a
>> C_API from the current package (the package I'm a member of) seems much more
>> elegant and robust for cooperating modules but this semantic isn't supported
>> (in fact the leading dot syntax completely confuses PyCapsule_Import, doc
>> should clarify this).

Until PEP 489 (multi-phase initialisation) was implemented for Python
3.5, extension modules didn't know their actual runtime place in the
module hierarchy, so there was no easy way to provide a module name to
the API to anchor relative lookups.

Given PEP 489, it may be feasible to offer a PyCapsule_ImportRelative
for 3.6+, but it would require someone interested in working through
the details of such an API.

>> 5. The requirement that a module specifies it's name as unqualified when
>> it is initializing but then also has to use a fully qualified package name
>> for PyCapsule_New, both of which occur inside the same initialization
>> function seems like an odd inconsistency (documentation clarification would
>> help here). Also, depending on your point of view package names could be
>> considered a deployment/packaging decision, a module obtains it's fully
>> qualified name by virtue of it's position in the filesystem, something at
>> compile time the module will not be aware of, another reason why relative
>> imports make sense. Note the identical comment regarding _Py_PackageContext
>> in  modsupport.c (Py2) and moduleobject.c (Py3) regarding how a module
>> obtains it's fully qualified package name (see item 1).

Yes, these weird limitations were the genesis of Petr Viktorin's
efforts in implementing a new approach to import extension modules for
Python 3.5: https://www.python.org/dev/peps/pep-0489/

While the old single-phase initialisation mechanism is still supported
for backwards compatibility, the new multi-phase initialisation
approach brings the capabilities of extension modules much closer to
their pure Python counterparts.

Cheers,
Nick.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia