[Boost-users] [serialization][python] (ab)using serialization to read/write object attribute values from python

Subject: [Boost-users] [serialization][python] (ab)using serialization to read/write object attribute values from python
From: Václav Šmilauer (eudoxos_at_[hidden])
Date: 2009-07-16 04:38:08

Hello boosters,

I am working on http://yade.wikia.com project and we plan to switch to
boost::serialization. boost::python is already used (many thanks for
that fantastic piece of code).

Let me explain the current state first. We use custom serialization
interface: each class registers various attributes of objects; at
serialization, archive of pointers to the attributes associated with
attribute names is created and serialized to XML (and vice versa for
deserialization).

Boost.python is used to wrap c++ objects in python. Each wrapper class
defined __getattr__ and __setattr__; if called, the wrapper:

1. Checks that the underlying c++ object has an associated archive of
attributes created.

2.  Gets/sets the value of the attribute based on its registered name
(because of current limitations, the value is passed to the archive as
string)

The advantage is that there is no need to wrap every single class in
python, just a few top-level classes; and access to all other attributes
of the derived types is automatically provided by the serialization
framework.

My questions are:

1. Do boost::serialization archives keep pointers to attributes, in such
way that it is possible to keep one archive instance per object and use
it to write an attribute continuously, with the change being immediately
propagated to the c++ objects, and ask the archive for attribute value
based on its name?

2. Is it possible (or has it even been done) to define python_iarchive
and python_oarchive that would (de)serialize between python and c++
types (such as std::map to dicitonary, [] to vector etc.)?

I am sorry if those questions are trivial or nonsensical... I am trying
to foresee problems that we could have with boost::serialization. If you
have some proposition how to solve this xml+python attribute access in a
better unified way, don't hesitate to post, please.

(Switching the wrapper to py++ doesn't seem very practical at the
moment, as it works quite nicely the current way; (frequent) adding of a
class or adding/removing attributes doesn't require the wrapper to be
regenerated/recompiled at all)

Regards,

Václav Šmilauer