Re: [Boost-users] [serialization] strange assert with wrong cid

From: Sergey Skorniakov (s.skorniakov_at_[hidden])
Date: 2005-07-05 05:51:29

Hello, Robert!
You wrote  on Mon, 4 Jul 2005 08:59:22 -0700:

Thanks, Robert. Your overwiew of versioning should help me to clean up what 
going wrong.

 ??>>> Trying to load an archive with one set of traits from a archive saved
 ??>>> with another set of traits would generally not work without making
 ??>>> special arrangements to read the "old" archives.  That is, saving one
 ??>>> way and loading another way cannot be generally expected to work and
 ??>>> should give an exception.
 ??>>
 ??>> Ok, we are using class versioning to support loading of older
 ??>> versions of classes. Can we remove obsolete classes (serialized by
 ??>> pointer to base) from time to time or this will break reading
 ??>> archives even if no instances of such classes was used during
 ??>> writing?

 RR> I would expect this to break any archives that contain any instances of
 RR> the obsolete classes but not break any that didn't contain the obsolete
 RR> classes. The question arises as to what an "obsolete class" means.  How
 RR> can one be sure it was never written to an archive.  It had
 RR> serialization in it for a reason at one time.  At the very least I
 RR> would be inclined to leave it with an assertion in the serialization
 RR> function.  Hopefully, this would trap any erroneous presumption that
 RR> that an old archive doesn't contain an instance of the obsolete class.

 RR> Note that handling obsolete classes through versioning can sometimes be
 RR> tricky.  For an example take a look at the new shared_ptr serialization
 RR> which inlcludes code to handle the shared_ptr's serialized by v 1.32.
 RR> The ways of doing it are totally different between the two versions.
 RR> So mapping the old serialization to the new one wasn't at all straight
 RR> forward.

 ??>> I see some class names and numeric ids for classes in
 ??>> archives but I'm not sure that these names or ids used for
 ??>> polymorphic loading.

 RR> ids are used when possible.  That is when a type is automatically
 RR> "registered" by its first use.  Otherwise it is identified by its class
 RR> name declared with export.

 ??>> Of course, if ids used, we should not remove any
 ??>> class that marked with BOOST_CLASS_EXPORT macro, but if class names
 ??>> used, I see no reasons why we can't drop obsolete classes.

 RR> well, the archive will contain a name of the obsolete class.  when that
 RR> name is looked up in the table which maps class names to types
 RR> (extended_type_info) a matching type won't be found as the program
 RR> expects and you'll get an error like class id not found - as it seems
 RR> you have done.

With best regards, Sergey.