On Sun, 23 Apr 2006 09:19:54 -0700, Robert Ramey wrote
Jeff Garland wrote:
Well, the order of fields only matters in boost::serialization because all the current archive classes read in order from the serialized data.
I did it this way for two reasons: a) it was easier b) it supports arbitrary sized archives - they don't have to be loaded all into memory.
Yep, it's the way most serialization libs work -- nothing wrong with it at all.
I assure you that a serialization archive can be built that doesn't depend directly on order. Essentially the archive would read and parse the data into an intermediate structure (property tree might be good :-) and then responds to the deserialization requests using a mapped based lookup instead of getting the next data from the file.
I'm surprised that no one (that I know of) hasn't done this.
Me too...
The same exact implementation can be used to make a binding to a database buffer derived from an sql command -- this also might not have the fields in the same order as the object being serialized.
Not providing support for this case was one of the reasons that boost::serialization was orginally rejected.
Hmm it still doesn't permit re-ordering of the data items at any level - still it got accepted. Maybe I just outlasted my critics.
Not sure what you mean. Surely if I write MyFancyArchive class I can simply load the data and lookup on the fieldname during the load step, right? I agree this isn't how the current archives work, but I don't believe this design is precluded. Of course it requires nvp style in the object eg: ar & make_nvp("field1", obj.field1); Jeff