LOL - well it IS surprising. Take a look at the documentation for the class serialization trait "Implementation Level". This touches upon a really big issue with a library such as this. Question is which do you do: a) always the right thing b) always the same thing - and maybe emit a warning or error when its not a good idea Choosing b) makes for a transparent system and this is what I usually prefer. Generally I detest hidden features, attributes, which make things look easier but in fact produce surprising side-effects. When tthings fail it means a huge effort trying to get to the bottom of things - often the only recourse is by trial and error. a) is much more popular today. It seems easier and makes a better demo. And its seductive - let's design it so "it just works". If there is one program in my whole life which has tested my sanity its Microsoft Word. (next in line might be bjam). In practice I usually prefer b) but sometimes - like this one - I give myself a pass and slip into a). In this case it was deliberate design decision to "do what the user probably wants if he doesn't otherwise specify it". In spite of my prejudice generally against this kind of thing, I think in this case it has worked to advantage as I have gotten very few complaints and problems about it. There might be a few other places where something similar has been done - but these would also be exeptional cases. So I sympathize with your point of view in general, I just don't think its the correct one in this particular case. Of course, if one is going to do something like run 1 TB through the serializer, he really should carefully read the manual, and carefully consider what he's doing. I wouldn't trust me to have done the right thing . Robert Ramey gast128 wrote:

Thx for the answer. My point is that if you nowhere specify a load/store through pointer, the following code stores just multiple copies:

for (int i = 0; i < 100; ++i) { const SomeStruct s; oa << s; }

but this behavior seems to change if somewhere a store through is performed. This was for my surprising.

Unfortunately, current tracking implementation can also leads to weird version incompatibility. The unlucky combination is object_serializable + track_selectivly. Suppose, someone have code that serizalizes vector<int> to file (by value). After some time, some code that serializes vector<int> as pointer is added to program. Its enough to make old files unreadable (or readable with errors). Sample code: void TestTrack2Err(boost::archive::polymorphic_xml_iarchive &ar) { std::vector<int> *p; ar & BOOST_SERIALIZATION_NVP(p); // saving of vector<int>* is absent intentionaly to simulate different versions of saving and loading code } void TestTrack2() { std::stringstream s; std::vector<int> x; x.push_back(4); { boost::archive::polymorphic_text_oarchive ar(s); ar & BOOST_SERIALIZATION_NVP(x); } x.clear(); { boost::archive::polymorphic_text_iarchive ar(s); ar & BOOST_SERIALIZATION_NVP(x); assert(x.size() == 1); // The vector is empty at this point - unless loading of vector<int>* is commented out } } The consequences of error can vary from noting (especially with xml archives) to program crashes in some more difficult cases. I have recently encountered this error when dealing with a big project (more than 500 serialized types) and have no idea how I can get out of a difficulty. "Robert Ramey" wrote in message news:f9qnba$req$1@sea.gmane.org...

LOL - well it IS surprising.

Take a look at the documentation for the class serialization trait "Implementation Level".

This touches upon a really big issue with a library such as this. Question is which do you do:

a) always the right thing b) always the same thing - and maybe emit a warning or error when its not a good idea

Choosing b) makes for a transparent system and this is what I usually prefer. Generally I detest hidden features, attributes, which make things look easier but in fact produce surprising side-effects. When tthings fail it means a huge effort trying to get to the bottom of things - often the only recourse is by trial and error.

a) is much more popular today. It seems easier and makes a better demo. And its seductive - let's design it so "it just works". If there is one program in my whole life which has tested my sanity its Microsoft Word. (next in line might be bjam).

In practice I usually prefer b) but sometimes - like this one - I give myself a pass and slip into a). In this case it was deliberate design decision to "do what the user probably wants if he doesn't otherwise specify it". In spite of my prejudice generally against this kind of thing, I think in this case it has worked to advantage as I have gotten very few complaints and problems about it.

There might be a few other places where something similar has been done - but these would also be exeptional cases.

So I sympathize with your point of view in general, I just don't think its the correct one in this particular case.

Of course, if one is going to do something like run 1 TB through the serializer, he really should carefully read the manual, and carefully consider what he's doing. I wouldn't trust me to have done the right thing .

Robert Ramey

gast128 wrote:

...

Thx for the answer. My point is that if you nowhere specify a load/store through pointer, the following code stores just multiple copies:

for (int i = 0; i < 100; ++i) { const SomeStruct s; oa << s; }

but this behavior seems to change if somewhere a store through is performed. This was for my surprising.

Robert Ramey wrote:

OK - I see the problem - good example. I don't see an obvious or easy fix. I'll look into it.

Thinking about this some more.

Sergey Skorniakov wrote:

...

Unfortunately, current tracking implementation can also leads to weird version incompatibility. The unlucky combination is object_serializable + track_selectivly. Suppose, someone have code that serizalizes vector<int> to file (by value). After some time, some code that serializes vector<int> as pointer is added to program. Its enough to make old files unreadable (or readable with errors).

This issue was considered in the implemenation. The fact that at type is tracked/untracked is written to the archive when the data is saved. Tracking at load time is determined by the corresponding flag in the archive and NOT the current attribute. Soooo this will have to be looked into. standard library collections are a little different in that they are unversioned. I don't know if that makes a difference, but we'll look into it. Its also very odd the behavior is different depending upon the type of archive. Robert Ramey

...

Sample code:

void TestTrack2Err(boost::archive::polymorphic_xml_iarchive &ar) { std::vector<int> *p; ar & BOOST_SERIALIZATION_NVP(p); // saving of vector<int>* is absent intentionaly to simulate different versions of saving and loading code } void TestTrack2() { std::stringstream s; std::vector<int> x; x.push_back(4); { boost::archive::polymorphic_text_oarchive ar(s); ar & BOOST_SERIALIZATION_NVP(x); } x.clear(); { boost::archive::polymorphic_text_iarchive ar(s); ar & BOOST_SERIALIZATION_NVP(x); assert(x.size() == 1); // The vector is empty at this point - unless loading of vector<int>* is commented out } }

The consequences of error can vary from noting (especially with xml archives) to program crashes in some more difficult cases. I have recently encountered this error when dealing with a big project (more than 500 serialized types) and have no idea how I can get out of a difficulty.

Thanks for looking into this. Sergey Skorniakov wrote:

I had looked into code and found that tracking and version information loaded from archive only if basic_iserializer::class_info returns true:

void basic_iarchive_impl::load_preamble( basic_iarchive & ar, cobject_id & co ){ if(! co.initialized){ if(co.bis_ptr->class_info()){ class_id_optional_type cid; load(ar, cid); // to be thrown away load(ar, co.tracking_level); load(ar, co.file_version); } else{ // override tracking with indicator from class information co.tracking_level = co.bis_ptr->tracking(m_flags); co.file_version = version_type( co.bis_ptr->version() ); } co.initialized = true; } }

So, for all types that has implementation_level less than object_class_info (standard library collections of primitive types has object_serializable implementation_level) information about tracking (and version) are generated in-place depending of presence appropriate instance of pointer_iserializer in code. I think, it is very dangerous solution. May be, it's better to consider track_selectivly as track_always in such a situation.

That would raise howls from those concerned about performance. I'm thinking that tracking behavior shouldn't be tied to implementation level at all.

The different behavior of xml archives explained by diferent implementation of loading object_id - xml archive just returns already parsed attribute (it gives object_id of last loaded object with object_id or uninitialized value), but other archives performs actual reading that leads into various troubles, depending of archive content.

Its the intention that all archives behave alike. I'll look into it. Robert Ramey