[xml] redicting parts of trees

[Martijn Faassen <faassen infrae com>]

Hi there,

libxml2 documents use dictionaries to hash commonly used strings in XML 
documents, for reasons of space and time efficiency.

Unfortunately this makes it impossible to safely move a node from one 
tree to another.

In the process of writing lxml, an alternative Python wrapper for 
libxml2 that among other things aims at automatic memory management, 
I've so far worked around this issue by sharing dictionaries between all 
trees, whether they be created through reading in XMl, or created from 
scratch.

This works, though with the disadvantage of an endlessly growing 
dictionary as more and more trees are read in, and potential threading 
issues.

I recently discovered another disadvantage: if I understand it 
correctly, XSLT processing, in xsltNewTransformContext, creates a new 
dictionary in its context, with the style's signature as the subdictionary.

The style's signature is based on the style document's signature.

The style document's signature in lxml is of course the single shared 
dictionary, but unfortunately it seems to be unavoidable (except by 
rewriting xsltNewTransformContext itself) to use the shared dictionary 
for the result of XSLT processing. This means that nodes from the tree 
resulting from the XSLT processing cannot be safely moved to other trees 
which do use the single shared dictionary.

Would the developers be open to me suggesting changes to the XSLT 
codebase to make this possible again? I suppose I should ask on the XSLT 
list, so let's move on to the real purpose of this mail.

Exploring these issues made me conclude that it's time to at least look 
at the alternative to sharing a single global dictionary, redicting 
parts of trees. A redicting operation would take place whenever a node 
is moved into a new tree. All the strings in the subtree below this node 
will be traced to the originating document's dictionary, and the entries 
will be copied into the target document's dictionary. Additionally, all 
string references in the subtree will be made to point to the new 
document's dictionary.

Redicting is a potentially expensive operation, but I think that for 
lxml's purposes, it's may be worth it to incur this cost. I also have 
hopes that there are still ways to optimize this, so that redicting 
doesn't have to happen in all cases -- after all, dictionary sharing 
works and pools of documents might end up sharing a single dictionary.

In order to write a good redicting operation, I'd need a bit more 
information about which information in a tree exactly can end up in a 
dictionary. If someone would be able to give me a list of what ends up 
in the dictionary, that would be extremely helpful.

Thank you,

Martijn