Le mardi 30 octobre 2007 13:36, Stefan Behnel a écrit :
Before using lxml, I was using lxml2, which is also based on libxml2 and libxslt, and I had not this thread problem.
I assume you meant "libxml2", the Python wrapper around the libxml2 C library.
Yes, sorry.
Is this optimisation in your C/python code, or in the original C libxslt code?
We are talking about two optimisations here, one in libxslt and one in lxml. libxslt uses a hash table for XML names to avoid re-allocation of memory. AFAIR, the optimisation is that the dict used for the generated document inherits from the dict of the stylesheet document, which is treated as read-only dictionary fallback.
The optimisation in lxml is that it uses one dictionary per thread, for all documents that are parsed in that thread. So, if you take a stylesheet that was parsed in one thread (and thus depends on the dictionary of that thread), and use it in a different thread that uses a different dictionary, you end up with a result document that tries to mix entries from different dictionaries and will therefore free some of them although they are still referenced in the dictionary of another thread. A sure way to crash your system.
Sadly, the dictionary used by libxslt is not configurable, so all we can do is raise an exception if we detect this problem. I don't really see that as a disadvantage, as it is a fast and safe solution, and you can usually work around the restriction without major hassle.
Ok, I understand. I also prefer a fast XSLT solution ;o) Thanks, -- Frédéric