Pl/Python runtime overhead

Somehow I have failed to find the appropriate keywords for successful results for my question. 

When a pl/python based function is invoked, does it keep a python runtime running across calls to same function? That is, if I use connection pooling, can I save on the python runtime initialization and loading costs? 

Are there any documents/books etc you'd recommend to get a good understanding of extending postgres with languages like python? I'd really like to get a good grip of the architecture of this type of extension, and possibly attempt to introduce a language of my own choosing. The docs I've seen so far are mostly too specific, making it a bit for hard for me to see the forest from the trees.

Regards

Seref