Monday, August 01, 2005

Week 5: Locks

Locks got off to a good start.

Locks Managing

Unlike dead properties, locks was slightly more complex to model as a shelve dict structure, since its really two structures - locks and urls. Each lock is applied over a number of urls, and one url could have several locks over it.

What I did was to store the information in a slightly redundant way. Each lock stored which urls it locked, and each url stored which locks was applicable over it.

Since shelve sub-dicts (that is, dictionaries and other data structures that are themselves value elements within the shelve dict) are retrieved, manipulated in memory and then written back to the repository on sync, this duplicated structure meant that two concurrent lock operations over the same resource could make the two sub-dicts inconsistent with each other by writing the wrong structure at the wrong time. What I did was to use threading such that:

def writinglockoperation(self, args):

where self._dict is the shelve dict and self._write_lock is a threading.RLock

This does mean no support for concurrency - only one writing Lock operation can be performed at any one time.

Test Status
Is too ugly to put on the blog at the moment. Let's just say that I got what I want for locks coded, except for write locks over null resources and adding new resources to existing collection locks. But there is a lot of debugging to be done.

One big realization coming out of this is how important tests are. Initially I was writing code that I knew worked and letting the tests verify what I knew (something like a rubberstamp), but as the project goes along I am increasingly relying on tests to catch what I should have caught in coding.

One thing for sure, I could not have made such progress without the litmus test suite. Having to write out, code and conduct the tests alone would have taken me about as much time as I spend on the server.

No comments: