The way I see this working is something like (vaguely):

def transaction_middleware(app): def wrapper(environ, start_response): manager = TransactionManager() environ['transaction.manager'] = manager try: app_iter = app(environ, start_response) except: manager.rollback() raise else: manager.commit()

The manager is basically a container of *actual* transactions, and calling rollback or commit on it gets passed on to all the transactions in the manager.

If you don't do anything that needs a transaction (e.g., read-only), you shouldn't put your transaction in the manager.

The last thing that Ian requests here would let me do that to a certain extent:

I'd kind of like a way of getting the "current" WSGI environment. Then one possible implementation of this config-getter is:

def get_dburi(): environ = get_environ() return environ['paste.config']['dburi']

Or it could get it out of another key, of course. And in a non-web context you have a different get_dburi() implementation. The only annoying part is actually figuring out how you get that function, and how you provide that function.

The way they do it in Zope, which is similar in ways to Paste's StackedObjectProxy and paste.registry, is basically something like:

dburi = getUtility('get_dburi')()

Except in Zope they use an interface instead of 'get_dburi'. But I think we should use a string.

We might want to look at PEAK's contextual stuff, as this is basically addressing the same problem. I believe Phillip recently extracted that from PEAK. (And all this config stuff is exactly the same issue as getting the current transaction manager.) Personally I'm not terribly comfortable with paste.registry and StackedObjectProxy, as I feel it pretends to be more transparent than it really is; I prefer something more explicit like getUtility().