[ZODB-Dev] Re: [Zope-dev] Zope 2.4.0 CopySupport changes seem wrong

[Steve Alexander]

Hi John,

John D.Heintz wrote:
> 
> I'll offer any help I can to you.

Thanks. That's much appreciated.

> I have some ideas to contribute, all of which have been coded into the 
> TransactionManagement.py file include in ZCF 0.5.
> See: http://www.zope.org/Members/jheintz/ZODB_CORBA_Connection/index_html

I'll take a look at this.


>     def attach(self, _id=thread.get_ident, force=0):
>           "map _id() to this transactions for get_transaction(_id)"
> 
>     def detach(self, _id=thread.get_ident):
>           "un-map _id() to this transactions for get_transaction(_id)"

So I guess this is for having many threads to one transaction. I suppose 
that's important for the way CORBA works.


> def registerCommitCallback(self, func, *args, **kw):
>       "Register callback when Transaction has totally finished commit"

Does this run only on a successful finished commit, or does it run on 
anything that causes tpc_finish?


> def commit(self, subtransaction=None):
>     ZODB.Transaction.Transaction.commit(self, subtransaction)
>     if subtransaction is None and self._onCommitCallbacks is not None:
>         for callback, args, kw in self._onCommitCallbacks:
>             try:
>                 apply(callback, args, kw)
>             except:
>                 LOG('ThorTransaction',ERROR, 'Error processing ?'
>                         callbacks',
>                         error=sys.exc_info())
>             self._onCommitCallbacks = None

I guess I can answer my own question -- you only run this on a 
successful commit. What commits the work done in the callbacks?


> Note:  The reason we needed postCommit() and registerCommitCallback() 
> was that some of our _finish() processing was trying to send the ID of 
> a newly created object to an external system that was synchronously 
> calling back into  our ZODB (on another connection) and was unable to 
> find the new object.  This was the side-effect that until all of the 
> finish() calls have been made a callback like this will likely fail.

I had a similar problem with a race condition when I wanted to 
invalidate a RAM Cache Manager after certain things had changed in a 
transaction. I solved it by running the invalidation in a separate 
transaction that runs after the original transaction is completely 
successfully committed.


--
Steve Alexander
Software Engineer
Cat-Box limited