On 18 Mar 2009, at 14:07, Jason T. Greene wrote:

Manik Surtani wrote:
On 18 Mar 2009, at 13:54, Jason T. Greene wrote:
Manik Surtani wrote:
On 17 Mar 2009, at 20:33, Jason T. Greene wrote:
Brian Stansberry wrote:

However, this sounds like a problem with PFER. If someone calls PFER, I think the original transaction should resync the node snapshot.
How would this be done? AFAIK the application has no control over the data in JBCs transaction context.

The PFER implementation, not the application, would just drop the node from the tx context which invoked pfer. That would mean that any subsequent read would fetch the most current data.
No, that is not correct.  PFER suspends ongoing TXs and runs outside of any TX, to prevent a failure rolling back the TX.  And this is the root of the problem.

"correctness" I think is in the eye of the beholder :)

To me it does not seem correct that i can do

pfer(k, 7)
get(k) == null
The above would only happen if you did:
tx.start() // ensure this in a transactional context
assert get(k) == null // initially empty
pfer(k, 7) // this *always* happens outside of the context of a tx
assert get(k) == null // this still holds true since we initially read this as a null.

Yep

And I would say that this is expected since pfer is clearly documented as an operation that happens outside of the current transactional context.

Of course, we *could* hack it to provide more natural behaviour in *some* cases, by actually tracking any ongoing transaction, suspending it, completing the pfer, and if the pfer is successful remove any cached nodes in the *original* transaction's context so this is refreshed.  But this will only work if the node has not been first modified by the transaction, otherwise you will end up either losing changes made by the transaction, or flushing transaction changes prematurely.

E..g., this case:

get(fqn, k) == null
put(fqn, k1, 1)
pfer(fqn, k2, 2)

If we only deal with some cases (node not modified before pfer) and not others, that would lead to even more confusion in what is and is not expected behaviour.  :-)


Anyway, this pattern has nothing to do with the problem at hand, or the correctness/consistency I discussed, which has to do with handling a remove() on a null entry with repeatable read.

Sure, remove was also broken, which fixing the above would have hidden (since the issues are related). However, that doesn't mean this should not be addressed at some point.

--
Jason T. Greene
JBoss, a division of Red Hat

--
Manik Surtani
Lead, JBoss Cache