| Hi Jörg,
What I'm looking for first of all is a description of what replication means to you.
Is it a master/slave configuration where only one of the datastores is updating instances, and you're pushing out updates?
Is bidirectional update supported?
Does it support application identity as well as datastore identity?
Is versioning supported?
Is there a need for user-written callbacks in case of version failure?
It appears from the description of your proposed feature enhancements that you're only covering master/slave, application identity, no versioning, no bidirectional update.
Craig On Mar 10, 2006, at 9:22 AM, Jörg von Frantzius wrote: Hi Craig,
sorry I didn't understand what you wanted to see here exactly. Maybe something like this:
Replication does work perfectly with detaching using an attachCopy() that inserts new instances, and it does so in a productive environment using an older JPOX version, however compliant with the spec that was.
With the current specification, there is one thinkable approach that might work (as pointed out by Matthew), and it has two drawbacks. Here's what it could look like:
Object detachedGraph = pm1.detachCopy(root);
try {
pm2.makePersistent(detachedGraph);
} catch (whatever Exception it is exactly) {
Collection failedInstances = ... // failed are contained in exception
// find out the original instances of the failed detached instances
Collection originalFailedInstances;
for (Object failed: failedInstances) {
Object originalFailed = pm1.getObjectById(JDOHelper.getObjectId(failed));
originalFailedInstances.add(originalFailed);
}
pm1.makeTransientAll(originalFailedInstances);
pm2.makePersistentAll(originalFailedInstances);
// now that the new objects are made persistent, this should succeed:
pm2.makePersistent(detachedGraph);
} The drawbacks are:
- it won't work for datastore identity, as the identities of the instances new to the second PM/datastore are lost upon makeTransient(), and there is no reliable way to assure that they will be assigned the same identities upon makePersistent() on the second PM,
- it certainly is not very effective.
The remedy could be the approach proposed by Marco: to have an additional method
PersistenceManager.makePersistent(Object o, boolean insertNew) and to have PersistenceManager.makePersistent(Object) default to PersistenceManager.makePersistent(o, false).
Regards,
Jörg
Craig L Russell schrieb: Hi Jörg,
As I said earlier, I'd like to see the details of using this feature for replication. What cases are covered, what cases are still problematic, what modifications to the specification are needed in order to accomplish the task?
Craig
On Mar 10, 2006, at 3:14 AM, Jörg von Frantzius wrote:
Hi Craig,
replication really is lost in a specification gap: makePersistent() on transient instances won't update existing data, and on detached instances it won't insert new. For replication, you need both behaviours at the same time.
That's really misfortunate for such a nice feature! Even more so as it is not just theory, but it proves to be working in production with JPOX' old implementation of attachCopy().
Regards,
Jörg
Craig L Russell schrieb:
Hi Jörg,
Using detachment for replication is an interesting use case, and I'd like to see more in-depth analysis of the issues that you encounter once you've done with it.
The use-case for detachment is long-running optimistic transactions, as you have noted below. We did add makeTransient(Object, useFetchPlan) as a way to disconnect objects from one datastore that could be used with another, but I really doubt that we are going to be able to incorporate into the JDO API all the policy algorithms needed by a general-purpose replication scheme.
Craig
On Mar 9, 2006, at 9:48 AM, Jörg von Frantzius wrote:
Craig L Russell schrieb:
Hi Jörg,
There are no tests planned for this behavior.
That's good ;-)
The issue is that it violates the contract of detachment. Detachment is intended to provide a "long-running optimistic transaction" in which conflicts are detected in a subsequent transaction.
I'd find it a little sad if a great feature like easy replication was sacrificed in favor of that. Unless replication should be reserved for JPOX (using a vendor extension), then maybe a future version of the spec could have something along the lines of the solution described by Marco in http://www.jpox.org/servlet/jira/browse/CORE-2741
That would be great.
Just for completeness, and maybe it's just me, but the only sentence about detaching in general that I could find is
"These methods provide a way for an application to identify persistent instances, obtain
copies of these persistent instances, modify the detached instances either in the same JVM
or in a different JVM, apply the changes to the same or different PersistenceManager,
and commit the changes."
It's not really talking about an equivalent to long-running optimistic transactions, I find.
If an instance is detached and then the underlying datastore instance is deleted, this is a consistency violation that should be detected by the transaction semantics. For example, in an order system, if a customer is in a long-running transaction with "groovy beads" in the shopping cart, and the administrators decide that "groovy beads" are no longer to be sold, you want the order that contains "groovy beads" to be rejected when the shopping cart arrives at checkout. You don't want that order to reinsert "groovy beads" into the database.
I agree that this surely must be catered for.
Craig
On Mar 9, 2006, at 8:40 AM, Jörg von Frantzius wrote:
Hi Craig,
I was already afraid that "create a persistent instance" might only apply to the PM cache, not the datastore (but only after second read). However, would you say that JPOX is not JDO2 compliant if it created missing instances in the datastore anyway? Will there be a test in the TCK2 that expects an exception to be thrown if a detached instances does not exist in the datastore?
And, most of all, what sense would it make to forbid the creation of missing detached instances in the datastore? There is lots of application for that behaviour, and at least I don't know of any problem with it.
Regards,
Jörg
Craig L Russell schrieb:
Hi Jörg,
On Mar 9, 2006, at 1:43 AM, Jörg von Frantzius wrote:
Craig L Russell schrieb:
Also I find it confusing that the method most prominently used for inserting new objects shouldn't do so for detached instances.
There is a bunch of history that you should look at, most of which is in the jdo-dev archives. Bottom line, we used to have a different API, attachCopy, but we looked at what it had to do for transient and detached instances and decided that it wasn't worth making a different API for attaching detached instances.
That particular behaviour of attachCopy() wasn't really specified, but it was pleasant JPOX-specific behaviour, if I remember correctly. I saw the discussion and I didn't see where inserting the instances would be forbidden by the spec, and still I don't see where it says that, especially in the light of 12.6.7. Please excuse my ignorance, where does it say that?
<spec>
These methods make transient instances persistent and apply detached instance changes
to the cache.
...
For a detached instance, they locate or create a persistent
instance with the same JDO identity as the detached instance, and merge the persistent
state of the detached instance into the persistent instance. Only the state of persistent fields
is merged.
</spec>
This means that if there is already a persistent instance in the cache with the same object id as the detached instance, the detached state will be merged. If there is not a persistent instance in the cache, a cache instance is created and the detached state is merged with the persistent instance.
But there is no creation aspect of makePersistent on a detached instance.
Craig
Regards,
Craig
Craig
On Mar 8, 2006, at 7:14 AM, Erik Bengtson wrote:
Hi,
What happens when we invoke makePersistent on a detached instance that was
deleted by another isolated process? I suspect that we raise an exception
instead of reinserting it for a second time. Is that right?
Maybe this can be clarified in the spec.
Regards,
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!
|
smime.p7s