[
https://issues.apache.org/jira/browse/OPENJPA-514?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
David Ezzio updated OPENJPA-514:
--------------------------------
Let me try to explain more clearly.
Section 3.3.2 of the JPA 1.0 spec is very clear: "For both transaction-scoped
and extended persistence contexts, transaction rollback causes all pre-existing
managed instances and removed instances to become detached."
Thus on rollback, under any circumstances, the result is a clearing of the
persistence context. This is verified by checking the AutoDetachType
properties of a simply opened EntityManager in a JSE test case. ROLLBACK and
CLOSE are the two settings. In extended contexts with OpenJPA, autodetachment
occurs upon rollback and close, but not upon commit or non-transactional-read.
In a transactionally-scoped context, autodetachment occurs in all four events.
So the notion that there will be any entities around to reload after rollback
is flat out wrong. The OpenJPA documentation at
http://openjpa.apache.org/docs/latest/manual/manual.html#ref_guide_pc_scos_restore
is therefore incorrect.
Also, as the test cases show when you play with the RestoreState property, the
default value appears to be "immutable" rather than "none". Consequently,
either the implementation or the documentation is incorrect on this point. I
don't know which.
Finally, either the documentation or the implementation is incorrect regarding
the implication in the OpenJPA documentation that the RestoreState property
controls the behavior. As the test cases show, the behavior depends at least
in part on whether the fields are eagerly or lazily loaded, and whether an
eagerly loaded field is dirtied before the lazily loaded field is dirtied.
Since I believe that regimentation is a sure path to simplicity (because it
reduces the learning curve), I think the implementation may be in error here as
well as the documentation.
In terms of what the behavior should be, my bias is for useful behavior that
can be explained in the fewest words. So, I think the property should always
control the behavior (no possibility of other factors affecting the outcome),
and I would think that three states would be acceptable: none (as the default,
with perhaps the name changed to "unloaded") resulting in all fields (except ID
fields) being unloaded, a new state "before-image" resulting in a roll back to
the state that existed before the object was first dirtied, and a new state
"no-action" resulting in no action by OpenJPA on the loaded or unloaded state
of the fields. None of these behaviors implies a "reloading" during rollback.
Reloading is probably useless.
> Problems with the openjpa.RestoreState property
> -----------------------------------------------
>
> Key: OPENJPA-514
> URL: https://issues.apache.org/jira/browse/OPENJPA-514
> Project: OpenJPA
> Issue Type: Bug
> Affects Versions: 1.1.0
> Environment: Windows, at OpenJPA revision 619978
> Reporter: David Ezzio
> Attachments: TestCasesOpenJPA-514.zip
>
>
> Section 3.3.2 of the JPA spec is clear that upon rollback, all managed
> instances become detached, and in their detached state they may not be
> consistent or reusable as normally detached instances would be.
> OpenJPA has a property "openjpa.RestoreState" for which it claims a much
> stronger behavior. In section 5.6.1, the OpenJPA manual says:
> "While the JPA specification says that you should not use rolled
> back objects, such objects are perfectly valid in OpenJPA. You can
> control whether the objects' managed state is rolled back to its
> pre-transaction values with the openjpa.RestoreState configuration
> property. none does not roll back state (the object becomes hollow,
> and will re-load its state the next time it is accessed), immutable
> restores immutable values (primitives, primitive wrappers, strings)
> and clears mutable values so that they are reloaded on next access,
> and all restores all managed values to their pre-transaction state."
>
> In section 2.5.53, the OpenJPA manual identifies three values for the
> the openjpa.RestoreState value. It identifies "none" as the default
> value.
> In fact, two central statements in the above cited passages from the
> OpenJPA manual are not true. In addition, the behavior in restoring
> fields is controlled at least in part by other factors.
> To begin with, a detached object will not reload any state. In the
> second instance, the default value appears to be "immutable" instead of
> "none".
> Two test cases were run with the entity class Dohickey. Dohickey has two
> string fields, lazy and eager, as well as an integer identity field.
> Both test cases use a new entity manager. In the first test case, the
> lazy field is dirtied in a newly found a Dohickey. In the second test
> case, the lazy field is dirtied after dirtying the eager field in a
> newly found Dohickey. In both test cases, the transaction is rolled
> back. The detached Dohickey is then tested for detachment, and for
> whether the values of the eager and lazy fields match the values in the
> database.
> Number Test case RestoreState Restored Eager Restored Lazy
> ------+------------+---------------+--------------+--------------
> 1 Lazy only undefined db value dirty value
> 2 Lazy only none null null
> 3 Lazy only immutable db value dirty value
> 4 Eager first undefined db value null
> 5 Eager first none null null
> 6 Eager first immutable db value null
> Test cases 1, 3, 4, and 6 indicate that the default value is likely
> "immutable".
> Test cases 2 and 5 indicate that the "none" value behaves as described
> after correcting the statement that fields will be reloaded.
> The test cases indicate that the behavior for "immutable" setting does
> not work as expected, since the behavior depends on whether the field is
> eagerly or lazily loaded (and perhaps on fetch group configurations as
> well.)
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
