Details
-
New Feature
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
From the jdo-dev mailing list
Craig Russell :-
"http://openjpa.apache.org/docs/latest/manual/manual.html#ref_guide_fetch_impl
says that "Even when a direct relation is not eagerly fetched,
OpenJPA selects the foreign key columns and caches the values. This way
when you do traverse the relation, OpenJPA can often find the related
object in its cache, or at least avoid joins when loading the related
object from the database." - that's very kind, but may be we have an
Entity with lots of foreign key columns and I KNOW I only want a list
showing a few attributes, and I KNOW that the direct relations won't be
traversed (say because the Entites are detached, may be). Is there any
way to configure OpenJPA to NOT eagerly fetch direct relations, unless
they are explicitly part of the Fetch Group? Actually what I'd want is
to be able to by default not include foreign key columns, but have a way
to explicitly include them - which is not the same as having the field
as part of the FetchGroup (that will lead to a JOIN, I may just want to
control whether or not to selects the foreign key column).
Currently we don't have the notion in JDO of fetching key values of
relationships, just fetching the targets of these keys.
Would it be useful to consider adding to the JDO specification the
ability to specify fetching the keys for relationships and not the
relationships themselves? This would probably include adding some
metadata to fetch-group and to FetchGroup.
One thing to kick around, is changing the meaning of recursion-
depth==0 so changing:
-1 recurse indefinitely
1 just fetch the target of this field
n fetch the target and recurse n times
0 just fetch the identifier (foreign key value) for the field but not
the field itself
This is a "spare bit" that is currently not particularly useful, since
the specified behavior is identical if "the member is not specified in
the FetchGroup" and "the member is specified with a recursion-depth of
0".
Matthew Adams :-
Seems like a good idea. Could foreign keys be considered the same as basic
fields for the purposes of the default fetch group? That is, fetched by
default unless the user says don't fetch?
I assume that if an object is loaded using a fetch group that doesn't
include foreign keys and the foreign key field is dereferenced, the impl
would go back to the datastore and load the referenced object's foreign key
via the referencing object's key (simple lazy loading). What's the
behavior of a JDO impl that supports change of identity if I load an object
that uses application identity with a fetch group that doesn't include
foreign keys of direct relationships, I then change the identity value of
the referencing object, and then I dereference a direct relationship that
was not loaded? The impl would not have a way to get back to the datastore
to get the referenced object because the referencing object's key has been
changed. It seems to me that a dumb implementation should throw
JDOUserException; a smart impl could cache the original value of the
referencing object's key if changed during the transaction and use the
cached id value to go get the referenced foreign key & object. I could see
this as being a JDO option:
javax.jdo.option.LazilyLoadDirectRelationshipWithoutLoadedForeignKey.
Craig Russell :-
> Seems like a good idea. Could foreign keys be considered the same
> as basic
> fields for the purposes of the default fetch group? That is,
> fetched by
> default unless the user says don't fetch?
That's where I would start. We need to decide whether this is just a
relational database issue and use mapping metadata or if it's a
general datastore issue and use jdo metadata. In other words, we could
put columns into the fetch group or define a new jdo metadata concept
to describe "the many-to-one artifacts that aren't visible as fields
but are required to instantiate relationships".
>
> I assume that if an object is loaded using a fetch group that doesn't
> include foreign keys and the foreign key field is dereferenced, the
> impl
> would go back to the datastore and load the referenced object's
> foreign key
> via the referencing object's key (simple lazy loading).
Actually, what I expect the implementation to do is to load the
referenced object directly using a join of the primary table (where
t0.pk = ?) with the referenced table (and t0.fk = t1.pk). One round
trip to the datastore instead of two.
But a compliant implementation can eagerly load the foreign keys
today, since there's nothing in the spec to prohibit it.
I think this would be a good feature to discuss via JIRA. There are
specification issues as well as technical issues to work through.
> What's the
> behavior of a JDO impl that supports change of identity if I load an
> object
> that uses application identity with a fetch group that doesn't include
> foreign keys of direct relationships, I then change the identity
> value of
> the referencing object, and then I dereference a direct relationship
> that
> was not loaded? The impl would not have a way to get back to the
> datastore
> to get the referenced object because the referencing object's key
> has been
> changed.
I'd expect that the behavior would depend on whether the change of pk
was flushed or not. If not flushed, the original pk is still visible
in the datastore. If flushed, the changed pk is visible (along with
all the consequent fk's that refer to the changed pk).
> It seems to me that a dumb implementation should throw
> JDOUserException; a smart impl could cache the original value of the
> referencing object's key if changed during the transaction and use the
> cached id value to go get the referenced foreign key & object. I
> could see
> this as being a JDO option:
> javax.jdo.option.LazilyLoadDirectRelationshipWithoutLoadedForeignKey.
I think that an impl that supports change of pk should also be able to
handle this case without being explicit about it.