OSGi Support in JPA 2.0 Persistence Providers

About three years ago, I started working with Hibernate in an OSGi environment, and at that time, I had to roll my own solution for resolving all sorts of classloader issues.

Since then, with OSGi 4.2 and JPA 2.0, there have been major specification releases, followed by implementations; OSGi is growing ever more popular, and the current release of the JPA standard has integrated a lot of functionality which used to be available in proprietary API extensions only.

The OSGi 4.2 Enterprise Specification includes JPA, JTA and JDBC services, so in theory, using JPA in an OSGi environment should stop being an issue, as soon as this specification is supported by all OSGi and JPA implementations.

In practice, the specifications are not yet fully covered by the available implementations, and the degree of OSGi support varies a lot.

In the past few months, I have worked with the current versions of the three certified JPA 2.0 persistence providers, and in my opinion, the ranking for OSGi support is quite clear:

1. Eclipselink
2. OpenJPA
3. Hibernate

Some more details on each provider:

Eclipselink

Eclipselink has a fully osgified distribution containing OSGi bundles of Eclipselink itself and all dependencies. To make your entity classes visible to Eclipselink, all you need to do is add a JPA-PersistenceUnits manifest header to each bundle containing one or more persistence units. This is enough for Eclipselink to load all your XML metadata and entity classes.

No ugly hacks, no buddy policies or fragments, it just works out of the box.

So for a quick start with JPA under OSGi, Eclipslink is definitely the best choice, even though the OSGi 4.2 JPA Service Specification is not yet fully supported (for instance, the Meta-Persistence header does not work), but this will be of little or no concern to most users.

Leaving OSGi aside, I have been bitten by a number of bugs in Eclipselink, discussed in earlier articles. If you do not use any of these features, you may be very happy with Eclipselink. For me, this set of bugs is currently a no-go, so I have migrated my OSGi applications from Eclipselink to OpenJPA.

OpenJPA

The silver medal goes to OpenJPA: OSGi does not appear to be on the top of the agenda of the OpenJPA core developers, but since OpenJPA is used under OSGi by Aries, another Apache project, there is at least a certain level of OSGi support.

In the binary distribution, the OpenJPA aggregate JAR is an OSGi bundle, but most of the dependencies are plain old JARs, so you need to find osgified versions of the dependencies on your own. The same goes for the Maven artifacts.

Once you have downloaded the appropriate OSGi bundles, you still have to do some extra work to ensure that OpenJPA finds your persistence units and your entity classes, and there are a couple of minor issues with the enhancer and with user-defined value handlers.

I will explain the details of my setup and some of the problems that should be addressed in OpenJPA in my next post.

Hibernate

Sadly, regarding OSGi support, Hibernate 3.5.x is not much different from the 3.2.x release I discussed in my earlier articles. The distribution contains plain old JARs, no OSGi bundles. The latest osgified version available from the SpringSource Enterprise Bundle Repository is 3.4.0, and the SpringSource versions never used to work for me, so I expect that the steps described in my 2008 article are still required to make Hibernate work on OSGi.

Disclaimer: I did not try to make Hibernate 3.5 work on OSGi, due to its lack of support of a number of JPA 2.0 features used heavily by my applications. For this reason, I have stopped using Hibernate, both on OSGi and on Java EE.

JPA 2.0: Ordered Collections

In earlier posts, I wrote about problems with persistent maps, a new feature introduced in JPA 2.0. Ordered collections with explicit order columns are another JPA 2.0 feature which, again, is not yet fully robust in all JPA 2.0 compliant and TCK tested implementations.

I'm currently working with a simple JPA model for OpenStreetMap (OSM). There are ways and nodes, and each way has a sequence of nodes. Of course the order of the nodes is important, and a node may be used more than once for a given way.

Thus, the nodes collection has to be a List, not a Set, and we need an explicit order column specifying the sequence number of the nodes along the way. Sorting the nodes by ID would not make any sense, obviously.

Here is a snippet from the model:

@Entity
@Table(name = "ways")
public class OsmWay
{
    @Id
    private long id;
    
    @ManyToMany(cascade = CascadeType.ALL)
    @OrderColumn(name = "sequence_id")
    @JoinTable(name = "way_nodes", 
               joinColumns = @JoinColumn(name = "id"), 
               inverseJoinColumns = @JoinColumn(name = "node_id"))
    private List<OsmNode> nodes = new ArrayList<OsmNode>();
}    

This entity is represented by two tables:

CREATE TABLE ways
(
  id bigint NOT NULL
);

CREATE TABLE way_nodes
(
  id bigint NOT NULL,
  node_id bigint NOT NULL,
  sequence_id integer,
);

I've tested this scenario on Hibernate 3.5.3, Eclipselink 2.1.1 and OpenJPA 2.0.1: Hibernate and OpenJPA pass, Eclipselink fails with 2 issues.

First, the generated DDL is incorrect: There is a PRIMARY KEY (id, node_id) for way_nodes. This should be (id, sequence_id).

Second, the order of the list items is not maintained in all contexts. It is ok when the collection is lazily loaded, e.g.

OsmWay way = em.find(OsmWay.class, wayId);
List<OsmNode> nodes = way.getNodes();

but it is broken when using a fetch join:

String jpql = "select distinct w from OsmWay w join fetch w.nodes";
TypedQuery<OsmWay> query = em.createQuery(jpql, OsmWay.class);
List<OsmWay> ways = query.getResultList();
OsmWay way = ways.get(0);
List<OsmNode> nodes = way.getNodes();

The funny thing is, the nodes are neither ordered by sequence_id nor by node_id. In my test case, the way has 5 nodes, one of which appears twice in different positions.

So here is some more evidence for my claim that the JPA 2.0 TCK is insufficient.

JPA 2.0: Querying a Map

Welcome back to more merriment with Maps in JPA 2.0!

After watching 3 out of 4 persistence providers choke on a model with a map in the previous post, let us now continue our experiments and see how our guinea pigs can handle JPQL queries for maps.

Recall that the JPQL query language has three special operators for building map queries: KEY(), VALUE() and ENTRY().

Now let us try and run the following query on a slightly modified model, compared to the previous post.

select m.text from MultilingualString s join s.map m where KEY(m) = 'de'

The corresponding model is:

@Embeddable
public class LocalizedString {

    private String language;

    private String text;

} 

@Entity
@Table(schema = "jpa", name = "multilingual_string")
public class MultilingualString {

    @Id
    @GeneratedValue(strategy = GenerationType.SEQUENCE)
    @Column(name = "string_id")
    private long id;

    @ElementCollection(fetch=FetchType.EAGER)
    @MapKeyColumn(name = "language_key")
    @CollectionTable(schema = "jpa", name = "multilingual_string_map", 
                     joinColumns = @JoinColumn(name = "string_id"))
    private Map<String, LocalizedString> map = new HashMap<String, LocalizedString>();
}

This time I've changed the model so that the map key is stored in its own column, which gives Hibernate and Eclipselink at least a chance to digest the model and proceed to the query. OpenJPA is fine with either version of the model.

DataNucleus is out of the game by now. I even tried replacing the @Embeddable by an @Entity and a few other things to cheat it into accepting my model, but in the end I gave up.

Now, Ladies and Gentleman, the winner and sole survivor is: OpenJPA again!

Both Hibernate and Eclipselink fail, merrily throwing exceptions. Hibernate only seems to have stubbed out the KEY() and VALUE() operators in their parser code (see HHH-5396 for the gory details and elaborate stack traces).

And Eclipselink's famous last words are:

Error compiling the query [select m.text from MultilingualString s join s.map m where KEY(m) = 'de'], 
line 1, column 9: unknown state or association field [text] of class [LocalizedString]. 

Not sure what the poor soul is trying to tell me.

To sum up: Should you ever consider working with persistent maps à la JPA 2.0, beware! Here be dragons...

JPA 2.0: Mapping a Map

JPA 2.0 has added support for persistent maps where keys and values may be any combination of basic types, embeddables or entities.

Let's start with a use case:

The Use Case

In an internationalized application, working with plain old Strings is not enough, sometimes you also need to know the language of a string, and given a string in English, you may need to find an equivalent string in German.

So you come up with a LocalizedString, which is nothing but a plain old String together with a language code, and then you build a MultilingualString as a map of language codes to LocalizedStrings. Since you want to reuse LocalizedStrings in other contexts, and you don't need to address them individually, you model them as an embeddable class, not as an entity.

The special thing about this map is that the keys are part of the value. The map contents look like

'de' -> ('de', 'Hallo')
'en' -> ('en', 'Hello')

The Model

This is the resulting model:


[Update 20 July 2010: There is a slight misconception in my model as pointed out by Mike Keith in his first comment on this post. Editing the post in-place would turn the comments meaningless, so I think I'd better leave the original text unchanged and insert a few Editor's Notes. The @MapKey annotation below should be replaced by @MapKeyColumn(name = "language", insertable = false, updatable = false) to make the model JPA 2.0 compliant.]

@Embeddable
public class LocalizedString {

    private String language;

    private String text;

    public LocalizedString() {}

    public LocalizedString(String language, String text) {
        this.language = language;
        this.text = text;
    }
    
    // autogenerated getters and setters, hashCode(), equals()
} 

@Entity
@Table(schema = "jpa", name = "multilingual_string")
public class MultilingualString {

    @Id
    @GeneratedValue(strategy = GenerationType.SEQUENCE)
    @Column(name = "string_id")
    private long id;

    @ElementCollection(fetch=FetchType.EAGER)
    @MapKey(name = "language")
    @CollectionTable(schema = "jpa", name = "multilingual_string_map", 
                     joinColumns = @JoinColumn(name = "string_id"))
    private Map<String, LocalizedString> map = new HashMap<String, LocalizedString>();

    public MultilingualString() {}
    
    public MultilingualString(String lang, String text) {
        addText(lang, text);
    }
    
    public void addText(String lang, String text) {
        map.put(lang, new LocalizedString(lang, text));
    }

    public String getText(String lang) {
        if (map.containsKey(lang)) {
            return map.get(lang).getText();
        }
        return null;
    }
    
    // autogenerated getters and setters, hashCode(), equals()
}

The SQL statements for creating the corresponding tables:

CREATE TABLE jpa.multilingual_string
(
  string_id bigint NOT NULL,
  CONSTRAINT multilingual_string_pkey PRIMARY KEY (string_id)
)

CREATE TABLE jpa.multilingual_string_map
(
  string_id bigint,
  language character varying(255) NOT NULL,
  text character varying(255)
)

The Specification

The most important and most difficult annotation in this example is @MapKey. According to JSR-317, section 2.1.7 Map Keys:

If the map key type is a basic type, the MapKeyColumn annotation can be used to specify the column mapping for the map key. [...]
The MapKey annotation is used to specify the special case where the map key is itself the primary key or a persistent field or property of the entity that is the value of the map.

Unfortunately, in our case it is not quite clear whether we should use @MapKey or @MapKeyColumn to define the table column for our map key. Our map key is a basic type and our map value is not an entity, so this seems to imply we should use @MapKeyColumn.

On the other hand, our key is a persistent field of the map value, and I think the whole point of the @MapKey annotation is to indicate the fact that we simply reuse a property of the map value as the map key, so we do not need to provide an extra table column, as the given property is already mapped to a column.

The way I see it, replacing @MapKey by @MapKeyColumn(name = "language_key") - note the _key suffix! - is also legal, but then we get a different table model and different semantics: The table jpa.multilingual_string_map would have a fourth column language_key, this language_key would not necessarily have to be equal to the language of the map value.

Another open question: Is it legal to write @MapKeyColumn(name = "language")? If so, this should indicate that the language column is to be used as the map key, so this would be equivalent to the @MapKey annotation. On the other hand, you might say that this annotation indicates that the application is free to use map keys that are independent of the map values, so this contract would be violated if the column name indicated by the annotation is already mapped.

The Persistence Providers

I've tried implementing this example with the current versions of Hibernate, Eclipselink, OpenJPA and DataNucleus. I did not succeed with any of them. Only OpenJPA provided a workable solution using @MapKeyColumn, but as I said, I'm not sure if this usage is really intended by the specification.

[Update 20 July 2010: With the corrected model, the updated verdict is: Only OpenJPA passes the test, the other three bail out for various reasons.]

Let's look at the contestants in turn:

Hibernate

Using the mapping defined above, Hibernate 3.5.3-Final complains:

org.hibernate.AnnotationException: Associated class not found: LocalizedString

Apparently Hibernate is expecting the map value to be an entity not an embeddable.

Using @MapKeyColumn(name = "language"), the exception is

org.hibernate.MappingException: Repeated column in mapping for collection: MultilingualString.map column: language

Finally, with @MapKeyColumn(name = "language_key"), Hibernate no longer complains about duplicate columns, but I end up with a redundant table column in my database which I was trying to avoid.

Another problem with Hibernate is different behaviour when working with XML mapping data instead of annotations (which is what I prefer for various reasons, but that's a topic for another post).

Using XML metadata for this example, Hibernate happily ignores the table names from the metadata and simply uses the default names. I filed a bug report in April 2010 (HHH-5136), with no reaction ever since.

Eclipselink

Using Eclipselink 2.1.0, I simply get a rather cryptic exception

java.lang.NullPointerException
 at org.eclipse.persistence.internal.queries.MapContainerPolicy.compareKeys(MapContainerPolicy.java:234)

With @MapKeyColumn=(name = "language"), Eclipselink also complains about a duplicate column, and changing the name to language_key, my test finally passes, at the expense of a redundant column, as with Hibernate.

OpenJPA

With OpenJPA 2.0.0, the message is

org.apache.openjpa.persistence.ArgumentException: Map field "MultilingualString.map" is attempting to use a map table, 
but its key is mapped by another field.  Use an inverse key or join table mapping.

which I can't make sense of. Switching to @MapKeyColumn=(name = "language"), the new message is

org.apache.openjpa.persistence.ArgumentException: 
"LocalizedString.text" declares a column that is not compatible with the expected type "varchar".  

Its seems OpenJPA is confused by the column name text which sounds like a column data type. After adding @Column(name = "_text") to LocalizedString.text, my test case works and my database table only has three columns.

DataNucleus

DataNucleus 2.1.1 complains

javax.persistence.PersistenceException: Persistent class "LocalizedString" has no table in the database, 
but the operation requires it. Please check the specification of the MetaData for this class.

I'm getting the same message with all three variants of the annotation, so it appears that DataNucleus simply cannot handle embeddable map value and expects them to be entities.

Conclusion

Mapping maps with JPA is much harder than you would think, both for the user and for the implementor. Hibernate, Eclipselink and OpenJPA have all passed the JPA TCK. DataNucleus would have liked to do so, but they have not yet been granted access to the TCK.

All four implementors failed this simple map example to various degrees, which implies that there are features in the JPA 2.0 specification which are not sufficiently covered by the TCK.

An Open Source TCK for JPA would help in detecting and eliminating such gaps instead of leaving that to the initiative of individuals.