Continuous Enterprise Development in Java (2014)
Chapter 6. NoSQL: Data Grids and Graph Databases
I’m more of an adventurous type than a relationship type.
— Bob Dylan
Until relatively recently, the relational database (RDBMS) reigned over data in enterprise applications by a wide margin when contrasted with other approaches. RDBMSs follow a premise of storing heavily normalized data in relational structures such as tables, and are heavily based on rigorous mathematical set theory. Commercial offerings from Oracle and established open source projects like MySQL (reborn as MariaDB) and PostgreSQL became de facto choices when it came to storing, archiving, retrieving, and searching for data. In retrospect, it’s shocking that given the varying requirements from those operations, one solution was so widely lauded for so long.
In the late 2000s, a trend away from the strict ACID transactional properties could be clearly observed, given the emergence of data stores that organized information differently from the traditional row-based table model. In addition, many programmers were beginning to advocate for a release from strict transactions; in many use cases it appeared that this level of isolation wasn’t enough of a priority to warrant the computational expense necessary to provide ACID guarantees, and the often severe performance penalties it imposed on storage systems distributed across more than one server.
In 2009, Amazon’s Werner Vogels published Eventually Consistent, an article that advocated for the tolerance of inconsistent data in large, distributed systems. He argued that this was central to providing a system that could continue to perform effectively under load and could withstand times when part of a distributed data model was effectively unavailable for use. In contrast with the rigid ACID properties, this system could be described as BASE (Basically Available, Soft state, Eventual consistency).
The tenets of eventual consistency are as the name implies: in a distributed system, updates to a data item will, over time, be reflected in all nodes. There is no guarantee of when, or that this will happen immediately, so it’s possible that disparate nodes in a replicated database may not all be in sync at any given point in time. Vogels argued that this was a condition to embrace, not fear.
Thus began the rise of the newly coined NoSQL systems, an umbrella term that encompasses, very generally, data solutions that do not adhere to the ACID properties or store information in a relational format.
It’s important to remember that as developers we’re not bound to any set of solutions before we’ve fully analyzed the problem. The overwhelming dominance of RDBMS is a clear sign that as an industry we stopped focusing on the operations needed by our data model, and instead were throwing all persistent storage features into solutions that perhaps were not the most efficient at solving a particular storage or querying problem.
Let’s have a look at a few scenarios in which an RDBMS might not be the best-suited solution.
RDBMS: Bad at Binary Data
An RDBMS is absolutely capable of storing binary data of arbitrary type. Typically done in a column of type BLOB (Binary Large OBject), CLOB (Character Large OBject), or some related variant, these fields can hold images, PDFs, or anything else we’d like to persist. There are a few caveats, however.
Most RDBMS engines have mechanisms to ensure that queries are executed quickly. One device is a query cache, a temporary location held in memory that remembers the result of recently or often-executed queries and holds them until the data changes, invalidating the cached result and evicting the result from the cache. This space is precious and valuable; it’s typically limited by the amount of RAM available and configuration parameters supported by the vendor. When we add BLOB data into a query and it makes its way into the cache, this very quickly fills up space that’s better used for holding references or other small bits of useful data. When a query is not available in the cache or an index is not available, a full table scan must be performed to attain the result. It’s therefore best to ensure that our large bits of data stay clear of the cache.
The other issue with a traditional RDMBS is that its adherence to the ACID properties is, in most cases, overkill for fetching documents. It’s entirely possible (and probable) in a distributed environment with many database nodes that the user can see data that’s not entirely up-to-date. Consider the case of Twitter or Facebook; seeing the newest tweets or status updates in your feed is not a request that demands completely up-to-date information. Eventually you’ll want to catch all the posts of your friends, but this doesn’t need to be available to you immediately after the update is posted to the server. To provide complete consistency across the database implies that there is locking and blocking taking place; other writes and reads in this concurrent environment would have to wait until the new update is fully committed. Before long, we’d have requests queueing up at a rate likely to exceed that at which the writes could be committed, and the system could grind to a standstill. The mark of a functioning concurrent environment is to avoid blocking wherever possible.
Data Grids
In an era where big data is becoming more and more relevant, we’re faced with the problem of scaling. Our systems are continually asked to store and query upon larger and larger data sets, and one machine is unlikely to be able to handle the load for a non-trivial, public application.
Traditional RDBMS implementations typically offer one or two approaches. First, replication involves a write-only master instance where a single machine is knighted as the authoritative instance. Slaves then pull data from the master’s write operations and replay those writes locally, thus being able to serve read requests. This works well in an application where there is a low write-to-read ratio; the reads will scale out to new slaves while the writes remain centralized on the master.
Second, clustering is an option where a number of database instances keep state current over the network. This is generally a preferred approach in a write-heavy environment where scaling out only the read operations is unlikely to provide much performance benefit. Full clustering has significant overhead, however, so the costs should be weighed carefully.
Data grids work a little bit differently. They’re designed to store pieces of data across available nodes; each node cannot contain the entire data set. Because of this arrangement, they’re built to scale out by simply adding more nodes to the network. Configuration is typically available to control the amount of redundancy; should each piece of data live on two nodes, three, or four? If a node goes down, the system is built to re-distribute the information contained in the now-offline node. This makes data grids especially fault-tolerant and elastic; nodes can be provisioned at runtime. The key to a data grid lies in its capability to partition the data and distribute it across nodes on the network.
Infinispan
Infinispan is an open source data grid from the JBoss Community. Its API centers around a Cache, a Map structure that provides a mapping between a key and value (this does not actually extend the java.util.Map interface from the Java Collections library). The full javax.cache API is defined by the Java Community Process in JSR-107, and although Infinispan is not a direct implementor of this specification, it contains many ideas adjacent to and inspiring JSR-107 and the newer Data Grids Specification JSR-347. Infinispan bills itself as a “Transactional in-memory key/value NoSQL datastore & Data Grid.”
NOTE
Readers keen on gaining better insight into Infinispan and data grids are urged to check out Infinispan Data Grid Platform by Francesco Marchioni and Manik Surtani (Packt Publishing, 2012). A guide to the Infinispan User API is located at http://bit.ly/MB3ree.
The general idea behind Infinispan is that it aims to provide unlimited heap to keep all objects available in memory. For instance, if we have an environment of 50 servers each with 2GB of available RAM, the total heap size is 100GB. When the data set becomes too large for the currently deployed nodes, new memory can be added in the form of new nodes. This keeps all data quickly accessible, though it may be partitioned across nodes. This is not an issue, because every item remains accessible from every node in the grid, even though the current node may not be the one storing the data.
This makes Infinispan well-suited to holding large objects that may have otherwise been baked into a traditional RDBMS model. Its implementation is versatile enough to be used in other applications (for instance, as a local cache), but for the purposes of our GeekSeek application we’ll be leveraging it to handle the storage and retrieval of our binary data.
RDBMS: Bad at Relationships
The greatest irony of the relational database management system is that it starts to break down when we need to model complex relationships. As we’ve seen, a traditional RDBMS will associate data types by drawing pointers between tables using foreign-key relationships, as shown inFigure 6-1.
Figure 6-1. Foreign-key relationships in an RDBMS
When we go to query for these relationships, we often perform an operation called an SQL JOIN to return a result of rows from two or more tables. This process involves resolving any foreign keys to provide a denormalized view; one that combines all relevant data into a form that’s useful for us as humans to interpret, but might not be the most efficient for storage or searching purposes.
The problem is that running joins between data types is an inherently expensive operation, and often we need to join more than two tables. Sometimes we even need to join results. Consider the following example, which has now become a commonplace feature in social media.
Andy has a set of friends. His friends also each have a set of friends. To find all of the people who are friends with both Andy and his friends, we might do something like:
§ Find all of Andy’s friends.
§ For each of those friends, find their friends (third-degree friends).
§ For each of the third-degree friends, determine who is also friends directly with Andy.
That amounts to a lot of querying and joining. What makes this approach unworkable from a computer science standpoint is the use of the term for each, which indicates a loop. The preceding example has two of these, creating a computational problem with geometric complexity at best. As the size of the friend network increases linearly, the time it will take to determine a result increases by factors of magnitude. Eventually (and it doesn’t take a very large social network size), our system will be unable to perform these calculations in a reasonable amount of time, if at all.
Additionally, the approach outlined in the preceding example will need to either search entire tables for the correct foreign-key relationships or maintain a separate index for each type of query. Indexing adds some overhead to write operations; whenever a row is updated or added, the index must reflect that. And working devoid of an index will require the database to do a full table scan. If the size of the table is large enough that it cannot be contained in memory (RAM) or the query cannot be held in a cache, now we face another serious roadblock, because the system must resort to reading from physical disk, which is a far slower undertaking.
When it comes to complex relationships involving tables of any substantial size, the classic RDBMS approach is simply not the most intelligent way to model these resources.
Graph Theory
The preceding problem illustrates that we’re simply using the wrong tool for the job. An RDBMS excels at storage of tabular data, and even does a passable job of drawing simple relationships.
What we want to do here is easily explore transitive relationships without a geometric complexity problem, so we need to tackle the problem from a different angle. Students of computer science will remember studying various data structures and their strengths and weaknesses. In this case, we benefit from turning to the writings of mathematician Leonhard Euler on the Seven Bridges of Königsberg, which in 1735 established the roots of graph theory.
Graphs are data structures comprised from nodes or vertices and edges; the node/vertex represents our data, while the edge defines the relationship.
Using this view of our data points and the relationships between them, we can apply much more efficient algorithms for:
§ Calculating the shortest distance between two nodes
§ Determining a path from one node to another
§ Finding subgraphs and intersections based on query criteria
We’ll be using a graph database to represent some of the relationships between the data held in our RDBMS; we can think of this as a “relationship layer” atop our pure data storage model.
Neo4j
Neo4j is is an open source, transactional graph database that does adhere to the ACID properties. Both its user view and its backing storage engine use underlying graph structures, so it achieves the performance we’d expect from applying graph theory to queries it’s suited to serve. Because of this, the Neo4j documentation touts performance one thousand times faster than possible with an RDBMS for connected data problems.
NOTE
For those looking to understand graph databases and Neo4j in greater detail, we recommend Graph Databases by Robinson, Webber, and Eifrém (O’Reilly, 2013).
Because our GeekSeek application has a social component (who is attending which conferences, who is following speakers and attendees, etc.), we’d like to put in place a solution that will enable us to augment the data in our RDBMS to:
§ Draw relationships between data unrelated in the RDBMS schema
§ Quickly query recursive relationships
§ Efficiently seek out information relevant to users based on relationship data
Use Cases and Requirements
We’ve already seen the domain model for our GeekSeek application in the previous chapter; this encompasses all of our Conference, Session, User, and Venue entities. The link between Conference and Session is fairly restricted, so we use an RBDMS relationship to handle this.
We’d also like to be able to introduce the notion of an Attachment; this can be any bit of supporting documentation that may be associated with a Conference or Session. Therefore we have the requirement:
As a User I should be able to Add/Change/Delete an Attachment.
Because the Attachment is binary data (perhaps a PDF, .doc, or other related material), we’ll store these in a data grid backend using Infinispan.
Additionally, we’d like to introduce some relationships atop our existing data model.
Adding an Attachment is wonderful, but it won’t have much utility for us unless we somehow associate this information with the entity it represents. Therefore, we have the requirement:
As a User I should be able to Add/Delete an Attachment to a Conference.
As a User I should be able to Add/Delete an Attachment to a Session.
A User may attend or speak at a Conference, and it’ll be useful to see who might be nearby while we’re at the show. So we also have the general requirement:
As a User I should be able to SPEAK at a Conference.
As a User I should be able to ATTEND a Conference.
Because this represents a potentially recursive situation (“I want to see all the attendees at conferences in which I’m a speaker”), we’d be smart to use a graph structure to model these ties.
Implementation
We’ll be introducing two domain objects here that are not reflected in our relational model: Attachment and Relation.
Attachment
We’ll start by introducing the model for our Attachment. Because this will not be stored in our RDBMS engine, we’ll create a value object to hold the data describing this entity, and it will not be an @Entity under the management of JPA. We can accomplish this by making a simple class to hold our fields, org.cedj.geekseek.domain.attachment.model.Attachment:
public class Attachment implements Identifiable, Timestampable, Serializable {
private static final long serialVersionUID = 1L;
private final String id;
private final String title;
private final String mimeType;
private final URL url;
private final Date created;
private final Date updated;
This class declaration will adhere to the contracts we’ve seen before in Identifiable and Timestampable, and it has no JPA annotations or metadata because we’ll be delegating the persistent operations of this class to Infinispan.
We should also be sure that these Attachment objects are in a valid state, so we’ll add some assertion checks and intelligent defaults along the way:
public Attachment(String title, String mimeType, URL url) {
this(UUID.randomUUID().toString(),
title, mimeType, url, new Date());
}
private Attachment(String id, String title, String mimeType,
URL url, Date created) {
requireNonNull(title, "Title must be specified)");
requireNonNull(mimeType, "MimeType must be specified)");
requireNonNull(url, "Url must be specified)");
this.id = id;
this.created = created;
this.updated = new Date();
this.title = title;
this.mimeType = mimeType;
this.url = url;
}
@Override
public String getId() {
return id;
}
public String getTitle() {
return title;
}
public Attachment setTitle(String title) {
return new Attachment(this.id, title, this.mimeType, this.url,
this.created);
}
public String getMimeType() {
return mimeType;
}
public Attachment setMimeType(String mimeType) {
return new Attachment(this.id, this.title, mimeType, this.url,
this.created);
}
public URL getUrl() {
return url;
}
public Attachment setUrl(URL url) {
return new Attachment(this.id, this.title, this.mimeType, url,
this.created);
}
public Date getLastUpdated() {
return updated == null ? null:(Date)updated.clone();
}
@Override
public Date getCreated() {
return created == null ? null:(Date)created.clone();
}
@Override
public Date getLastModified() {
return getLastUpdated() == null ? getCreated():getLastUpdated();
}
}
Of note are the calls to our updated method, which will set the timestamp to the current time on any state change operation.
Recall that our persistence layer for objects, whether through JPA or other means, operates through the Repository abstraction; this provides hooks for all CRUD operations. The previous chapter illustrated a Repository backed by JPA and the EntityManager, but because we’ll be storing Attachment objects in a data grid, we need an implementation that will delegate those operations to Infinispan. org.cedj.geekseek.domain.attachment.AttachmentRepository handles this for us:
@Stateless
@LocalBean
@Typed(AttachmentRepository.class)
@TransactionAttribute(TransactionAttributeType.REQUIRED)
public class AttachmentRepository implements Repository<Attachment> {
We’re implementing this AttachmentRepository as a Stateless Session EJB, where all business methods are executed inside the context of a transaction. If a transaction is already in flight, it will be used; otherwise, a new one will be started at the onset of the method invocation and committed when complete.
Our storage engine will be accessed via the Infinispan API’s org.infinispan.AdvancedCache, so we’ll inject this using CDI:
@Inject
private AdvancedCache<String, Attachment> cache;
Armed with a hook to the Infinispan grid, we can then implement the methods of the Repository contract using the Infinispan API:
@Override
public Class<Attachment> getType() {
return Attachment.class;
}
@Override
public Attachment store(Attachment entity) {
try {
cache.withFlags(Flag.SKIP_REMOTE_LOOKUP,
Flag.SKIP_CACHE_LOAD,
Flag.IGNORE_RETURN_VALUES)
.put(entity.getId(), entity);
return entity;
} catch (Exception e) {
throw new RuntimeException("Could not store Attachment with id " +
entity.getId(), e);
}
}
@Override
public Attachment get(String id) {
try {
return cache.get(id);
} catch (Exception e) {
throw new RuntimeException(
"Could not retreive Attachment with id "
+ id, e);
}
}
@Override
public void remove(Attachment entity) {
cache.withFlags(Flag.SKIP_REMOTE_LOOKUP,
Flag.SKIP_CACHE_LOAD,
Flag.IGNORE_RETURN_VALUES)
.remove(entity.getId());
}
Our AttachmentRepository relies upon an Infinispan AdvancedCache, so we must make a CDI producer to create the cache instance to be injected. This is handled by org.cedj.geekseek.domain.attachment.infinispan.CacheProducer:
public class CacheProducer {
@Produces @ApplicationScoped
public EmbeddedCacheManager create() {
GlobalConfiguration global = new GlobalConfigurationBuilder()
.globalJmxStatistics().cacheManagerName("geekseek")
.build();
Configuration local = new ConfigurationBuilder()
.clustering()
.cacheMode(CacheMode.LOCAL)
.transaction()
.transactionMode(TransactionMode.TRANSACTIONAL)
.transactionManagerLookup(new GenericTransactionManagerLookup())
.autoCommit(false)
.build();
return new DefaultCacheManager(global, local);
}
@Produces @ApplicationScoped
public AdvancedCache<String, Attachment> createAdvanced(
EmbeddedCacheManager manager) {
Cache<String, Attachment> cache =
manager.getCache();
return cache.getAdvancedCache();
}
public void destroy(@Disposes Cache<?, ?> cache) {
cache.stop();
}
...
}
CacheProducer does the business of creating and configuring the Infinispan AdvancedCache instance and makes it a valid injection source by use of CDI’s (technically javax.enterprise.inject) @Produces annotation.
This should be enough to fulfill our requirements to perform CRUD operations on an Attachment, and does so in a way that won’t bog down our RDBMS with binary data.
Relation
With our Attachment now modeled and capable of persistence in the data grid, we can move on to the task of associating it with a Session or Conference. Because we’ll handle relationships in a separate layer over the RDBMS, we can do this in a generic fashion that will also grant us the ability to let a User attend or speak at a Conference. The model for a relationship is reflected by org.cedj.geekseek.domain.relation.model. Relation:
public class Relation {
private Key key;
private Date created;
Relation is another standalone class with no additional metadata or dependencies. It contains a Date of creation and a Reference.Key:
private static class Key implements Serializable {
private static final long serialVersionUID = 1L;
private String sourceId;
private String targetId;
private String type;
private Key(String sourceId, String targetId, String type) {
this.sourceId = sourceId;
this.targetId = targetId;
this.type = type;
}
@Override
public int hashCode() {
final int prime = 31;
int result = 1;
result = prime * result + ((sourceId == null) ? 0 :
sourceId.hashCode());
result = prime * result + ((targetId == null) ? 0 :
targetId.hashCode());
result = prime * result + ((type == null) ? 0 : type.hashCode());
return result;
}
@Override
public boolean equals(Object obj) {
if (this == obj)
return true;
if (obj == null)
return false;
if (getClass() != obj.getClass())
return false;
Key other = (Key) obj;
if (sourceId == null) {
if (other.sourceId != null)
return false;
} else if (!sourceId.equals(other.sourceId))
return false;
if (targetId == null) {
if (other.targetId != null)
return false;
} else if (!targetId.equals(other.targetId))
return false;
if (type != other.type)
return false;
return true;
}
}
The Reference.Key very simply draws a link between a source primary key and a target primary key, the IDs of the entities it is linking. Additionally, we assign a type to note what the relationship is reflecting. Because we want to determine value equality using the Object.equalsmethod, we override the equals and hashCode methods (by Object contract, objects with equal values must have equal hashCodes).
The rest of the Relation class is straightforward:
public Relation(String sourceId, String targetId, String type) {
this.key = new Key(sourceId, targetId, type);
this.created = new Date();
}
public String getSourceId() {
return key.sourceId;
}
public String getTargetId() {
return key.targetId;
}
public String getType() {
return key.type;
}
public Date getCreated() {
return (Date) created.clone();
}
}
Now we need a mechanism to persist and remove Relation instances. Our Repository interface used on other objects doesn’t really fit the operations we need; relationships are not true entities, but are instead pointers from one entity to another. So inorg.cedj.geekseek.domain.relation.RelationRepository we’ll define a more fitting contract:
public interface RelationRepository {
Relation add(Identifiable source, String type, Identifiable target);
void remove(Identifiable source, String type, Identifiable target);
<T extends Identifiable> List<T> findTargets(Identifiable source,
String type, Class<T> targetType);
}
The RelationRepository will be used by the services layer, and acts as an abstraction above the data store provider persisting the relationships (a graph database in this case).
Now we’re free to implement RelationRepository with a Neo4j backend in org.cedj.geekseek.domain.relation.neo.GraphRelationRepository:
@ApplicationScoped
public class GraphRelationRepository implements RelationRepository {
private static final String PROP_INDEX_NODE = "all_nodes";
private static final String PROP_INDEX_REL = "all_relations";
private static final String PROP_ID = "id";
private static final String PROP_NODE_CLASS = "_classname";
private static final String PROP_CREATED = "created";
private static final String REL_TYPE_ALL = "all";
@Inject
private GraphDatabaseService graph;
@Inject
private BeanManager manager;
GraphRelationRepository is implemented as an application-scoped CDI bean; it contains a few constants, a hook to the backend graph database (Neo4j API’s GraphDatabaseService), and a reference to the CDI BeanManager.
The RelationRepository contract implementation looks like this:
@Override
public Relation add(Identifiable source, final String type,
Identifiable target) {
Transaction tx = graph.beginTx();
try {
Node root =graph.getNodeById(0);
String sourceTypeName = source.getClass().getSimpleName();
String targetTypeName = target.getClass().getSimpleName();
Node sourceTypeNode = getOrCreateNodeType(sourceTypeName);
Node targetTypeNode = getOrCreateNodeType(targetTypeName);
getOrCreateRelationship(root, sourceTypeNode,
Named.relation(sourceTypeName));
getOrCreateRelationship(root, targetTypeNode,
Named.relation(targetTypeName));
Node sourceNode = getOrCreateNode(source, sourceTypeName);
getOrCreateRelationship(sourceTypeNode, sourceNode,
Named.relation(REL_TYPE_ALL));
Node targetNode = getOrCreateNode(target, targetTypeName);
getOrCreateRelationship(targetTypeNode, targetNode,
Named.relation(REL_TYPE_ALL));
getOrCreateRelationship(sourceNode, targetNode, Named.relation(type));
tx.success();
} catch(Exception e) {
tx.failure();
throw new RuntimeException(
"Could not add relation of type " + type + " between " + source +
" and " + target, e);
} finally {
tx.finish();
}
return new Relation(source.getId(), target.getId(), type);
}
@Override
public void remove(Identifiable source, String type, Identifiable target) {
Transaction tx = graph.beginTx();
try {
Index<Node> nodeIndex = graph.index().forNodes(PROP_INDEX_NODE);
Index<Relationship> relationIndex = graph.index().forRelationships(
PROP_INDEX_REL);
Node sourceNode = nodeIndex.get(PROP_ID, source.getId()).getSingle();
Node targetNode = nodeIndex.get(PROP_ID, target.getId()).getSingle();
for(Relationship rel : sourceNode.getRelationships(
Named.relation(type))) {
if(rel.getEndNode().equals(targetNode)) {
rel.delete();
relationIndex.remove(rel);
}
}
tx.success();
} catch(Exception e) {
tx.failure();
throw new RuntimeException(
"Could not add relation of type " + type + " between " + source +
" and " + target, e);
} finally {
tx.finish();
}
}
@Override
public <T extends Identifiable> List<T> findTargets(Identifiable source,
final String type, final Class<T> targetType) {
Repository<T> repo = locateTargetRepository(targetType);
if(repo == null) {
throw new RuntimeException("Could not locate a " +
Repository.class.getName() + " instance for Type " +
targetType.getName());
}
List<T> targets = new ArrayList<T>();
Index<Node> index = graph.index().forNodes(PROP_INDEX_NODE);
Node node = index.get(PROP_ID, source.getId()).getSingle();
if(node == null) {
return targets;
}
Iterable<Relationship> relationships = node.getRelationships(
Named.relation(type));
List<String> targetIds = new ArrayList<String>();
for(Relationship relation : relationships) {
targetIds.add(relation.getEndNode().getProperty(PROP_ID).toString());
}
for(String targetId : targetIds) {
targets.add(repo.get(targetId));
}
return targets;
}
As shown, this is a fairly simple undertaking given a little research into proper use of the Neo4j API. We’ll also need a little help to resolve the proper Repository types from the types of the entities between which we’re drawing relationships. So we’ll add some internal helper methods toGraphRelationRepository to contain this logic:
/**
* Helper method that looks in the BeanManager for a Repository that
* matches signature Repository<T>.
*
* Used to dynamically find repository to load targets from.
*
* @param targetType Repository object type to locate
* @return Repository<T>
*/
private <T extends Identifiable> Repository<T> locateTargetRepository(
final Class<T> targetType) {
ParameterizedType paramType = new ParameterizedType() {
@Override
public Type getRawType() {
return Repository.class;
}
@Override
public Type getOwnerType() {
return null;
}
@Override
public Type[] getActualTypeArguments() {
return new Type[] {targetType};
}
};
Set<Bean<?>> beans = manager.getBeans(paramType);
Bean<?> bean = manager.resolve(beans);
CreationalContext<?> cc = manager.createCreationalContext(null);
@SuppressWarnings("unchecked")
Repository<T> repo = (Repository<T>)manager.getReference(bean,
paramType, cc);
return repo;
}
private Node getOrCreateNodeType(String type) {
UniqueFactory<Node> factory = new UniqueFactory.UniqueNodeFactory(
graph, PROP_INDEX_NODE) {
@Override
protected void initialize(Node created, Map<String, Object>
properties) {
created.setProperty(PROP_ID, properties.get(PROP_ID));
}
};
return factory.getOrCreate(PROP_ID, type);
}
private Node getOrCreateNode(Identifiable source,
final String nodeClassType) {
UniqueFactory<Node> factory = new UniqueFactory.UniqueNodeFactory(
graph, PROP_INDEX_NODE) {
@Override
protected void initialize(Node created, Map<String, Object>
properties) {
created.setProperty(PROP_ID, properties.get(PROP_ID));
created.setProperty(PROP_NODE_CLASS, nodeClassType);
}
};
return factory.getOrCreate(PROP_ID, source.getId());
}
private Relationship getOrCreateRelationship(final Node source,
final Node target, final RelationshipType type) {
final String key = generateKey(source, target, type);
UniqueFactory<Relationship> factory =
new UniqueFactory.UniqueRelationshipFactory(
graph, PROP_INDEX_REL) {
@Override
protected Relationship create(Map<String, Object> properties) {
Relationship rel = source.createRelationshipTo(target, type);
rel.setProperty(PROP_ID, properties.get(PROP_ID));
return rel;
}
@Override
protected void initialize(Relationship rel,
Map<String, Object> properties) {
rel.setProperty(PROP_CREATED, System.currentTimeMillis());
}
};
return factory.getOrCreate(PROP_ID, key);
}
/**
* Generate some unique key we can identify a relationship with.
*/
private String generateKey(Node source, Node target,
RelationshipType type) {
return source.getProperty(PROP_ID, "X") + "-" + type.name() + "-" +
target.getProperty(PROP_ID, "X");
}
private static class Named implements RelationshipType {
public static RelationshipType relation(String name) {
return new Named(name);
}
private String name;
private Named(String name) {
this.name = name;
}
@Override
public String name() {
return name;
}
}
}
Again, we’ve made an implementation class that depends upon injection of a backend provider’s API. To enable injection of the Neo4j GraphDatabaseService, we’ll create another CDI producer in org.cedj.geekseek.domain.relation.neo.GraphDatabaseProducer:
@ApplicationScoped
public class GraphDatabaseProducer {
private String DATABASE_PATH_PROPERTY = "neo4j.path";
private static Logger log = Logger.getLogger(
GraphDatabaseProducer.class.getName());
@Produces
public GraphDatabaseService createGraphInstance() throws Exception {
String databasePath = getDataBasePath();
log.info("Using Neo4j database at " + databasePath);
return new GraphDatabaseFactory().newEmbeddedDatabase(databasePath);
}
public void shutdownGraphInstance(@Disposes GraphDatabaseService service)
throws Exception {
service.shutdown();
}
private String getDataBasePath() {
String path = System.clearProperty(DATABASE_PATH_PROPERTY);
if(path == null || path.isEmpty()) {
try {
File tmp = File.createTempFile("neo", "geekseek");
File parent = tmp.getParentFile();
tmp.delete();
parent.mkdirs();
path = parent.getAbsolutePath();
}catch (IOException e) {
throw new RuntimeException(
"Could not create temp location for Nepo4j Database. " +
"Please provide system property " + DATABASE_PATH_PROPERTY +
" with a valid path", e);
}
}
return path;
}
}
With this in place we can inject a GraphDataBaseService instance into our GraphRelationRepository.
Our implementation is almost complete, though it’s our position that nothing truly exists until it’s been proven through tests.
Requirement Test Scenarios
Given our user requirements and the implementation choices we’ve made, it’s important we assert that a few areas are working as expected:
§ CRUD operations on Attachment objects
§ Transactional integrity of CRUD operations on Attachment objects
§ Create, Delete, and Find relationships between entities
Attachment CRUD Tests
First we’ll need to ensure that we can Create, Read, Update, and Delete Attachment instances using the data grid provided by Infinispan. To ensure these are working, we’ll use org.cedj.geekseek.domain.attachment.test.integration.AttachmentRepositoryTestCase:
@RunWith(Arquillian.class)
public class AttachmentRepositoryTestCase {
// Given
@Deployment
public static WebArchive deploy() {
return ShrinkWrap.create(WebArchive.class)
.addAsLibraries(
CoreDeployments.core(),
AttachmentDeployments.attachmentWithCache())
.addAsLibraries(AttachmentDeployments.resolveDependencies())
.addClass(TestUtils.class)
.addAsWebInfResource(EmptyAsset.INSTANCE, "beans.xml");
}
Here we have a simple Arquillian test defined with no additional extensions. We’ll deploy an attachmentWithCache, as defined by:
public static JavaArchive attachment() {
return ShrinkWrap.create(JavaArchive.class)
.addPackage(Attachment.class.getPackage())
.addAsManifestResource(EmptyAsset.INSTANCE, "beans.xml");
}
public static JavaArchive attachmentWithCache() {
return attachment()
.addPackage(AttachmentRepository.class.getPackage())
.addPackage(CacheProducer.class.getPackage());
}
This will give us our Attachment domain entity, the AttachmentRepository, and the CDI producer to inject hooks into an Infinispan Cache as shown before. Additionally, we’ll need to deploy the Infinispan API and implementation as a library, soAttachmentDeployments.resolveDependencies will bring this in for us:
public static File[] resolveDependencies() {
return Maven.resolver()
.offline()
.loadPomFromFile("pom.xml")
.resolve("org.infinispan:infinispan-core")
.withTransitivity()
.asFile();
}
This uses the ShrinkWrap Maven Resolver to pull the groupId:artifactId of org.infinispan:infinispan-core and all of its dependencies in from the Maven repository, returning the artifacts as files. We don’t need to define the version explicitly here; that will be configured from the definition contained in the project’s pom.xml file because we’ve told the resolver to loadPomFromFile("pom.xml").
Also as part of the deployment we’ll throw in a TestUtils class, which will let us easily create Attachment objects from the tests running inside the container:
public static Attachment createAttachment() {
try {
return new Attachment(
"Test Attachment",
"text/plain",
new URL("http://geekseek.org"));
} catch(MalformedURLException e) {
throw new RuntimeException(e);
}
}
The resulting deployment should have structure that looks similar to this:
749e9f51-d858-42a6-a06e-3f3d03fc32ad.war:
/WEB-INF/
/WEB-INF/lib/
/WEB-INF/lib/jgroups-3.3.1.Final.jar
/WEB-INF/lib/43322d61-32c4-444c-9681-079ac34c6e87.jar
/WEB-INF/lib/staxmapper-1.1.0.Final.jar
/WEB-INF/lib/jboss-marshalling-river-1.3.15.GA.jar
/WEB-INF/lib/56201983-371f-4ed5-8705-d4fd6ec8f936.jar
/WEB-INF/lib/infinispan-core-5.3.0.Final.jar
/WEB-INF/lib/jboss-marshalling-1.3.15.GA.jar
/WEB-INF/lib/jboss-logging-3.1.1.GA.jar
/WEB-INF/beans.xml
/WEB-INF/classes/
/WEB-INF/classes/org/
/WEB-INF/classes/org/cedj/
/WEB-INF/classes/org/cedj/geekseek/
/WEB-INF/classes/org/cedj/geekseek/domain/
/WEB-INF/classes/org/cedj/geekseek/domain/attachment/
/WEB-INF/classes/org/cedj/geekseek/domain/attachment/test/
/WEB-INF/classes/org/cedj/geekseek/domain/attachment/test/TestUtils.class
As we can see, Infinispan and all of its dependencies have made their way to WEB-INF/lib; our own libraries are not explicitly named, so they’re assigned a UUID filename.
NOTE
It’s useful to debug your deployments by simply printing out a listing of your archive; this is easily accomplished by throwing a statement like System.out.println(archive.toString(true)); in your @Deployment method before returning the archive. If you want to debug the content of the final deployment as seen by the container, you can set thedeploymentExportPath property under the engine element in arquillian.xml to the path where you want Arquillian to output the deployments. This is useful if you’re having deployment problems that you suspect are related to how Arquillian enriches the deployment, or if you’re generating file content dynamically.
Now let’s give our test a hook to the Repository we’ll use to perform CRUD operations on our Attachment objects:
@Inject
private Repository<Attachment> repository;
With the deployment and injection of the Repository done, we’re now free to implement our tests:
// Story: As a User I should be able to create an Attachment
@Test
public void shouldBeAbleToCreateAttachment() throws Exception {
Attachment attachment = createAttachment();
repository.store(attachment);
Attachment stored = repository.get(attachment.getId());
Assert.assertNotNull(stored);
Assert.assertEquals(attachment.getId(), stored.getId());
Assert.assertEquals(attachment.getTitle(), stored.getTitle());
Assert.assertEquals(attachment.getUrl(), stored.getUrl());
Assert.assertEquals(attachment.getMimeType(), stored.getMimeType());
Assert.assertNotNull(stored.getCreated());
}
// Story: As a User I should be able to update an Attachment
@Test
public void shouldBeAbleToUpdateAttachment() throws Exception {
String updatedTitle = "Test 2";
Attachment attachment = createAttachment();
attachment = repository.store(attachment);
attachment = attachment.setTitle(updatedTitle);
attachment = repository.store(attachment);
Attachment updated = repository.get(attachment.getId());
Assert.assertEquals(updated.getTitle(), updatedTitle);
Assert.assertNotNull(attachment.getLastUpdated());
}
// Story: As a User I should be able to remove an Attachment
@Test
public void shouldBeAbleToRemoveAttachment() throws Exception {
Attachment attachment = createAttachment();
attachment = repository.store(attachment);
repository.remove(attachment);
Attachment removed = repository.get(attachment.getId());
Assert.assertNull(removed);
}
@Test
public void shouldNotReflectNonStoredChanges() throws Exception {
tring updatedTitle = "Test Non Stored Changes";
Attachment attachment = createAttachment();
String originalTitle = attachment.getTitle();
Attachment stored = repository.store(attachment);
// tile change not stored to repository
stored = stored.setTitle(updatedTitle);
Attachment refreshed = repository.get(attachment.getId());
Assert.assertEquals(refreshed.getTitle(), originalTitle);
}
}
So here we have our CRUD tests using the injected Repository to perform their persistence operations. In turn, we’ve implemented the Repository with an Infinispan backend (which in this case is running in local embedded mode). We can now be assured that our repository layer is correctly hooked together and persistence to the data grid is working properly.
Transactional Integrity of Attachment Persistence
While we’re confident that the CRUD operations of our Attachment entity are in place, we should ensure that the transactional semantics are upheld if a transaction is in flight. This will essentially validate that Infinispan is respectful of the Java Transactions API (JTA), a specification under the direction of the JSR-907 Expert Group.
To accomplish this, we’re going to directly interact with JTA’s UserTransaction in our test. In fact, the Attachment entity is not the only one we should be verifying, so we’ll code this test in a way that will enable us to extend it to ensure that Conference, Session, and other entities can be exercised for transactional compliance.
Our goals are to assert that for any entity type T:
§ T is Stored on commit and can be read from another transaction.
§ T is Updated on commit and can be read from another transaction.
§ T is Removed on commit and cannot be read by another transaction.
§ T is not Stored on rollback and cannot be read by another transaction.
§ T is not Updated on rollback and cannot be read by another transaction.
§ T is not Removed on rollback and can be read by another transaction.
Therefore we’ll attempt to centralize these operations in a base test class that will, when provided a T and a Repository<T>, verify that T is committed and rolled back as required. Thus we introduceorg.cedj.geekseek.domain.test.integration.BaseTransactionalSpecification:
public abstract class BaseTransactionalSpecification<
DOMAIN extends Identifiable,
REPO extends Repository<DOMAIN>> {
We define some generic variables for easy extension; this test will deal with entity objects of type Identifiable and the Repository that interacts with them. Next we’ll gain access to the JTA UserTransaction:
@Inject
private UserTransaction tx;
Because this class is to be extended for each entity type we’d like to test, we’ll make a contract for those implementations to supply:
/**
* Get the Repository instance to use.
*/
protected abstract REPO getRepository();
/**
* Create a new unique instance of the Domain Object.
*/
protected abstract DOMAIN createNewDomainObject();
/**
* Update some domain object values.
*/
protected abstract void updateDomainObject(
DOMAIN domain);
/**
* Validate that the update change has occurred.
* Expecting Assert error when validation does not match.
*/
protected abstract void validateUpdatedDomainObject(
DOMAIN domain);
And now we’re free to write the tests backing the points listed earlier; we want to validate that objects are either accessible or not based on commit or rollback operations to the transaction in play. For instance, this test ensures that an object is stored after a commit:
@Test
public void shouldStoreObjectOnCommit() throws Exception {
final DOMAIN domain = createNewDomainObject();
commit(Void.class, new Store(domain));
DOMAIN stored = commit(new Get(domain.getId()));
Assert.assertNotNull(
"Object should be stored when transaction is committed",
stored);
}
protected DOMAIN commit(Callable<DOMAIN> callable) throws Exception {
return commit(getDomainClass(), callable);
}
protected <T> T commit(Class<T> type, Callable<T> callable)
throws Exception {
try {
tx.begin();
return callable.call();
} finally {
tx.commit();
}
}
private class Store implements Callable<Void> {
private DOMAIN domain;
public Store(DOMAIN domain) {
this.domain = domain;
}
@Override
public Void call() throws Exception {
getRepository().store(domain);
return null;
}
}
private class Get implements Callable<DOMAIN> {
private String id;
public Get(String id) {
this.id = id;
}
@Override
public DOMAIN call() throws Exception {
return getRepository().get(id);
}
}
Here we see that we manually manipulate the UserTransaction to our liking in the test method; the mechanics of this interaction are handled by the commit method.
We have similar tests in place to validate the other conditions:
@Test public void shouldUpdateObjectOnCommit() throws Exception {...}
@Test public void shouldRemoveObjectOnCommmit() throws Exception {...}
@Test public void shouldNotStoreObjectOnRollback() throws Exception {...}
@Test public void shouldNotUpdateObjectOnRollback() throws Exception {...}
@Test public void shouldNotRemoveObjectOnRollback() throws Exception {...}
@Test public void shouldSetCreatedDate() throws Exception {...}
@Test public void shouldSetUpdatedDate() throws Exception {...}
With our base class containing most of our support for the transactional specification tests, now we can provide a concrete implementation for our Attachment entities. We do this in org.cedj.geekseek.domain.attachment.test.integration.AttachmentRepositoryTransactionalTestCase:
@RunWith(Arquillian.class)
public class AttachmentRepositoryTransactionalTestCase
extends
BaseTransactionalSpecification<Attachment, Repository<Attachment>> {
We’ll extend BaseTransactionalSpecification and close the generic context to be relative to Attachment. By implementing the parent abstract methods of the parent class, we’ll then be done and able to run transactional tests on Attachment types:
private static final String UPDATED_TITLE = "TEST UPDATED";
...
@Inject
private Repository<Attachment> repository;
@Override
protected Attachment createNewDomainObject() {
return createAttachment();
}
@Override
protected Attachment updateDomainObject(
Attachment domain) {
return domain.setTitle(UPDATED_TITLE);
}
@Override
protected void validateUpdatedDomainObject(Attachment domain) {
Assert.assertEquals(UPDATED_TITLE, domain.getTitle());
}
@Override
protected Repository<Attachment> getRepository() {
return repository;
}
With these tests passing, we’re now satisfied that our Infinispan backend is complying with the semantics of a backing application transaction. We therefore have nicely abstracted the data grid from the perspective of the caller; it’s just another transactionally aware persistence engine representing itself as a Repository.
Validating Relationships
Armed with our Neo4j-backed RelationRepository, we’re able to draw relationships between entities that are not otherwise related in the schema, or may even be in separate data stores. Let’s construct a test to validate that our Relation edges in the graph are serving us well. We’ll do this in org.cedj.geekseek.domain.relation.test.integration.RelationTestCase:
@RunWith(Arquillian.class)
public class RelationTestCase {
This will be another relatively simple Arquillian test case, running inside the container. We’ll again define a deployment, this time including Neo4j as a dependency in place of Infinispan:
@Deployment
public static WebArchive deploy() {
return ShrinkWrap.create(WebArchive.class)
.addAsLibraries(
RelationDeployments.relationWithNeo(),
CoreDeployments.core())
.addAsLibraries(RelationDeployments.neo4j())
.addPackage(SourceObject.class.getPackage())
.addAsWebInfResource(EmptyAsset.INSTANCE, "beans.xml");
}
This deployment will include our GraphDatabaseProducer, so we’ll be able to inject a GraphRelationRepository in our test case to create, remove, and find Relation edges. We’ll obtain this easily via injection into the test instance:
@Inject
private GraphRelationRepository repository;
Now we’ll set up some constants and instance members, then populate them before each test runs using a JUnit lifecycle annotation:
private static final String SOURCE_ID = "11";
private static final String TARGET_ID = "1";
private SourceObject source;
private TargetObject target;
private String type;
@Before
public void createTypes() {
source = new SourceObject(SOURCE_ID);
target = new TargetObject(TARGET_ID);
type = "SPEAKING";
}
SourceObject and TargetObject are test-only objects we’ve introduced to represent entities. Again, we only care about relationships here, so there’s no sense tying this test to one of our real entities at this level of integration. At this point we want to test the Relation and its persistence mechanisms in as much isolation as possible, so it’s appropriate to tie together test-only entities.
Now we’ll want to run our tests to:
§ Create a relationship
§ Find the created relationship
§ Delete the relationship
§ Only find valid targets remaining
Rather than do this in one large test, we’ll make separate tests for each case. There are dependencies however, because the state of the system will change after each test is run. Therefore we need to make sure that these tests run in the proper order using Arquillian’s @InSequence annotation:
@Test @InSequence(0)
public void shouldBeAbleToCreateRelation() {
Relation relation = repository.add(source, type, target);
Assert.assertEquals("Verify returned object has same source id",
relation.getSourceId(), source.getId());
Assert.assertEquals("Verify returned object has same target id",
relation.getTargetId(), target.getId());
Assert.assertEquals("Verify returned object has same type",
relation.getType(), type);
Assert.assertNotNull("Verify created date was set",
relation.getCreated());
}
@Test @InSequence(1)
public void shouldBeAbleToFindTargetedRelations(
Repository<TargetObject> targetRepo,
Repository<SourceObject> sourceRepo) {...}
@Test @InSequence(2)
public void shouldBeAbleToDeleteRelations() {...}
@Test @InSequence(3)
public void shouldOnlyFindGivenRelation() {...}
With these passing, it’s now proven that we can perform all the contracted operations of RelationRepository against a real Neo4j graph database backend.
Our GeekSeek application now has many database layers at its disposal: CRUD operations in an RDBMS for most entities, a key/value store to hold onto Attachment objects, and a graph to draw ties among the entities such that their relationships can be explored in an efficient manner.