How to build a custom Lucene index for Neo4j graph? - neo4j

I am using Gremlin and Neo4j to manipulate the ENRON dataset from infochimps. This dataset has two types of vertexes, Message and Email Addresss and two types of edges, SENT and RECEVIED_BY. I would like to create a custom index on this dataset that creates a Lucene document for each vertex of type: 'Message' and incorporates information from associated vertexes (e.g., v.in(), v.out()) as additional fields in the Lucene document.
I am thinking of code along the lines of
g = new Neo4jGraph('enron');
PerFieldAnalyzerWrapper analyzer =
new PerFieldAnalyzerWrapper(new StandardAnalyzer());
analyzer.addAnalyzer("sender", new KeywordAnalyzer());
analyzer.addAnalyzer("recipient", new KeywordAnalyzer());
IndexWriter idx = new IndexWriter (dir,analyzer,IndexWriter.MaxFieldLength.UNLIMITED);
g.V.filter{it.type == 'Message'}.each { v ->
Document doc = new Document();
doc.add(new Field("subject", v.subject));
doc.add(new Field("body", v.body));
doc.add(new Field("sender", v.in().address);
v.out().each { recipient ->
doc.add(new Field("recipient", recipient.address));
}
idx.addDocument(doc);
}
idx.close();
My questions are:
Is there a better way to enumerate vertexes for indexing?
Can I use auto-indexing for this, and if so, how to I specify what should be indexed?
Can I specify my own Analyzer, or am I stuck with the default? What is the default?
If I must create my own index, should I be using gremlin for this, or am I better off with a Java program?

I will be talking about direct Neo4j access here since I'm not well travelled in Gremlin.
So you'd like to build a Lucene index "outside of" the graph itself? Otherwise you can use the built in graphDb.index().forNodes( "myIndex", configForMyIndex ) to get (created on demand) a Lucene index associated with neo4j. You can then add multiple fields to each document by calling index.add( node, key, value ), where each node will be represented by one document in that Lucene index.
1) In Gremiln... I don't know
2) See http://docs.neo4j.org/chunked/milestone/auto-indexing.html
3) See http://docs.neo4j.org/chunked/milestone/indexing-create-advanced.html
4) Do you need to create it outside of the db entirely? If so, why?

I just finished an import with a Java process and it's really easy, in my opinion better inclusive through Gremlin.
Anyway, if the process is failing is because of you CAN'T create a new object of StandardAnalyzer. All the constructors of that class require parameters, so you should create a wrapper class or create it with the right version of Lucene like paramater in the constructor.
Neo4J, until today, accepts only until the lucene version 36.

Related

Using Merge in BatchInserter?

I am using the BatchInserter in order to create some nodes and relationships, however I have unique nodes, and I wanted to make multiple relationships between them.
I can easily do that using the Cypher and in the very same time by using the Java Core API by:
ResourceIterator<Node> existedNodes = graphDBService.findNodesByLabelAndProperty( DynamicLabel.label( "BaseProduct" ), "code", source.getBaseProduct().getCode() ).iterator();
if ( !existedNodes.hasNext() )
{
//TO DO
}
else {
// create relationship with the retrieved node
}
and in Cypher I can easily use the merge.
is there any possible way to do the same with the BatchInserter ?
No it is not possible in the batch-inserter, as those APIs are not available there.
That's why I usually keep in-memory maps with the information I need to look up.
See this blog post for a groovy script:
http://jexp.de/blog/2014/10/flexible-neo4j-batch-import-with-groovy/

Grails: query or criteria against a string/value pairs map property

Grails gives the possibility of creating simple string/value map properties section "Maps of Objects", first paragraph.
I was wondering, is there a way to later query the domain class (using Gorm dynamic finders, criterias or HQL) using the map property as part of the query (i.e adding a condition for the key X to have the value Y)?
After playing with it a bit and almost give up, I discovered the map syntax to (surprisingly) work in HQL. Assuming the class looks like:
class SomeClass {
Map pairKeyProperty
}
You can build queries that look like the following:
select * from SomeClass sc where sc.pairKeyProperty['someKey'] = 'someValue' and sc.pairKeyProperty['someOtherKey'] = 'someOtherValue'
Pretty neat! I still would prefer to use criterias as they are much cleaner to compose, but they seem to not support the same syntax (or I couldn't find it).
I created a sample app in GitHub:
https://github.com/deigote/grails-simple-map-of-string-value-pairs
It can be visisted at:
http://grails-map-of-string-pairs.herokuapp.com/
The form above uses a cross join. To enforce an inner join use
join sc.pairKeyProperty pk1 on index(pk1) = 'someKey'
where 'someValue' in elements(pk1)

Neo4jClient: doubts about CRUD API

My persistency layer essentially uses Neo4jClient to access a Neo4j 1.9.4 database. More specifically, to create nodes I use IGraphClient#Create() in Neo4jClient's CRUD API and to query the graph I use Neo4jClient's Cypher support.
All was well until a friend of mine pointed out that for every query, I essentially did two HTTP requests:
one request to get a node reference from a legacy index by the node's unique ID (not its node ID! but a unique ID generated by SnowMaker)
one Cypher query that started from this node reference that does the actual work.
For read operations, I did the obvious thing and moved the index lookup into my Start() call, i.e.:
GraphClient.Cypher
.Start(new { user = Node.ByIndexLookup("User", "Id", userId) })
// ... the rest of the query ...
For create operations, on the other hand, I don't think this is actually possible. What I mean is: the Create() method takes a POCO, a couple of relationship instances and a couple of index entries in order to create a node, its relationships and its index entries in one transaction/HTTP request. The problem is the node references that you pass to the relationship instances: where do they come from? From previous HTTP requests, right?
My questions:
Can I use the CRUD API to look up node A by its ID, create node B from a POCO, create a relationship between A and B and add B's ID to a legacy index in one request?
If not, what is the alternative? Is the CRUD API considered legacy code and should we move towards a Cypher-based Neo4j 2.0 approach?
Does this Cypher-based approach mean that we lose POCO-to-node translation for create operations? That was very convenient.
Also, can Neo4jClient's documentation be updated because it is, frankly, quite poor. I do realize that Readify also offers commercial support so that might explain things.
Thanks!
I'm the author of Neo4jClient. (The guy who gives his software away for free.)
Q1a:
"Can I use the CRUD API to look up node A by its ID, create node B from a POCO, create a relationship between A and B"
Cypher is the way of not just the future, but also the 'now'.
Start with the Cypher (lots of resources for that):
START user=node:user(Id: 1234)
CREATE user-[:INVITED]->(user2 { Id: 4567, Name: "Jim" })
Return user2
Then convert it to C#:
graphClient.Cypher
.Start(new { user = Node.ByIndexLookup("User", "Id", userId) })
.Create("user-[:INVITED]->(user2 {newUser})")
.WithParam("newUser", new User { Id = 4567, Name = "Jim" })
.Return(user2 => user2.Node<User>())
.Results;
There are lots more similar examples here: https://github.com/Readify/Neo4jClient/wiki/cypher-examples
Q1b:
" and add B's ID to a legacy index in one request?"
No, legacy indexes are not supported in Cypher. If you really want to keep using them, then you should stick with the CRUD API. That's ok: if you want to use legacy indexes, use the legacy API.
Q2.
"If not, what is the alternative? Is the CRUD API considered legacy code and should we move towards a Cypher-based Neo4j 2.0 approach?"
That's exactly what you want to do. Cypher, with labels and automated indexes:
// One time op to create the index
// Yes, this syntax is a bit clunky in C# for now
graphClient.Cypher
.Create("INDEX ON :User(Id)")
.ExecuteWithoutResults();
// Find an existing user, create a new one, relate them,
// and index them, all in a single HTTP call
graphClient.Cypher
.Match("(user:User)")
.Where((User user) => user.Id == userId)
.Create("user-[:INVITED]->(user2 {newUser})")
.WithParam("newUser", new User { Id = 4567, Name = "Jim" })
.ExecuteWithoutResults();
More examples here: https://github.com/Readify/Neo4jClient/wiki/cypher-examples
Q3.
"Does this Cypher-based approach mean that we lose POCO-to-node translation for create operations? That was very convenient."
Correct. But that's what we collectively all want to do, where Neo4j is going, and where Neo4jClient is going too.
Think about SQL for a second (something that I assume you are familiar with). Do you run a query to find the internal identifier of a node, including its file offset on disk, then use this internal identifier in a second query to manipulate it? No. You run a single query that does all that in one hit.
Now, a common use case for why people like passing around Node<T> or NodeReference instances is to reduce repetition in queries. This is a legitimate concern, however because the fluent queries in .NET are immutable, we can just construct a base query:
public ICypherFluentQuery FindUserById(long userId)
{
return graphClient.Cypher
.Match("(user:User)")
.Where((User user) => user.Id == userId);
// Nothing has been executed here: we've just built a query object
}
Then use it like so:
public void DeleteUser(long userId)
{
FindUserById(userId)
.Delete("user")
.ExecuteWithoutResults();
}
Or, add even more Cypher logic to delete all the relationships too:
Then use it like so:
public void DeleteUser(long userId)
{
FindUserById(userId)
.Match("user-[:?rel]-()")
.Delete("rel, user")
.ExecuteWithoutResults();
}
This way, you can effectively reuse references, but without ever having to pull them back across the wire in the first place.

Neo4j Embedded Fulltext Automatic Node Index

When running Neo4j embedded, the default configuration doesn't have the automatic node index set as fulltext (meaning that all Lucene queries are case sensitive). How can I configure the automatic index to be fulltext?
For starters, you must perform this on a new database. The automatic index is lazily created, which means that it isn't created until the first access. You have until the first access to perform this configuration. If you attempt to change the property after it's already been created, it won't work. So the first step is to load the database with automatic indexing enabled (node or relationship).
val db = new GraphDatabaseFactory().newEmbedddedDatabaseBuilder("path/to/db").
setConfig(GraphDatabaseSettings.node_keys_indexable, "label,username").
setConfig(GraphDatabaseSettings.node_auto_indexing, "true").newGraphDatabase()
Now, before you do anything, you have to set the configuration properties. You can find out about the possible properties and values here. To do this, we just need two more lines.
val autoIndex = db.index.forNodes("node_auto_index")
db.index.setConfiguration(autoIndex, "type", "fulltext")
And that's all there is to it. You can now create vertices and relationships and the automatic index will be created and populated. You can get use the following code to query it using any Lucene query.
autoIndex.getAutoIndex.query("label:*caseinsensitive*")

Neo4j indexes and legacy data

I have a legacy dataset (ENRON data represented as GraphML) that I would like to query. In an comment in a related question, #StefanArmbruster suggests that I use Cypher to query the database. My query use case is simple: given a message id (a property of the Message node), retrieve the node that has that id, and also retrieve the sender and recipient nodes of that message.
It seems that to do this in Cypher, I first have to create an index of the nodes. Is there a way to do this automatically when the data is loaded from the graphML file? (I had used Gremlin to load the data and create the database.)
I also have an external Lucene index of the data (I need it for other purposes). Does it make sense to have two indexes? I could, for example, index the Neo4J node ids into my external index, and then query the graph based on those ids. My concern is about the persistence of these ids. (By analogy, Lucene document ids should not be treated as persistent.)
So, should I:
Index the Neo4j graph internally to query on message ids using Cypher? (If so, what is the best way to do that: regenerate the database with some suitable incantation to get the index built? Build the index on the already-existing db?)
Store Neo4j node ids in my external Lucene index and retrieve nodes via these stored ids?
UPDATE
I have been trying to get auto-indexing to work with Gremlin and an embedded server, but with no luck. In the documentation it says
The underlying database is auto-indexed, see Section 14.12, “Automatic Indexing” so the script can return the imported node by index lookup.
But when I examine the graph after loading a new database, no indexes seem to exist.
The Neo4j documentation on auto indexing says that a bunch of configuration is required. In addition to setting node_auto_indexing = true, you have to configure it
To actually auto index something, you have to set which properties
should get indexed. You do this by listing the property keys to index
on. In the configuration file, use the node_keys_indexable and
relationship_keys_indexable configuration keys. When using embedded
mode, use the GraphDatabaseSettings.node_keys_indexable and
GraphDatabaseSettings.relationship_keys_indexable configuration keys.
In all cases, the value should be a comma separated list of property
keys to index on.
So is Gremlin supposed to set the GraphDatabaseSettings parameters? I tried passing in a map into the Neo4jGraph constructor like this:
Map<String,String> config = [
'node_auto_indexing':'true',
'node_keys_indexable': 'emailID'
]
Neo4jGraph g = new Neo4jGraph(graphDB, config);
g.loadGraphML("../databases/data.graphml");
but that had no apparent effect on index creation.
UPDATE 2
Rather than configuring the database through Gremlin, I used the examples given in the Neo4j documentation so that my database creation was like this (in Groovy):
protected Neo4jGraph getGraph(String graphDBname, String databaseName) {
boolean populateDB = !new File(graphDBName).exists();
if(populateDB)
println "creating database";
else
println "opening database";
GraphDatabaseService graphDB = new GraphDatabaseFactory().
newEmbeddedDatabaseBuilder( graphDBName ).
setConfig( GraphDatabaseSettings.node_keys_indexable, "emailID" ).
setConfig( GraphDatabaseSettings.node_auto_indexing, "true" ).
setConfig( GraphDatabaseSettings.dump_configuration, "true").
newGraphDatabase();
Neo4jGraph g = new Neo4jGraph(graphDB);
if (populateDB) {
println "Populating graph"
g.loadGraphML(databaseName);
}
return g;
}
and my retrieval was done like this:
ReadableIndex<Node> autoNodeIndex = graph.rawGraph.index()
.getNodeAutoIndexer()
.getAutoIndex();
def node = autoNodeIndex.get( "emailID", "<2614099.1075839927264.JavaMail.evans#thyme>" ).getSingle();
And this seemed to work. Note, however, that the getIndices() call on the Neo4jGraph object still returned an empty list. So the upshot is that I can exercise the Neo4j API correctly, but the Gremlin wrapper seems to be unable to reflect the indexing state. The expression g.idx('node_auto_index') (documented in Gremlin Methods) returns null.
the auto indexes are created lazily. That is - when you have enabled the auto-indexing, the actual index is first created when you index your first property. Make sure you are inserting data before checking the existence of the index, otherwise it might not show up.
For some auto-indexing code (using programmatic configuration), see e.g. https://github.com/neo4j-contrib/rabbithole/blob/master/src/test/java/org/neo4j/community/console/IndexTest.java (this is working with Neo4j 1.8
/peter
Have you tried the automatic index feature? It's basically the use case you're looking for--unfortunately it needs to be enabled before you import the data. (Otherwise you have to remove/add the properties to reindex them.)
http://docs.neo4j.org/chunked/milestone/auto-indexing.html

Resources