Spring Data Neo4j or in short SDN is a next-generation Spring Data module, created and maintained by Neo4j, Inc. in close collaboration with VMware’s Spring Data Team.

SDN relies completely on the Neo4j Java Driver, without introducing another "driver" or "transport" layer between the mapping framework and the driver. The Neo4j Java Driver - sometimes dubbed Bolt or the Bolt driver - is used as a protocol much like JDBC is with relational databases.

Noteworthy features that differentiate the new SDN from Spring Data Neo4j + OGM are

SDN has several features not present in SDN+OGM, notably

Neo4j-OGM is an Object Graph Mapping library, which is mainly used by previous versions of Spring Data Neo4j as its backend for the heavy lifting of mapping nodes and relationships into domain object. The current SDN does not need and does not support Neo4j-OGM. SDN uses Spring Data’s mapping context exclusively for scanning classes and building the meta model.

While this pins SDN to the Spring ecosystem, it has several advantages, among them the smaller footprint regarding CPU and memory usage and especially, all the features of Spring’s mapping context.

No.

Embedded Neo4j has multiple facets to it:

You can issue a curl request against the Spring Initializer to create a basic Maven project:

This will create a new folder Neo4jSpringBootExample. As this starter is not yet on the initializer, you will have to add the following dependency manually to your pom.xml:

You would also add the dependency manually in case of an existing project.

The idea is the same, just generate a Gradle project:

The dependency for Gradle looks like this and must be added to build.gradle:

You would also add the dependency manually in case of an existing project.

SDN fully supports unmodifiable entities, for both Java and data classes in Kotlin. Therefore we will focus on immutable entities here, Listing 6 shows a such an entity.

As a general remark: immutable entities using internally generated ids are a bit contradictory, as SDN needs a way to set the field with the value generated by the database.

If you don’t find a good business key or don’t want to use a generator for IDs, here’s the same entity using the internally generated id together with a regular constructor and a so called wither-Method, that is used by SDN:

You can of course use SDN with Kotlin and model your domain with Kotlin’s data classes. Project Lombok is an alternative if you want or need to stay purely within Java.

You basically have two options here: you can work in a store-agnostic fashion with SDN and make your domain specific extend one of

Choose imperative and reactive accordingly.

The other option is to settle on a store specific implementation and gain all the methods we support out of the box. The advantage of this approach is also its biggest disadvantage: once out, all those methods will be part of your API. Most of the time it’s harder to take something away, than to add stuff afterwards. Furthermore, using store specifics leaks your store into your domain. From a performance point of view, there is no penalty.

A repository fitting to any of the movie entities above looks like this:

The @Node annotation is used to mark a class as a managed domain class, subject to the classpath scanning by the mapping context.

To map an Object to nodes in the graph and vice versa, we need a label to identify the class to map to and from.

@Node has an attribute labels that allows you to configure one or more labels to be used when reading and writing instances of the annotated class. The value attribute is an alias for labels. If you don’t specify a label, then the simple class name will be used as the primary label. In case you want to provide multiple labels, you could either:

The primary label should always be the most concrete label that reflects your domain class.

For each instance of an annotated class that is written through a repository or through the Neo4j template, one node in the graph with at least the primary label will be written. Vice versa, all nodes with the primary label will be mapped to the instances of the annotated class.

While @Node creates a mapping between a class and nodes having a specific label, we also need to make the connection between individual instances of that class (objects) and instances of the node.

This is where @Id comes into play. @Id marks an attribute of the class to be the unique identifier of the object. That unique identifier is in an optimal world a unique business key or in other words, a natural key. @Id can be used on all attributes with a supported simple type.

Natural keys are however pretty hard to find. Peoples names for example are seldom unique, change over time or worse, not everyone has a first and last name.

We therefore support two different kind of surrogate keys.

On an attribute of type long or Long, @Id can be used with @GeneratedValue. This maps the Neo4j internal id, which is not a property on a node or relationship and usually not visible, to the attribute and allows SDN to retrieve individual instances of the class.

@GeneratedValue provides the attribute generatorClass. generatorClass can be used to specify a class implementing IdGenerator. An IdGenerator is a functional interface and its generateId takes the primary label and the instance to generate an Id for. We support UUIDStringGenerator as one implementation out of the box.

You can also specify a Spring Bean from the application context on @GeneratedValue via generatorRef. That bean also needs to implement IdGenerator, but can make use of everything in the context, including the Neo4j client or template to interact with the database.

Spring Data Neo4j supports optimistic locking by using the @Version annotation on a Long typed field. This attribute will get incremented automatically during updates and must not be manually modified.

If, e.g., two transactions in different threads want to modify the same object with version x, the first operation will get successfully persisted to the database. At this moment, the version field will get incremented, so it is x+1. The second operation will fail with a OptimisticLockingFailureException because it wants to modify the object with the version x that does not exist anymore in the database. In such cases the operation needs to get retried, beginning with a fresh fetch of the object with the current version from the database.

All attributes of a @Node-annotated class will be persisted as properties of Neo4j nodes and relationships. Without further configuration, the name of the attribute in the Java or Kotlin class will be used as Neo4j property.

If you are working with an existing Neo4j schema or just like to adapt the mapping to your needs, you will need to use @Property. The name is used to specify the name of the property inside the database.

The @Relationship annotation can be used on all attributes that are not a simple type. It is applicable on attributes of other types annotated with @Node or collections and maps thereof.

The type or the value attribute allow configuration of the relationship’s type, direction allows specifying the direction. The default direction in SDN is Relationship.Direction#OUTGOING.

We support dynamic relationships. Dynamic relationships are represented as a Map<String, AnnotatedDomainClass> or Map<Enum, AnnotatedDomainClass>. In such a case, the type of the relationship to the other domain class is given by the maps key and must not be configured through the @Relationship.

Putting all those together, we can create a simple domain. We use movies and people with different roles:

People are mapped in two roles here, actors and directors. The domain class is the same:

The easiest way to give your domain classes a unique identifier is the combination of @Id and @GeneratedValue on a field of type Long (preferable the object, not the scalar long, as literal null is the better indicator whether an instance is new or not):

You don’t need to provide a setter for the field, SDN will use reflection to assign the field, but use a setter if there is one. If you want to create an immutable entity with an internally generated id, you have to provide a wither.

You either have to provide a setter for the id attribute or something like a wither, if you want to have

The @GeneratedValue annotation can take a class implementing org.springframework.data.neo4j.core.schema.IdGenerator as parameter. SDN provides InternalIdGenerator (the default) and UUIDStringGenerator out of the box. The latter generates new UUIDs for each entity and returns them as java.lang.String. An application entity using that would look like this:

We have to discuss two separate things regarding advantages and disadvantages. The assignment itself and the UUID-Strategy. A universally unique identifier is meant to be unique for practical purposes. To quote Wikipedia: “Thus, anyone can create a UUID and use it to identify something with near certainty that the identifier does not duplicate one that has already been, or will be, created to identify something else.” Our strategy uses Java internal UUID mechanism, employing a cryptographically strong pseudo random number generator. In most cases that should work fine, but your mileage might vary.

That leaves the assignment itself:

You have several options to roll out your own ID generator. One is a POJO implementing a generator:

Another option is to provide an additional Spring Bean like this:

The generator above would be configured as a bean reference like this:

We have been using a business key in the complete example’s MovieEntity and PersonEntity. The name of the person is assigned at construction time, both by your application and while being loaded through Spring Data.

This is only possible, if you find a stable, unique business key, but makes great immutable domain objects.

Spring Data automatically tries to detect a persistent entity’s constructor to be used to materialize objects of that type. The resolution algorithm works as follows:

The value resolution assumes constructor argument names to match the property names of the entity, i.e. the resolution will be performed as if the property was to be populated, including all customizations in mapping (different datastore column or field name etc.). This also requires either parameter names information available in the class file or an @ConstructorProperties annotation being present on the constructor.

Once an instance of the entity has been created, Spring Data populates all remaining persistent properties of that class. Unless already populated by the entity’s constructor (i.e. consumed through its constructor argument list), the identifier property will be populated first to allow the resolution of cyclic object references. After that, all non-transient properties that have not already been populated by the constructor are set on the entity instance. For that we use the following algorithm:

Let’s have a look at the following entity:

Spring Data adapts specifics of Kotlin to allow object creation and mutation.

You need the following dependencies to run Listing 24:

The dependencies for the enterprise version of Neo4j 3.5 are available under the com.neo4j.test:neo4j-harness-enterprise and an appropriate repository configuration.

You need the following dependencies to run Listing 26:

The dependencies for the enterprise version of Neo4j 4.x are available under the com.neo4j.test:neo4j-harness-enterprise and an appropriate repository configuration.

The principal of configuring the connection is of course still the same with Testcontainers as shown in Listing 27. You need the following dependencies:

And a complete test:

There are some scenarios in which the above annotation does not fit your use case. One of those might be that you want to have 100% control over how the driver is initialized. With a test container running, you could do this with a nested, static configuration class like this:

If you want to use the properties but cannot use a @DynamicPropertySource, you would use an initializer:

Unresolved directive in index.adoc - include::../../../../spring-data-commons/src/main/asciidoc/auditing.adoc[leveloffset=+1]

These days, a Spring boot deployment is the most likely deployment of a Spring Data based applications. Please use Spring Boots dependency management to change the driver version like this:

Or

Without Spring Boot, you would just manually declare the dependency. For Maven, we recommend using the <dependencyManagement /> section like this:

Configure the database name to use in your Spring Boot configuration like this (the same property applies of course for YML or environment based configuration, with Spring Boot’s conventions applied):

With that configuration in place, all queries generated by all instances of SDN repositories (both reactive and imperative) and by the ReactiveNeo4jTemplate respectively Neo4jTemplate will be executed against the database yourDatabase.

Provide a bean with the type Neo4jDatabaseNameProvider or ReactiveDatabaseSelectionProvider depending on the type of your Spring application.

That bean could use for example Spring’s security context to retrieve a tenant. Here is a working example for an imperative application secured with Spring Security:

Spring Boot comes with both imperative and reactive Neo4j health indicators. Both variants are able to detect multiple beans of org.neo4j.driver.Driver inside the application context and provide a contribution to the overall health for each instance. The Neo4j driver however does connect to a server and not to a specific database inside that server. Spring Boot is able to configure the driver without Spring Data Neo4j and as the information which database is to be used is tied to Spring Data Neo4j, this information is not available to the built-in health indicator.

This is most likely not a problem in many deployment scenarios. However, if configured database user does not have at least access rights to the default database, the health checks will fail.

This can be mitigated by custom Neo4j health contributors that are aware of the database selection.

The fragment DomainResults declares one additional method findMoviesAlongShortestPath:

This method is annotated with @Transactional(readOnly = true) to indicate that readers can answer it. It cannot be derived by SDN but would need a custom query. This custom query is provided by the one implementation of that interface. The implementation has the same name with the suffix Impl:

The template has overloads for String-based queries as well, so you could write down the query as String as well. The important takeaway here is:

Often times a custom query indicates custom results. Should all of those results be mapped as @Node? Of course not! Many times those objects represents read commands and are not meant to be used as write commands. It is also not unlikely that SDN 6 cannot or want not map everything that is possible with Cypher. It does however offer several hooks to run your own mapping: On the Neo4jClient. The benefit of using the SDN 6 Neo4jClient over the driver:

Declaring the fragment is exactly the same as before:

Without an implementation for that fragment, startup would fail, so here it is:

The whole query runs in the context of a Spring transaction, in this case, a read-only one.

We provide two abstract configuration classes to support you in bringing in the necessary beans: AbstractNeo4jConfig for imperative database access and AbstractReactiveNeo4jConfig for the reactive version. They are meant to be used with @EnableNeo4jRepositories and @EnableReactiveNeo4jRepositories respectively. See Listing 51 and Listing 52 for an example usage. Both classes require you to override driver() in which you are supposed to create the driver.

To get the imperative version of the Neo4j client, the template and support for imperative repositories, use something similar as shown here:

The following listing provides the reactive Neo4j client and template, enables reactive transaction management and discovers Neo4j related repositories:

For your convenience we provide a CDI extension with Neo4jCdiExtension. When run in a compatible CDI 2.0 container, it will be automatically be registered and loaded through Java’s service loader SPI.

The only thing you have to bring into your application is an annotated type that produces the Neo4j Java Driver:

If you are running in a SE Container - like the one Weld provides for example, you can enable the extension like that:

Interactions with a Neo4j Client usually ends with a call to

The imperative version will interact at this moment with the database and get the requested results or summary, wrapped in an Optional<> or a Collection.

The reactive version will in contrast return a publisher of the requested type. Interaction with the database and retrieval of the results will not happen until the publisher is subscribed to. The publisher can only be subscribed once.

As with most things in SDN, both clients depend on a configured driver instance.

The driver can only open a reactive session against a 4.0 database and will fail with an exception on any lower version.

Our Spring Boot starter provide a ready to use bean of the Neo4j Client that fits the environment (imperative or reactive) and you usually don’t have to configure your own instance.

Beside the find/load operations the save operation is one of the most used when working with data. A save operation call in general issues multiple statements against the database to ensure that the resulting graph model matches the given Java model.

The load documentation will not only show you how the MATCH part of the query looks like but also how the data gets returned.

The simplest kind of load operation is a findById call. It will match all nodes with the label of the type you queried for and does a filter on the id value.

MATCH (n:Person) WHERE id(n) = 1364

If there is a custom id provided SDN will use the property you have defined as the id.

MATCH (n:Person) WHERE n.customId = 'anId'

The data to return is defined as a map projection.

RETURN n{.first_name, .personNumber, __internalNeo4jId__: id(n), __nodeLabels__: labels(n)}

As you can see there are two special fields in there: The __internalNeo4jId__ and the __nodeLabels__. Both are critical when it comes to mapping the data to Java objects. The value of the __internalNeo4jId__ is either id(n) or the provided custom id but in the mapping process one known field to refer to has to exist. The __nodeLabels__ ensures that all defined labels on this node can be found and mapped. This is needed for situations when inheritance is used and you query not for the concrete classes or have relationships defined that only define a super-type.

Talking about relationships: If you have defined relationships in your entity, they will get added to the returned map as pattern comprehensions. The above return part will then look like:

RETURN n{.first_name, …​, Person_Has_Hobby: [(n)-[:Has]→(n_hobbies:Hobby)|n_hobbies{__internalNeo4jId__: id(n_hobbies), .name, nodeLabels: labels(n_hobbies)}]}

The map projection and pattern comprehension used by SDN ensures that only the properties and relationships you have defined are getting queried.

In cases where you have self-referencing nodes or creating schemas that potentially lead to cycles in the data that gets returned, SDN falls back to a cascading / data-driven query creation. Starting with an initial query that looks for the specific node and considering the conditions, it steps through the resulting nodes and, if their relationships are also mapped, would create further queries on the fly. This query creation and execution loop will continue until no query finds new relationships or nodes. The way of the creation can be seen analogue to the save/update process.

Given are a graph that looks like this:

and a domain model as shown in the mapping (Constructors and accessors have been omitted for brevity):

As you see, the relationships are only outgoing. Generated finder methods (including findById) will always try to match a root node to be mapped. From there on onwards, all related objects will be mapped. In queries that should return only one object, that root object is returned. In queries that return many objects, all matching objects are returned. Out- and incoming relationships from those objects returned are of course populated.

Assume the following Cypher query:

It follows the recommendation from Getting one record per root node and it works great for the leaf node you want to match here. However: That is only the case in all scenarios that return 0 or 1 mapped objects. While that query will populate all relationships like before, it won’t return all 4 objects.

This can be changed by returning the whole path:

Here we do want to use the fact that the path p actually returns 3 rows with paths to all 4 nodes. All 4 nodes will be populated, linked together and returned.

You do this exactly the same way as in a standard Cypher query issued in the Neo4j Browser or the Cypher-Shell, with the $ syntax (from Neo4j 4.0 on upwards, the old {foo} syntax for Cypher parameters has been removed from the database).

Spring Expression Language (SpEL) can be used in custom queries inside :#{}. This is the standard Spring Data way of defining a block of text inside a query that undergoes SpEL evaluation.

The following example basically defines the same query as above, but uses a WHERE clause to avoid even more curly braces:

The SpEL blocked starts with :#{ and then refers to the given String parameters by name (#pt1). Don’t confuse this with the above Cypher syntax! The SpEL expression concatenates both parameters into one single value that is eventually passed on to the Neo4jClient. The SpEL block ends with }.

SpEL also solves two additional problems. We provide two extensions that allow to pass in a Sort object into custom queries. Remember Listing 37 from custom queries? With the orderBy extension you can pass in a Pageable with a dynamic sort to a custom query:

The literal extension can be used to make things like labels or relationship-types "dynamic" in custom queries. Neither labels nor relationship types can be parameterized in Cypher, so they must be given literal.

Here, the literal value has been used to match dynamically on a Label. If you pass in SomeLabel as a parameter to the method, MATCH (n:SomeLabel) RETURN n will be generated. Ticks have been added to correctly escape values. SDN won’t do this for you as this is probably not what you want in all cases.

SDN+OGM has had quite a history over the years and we understand that migrating big application systems is neither fun nor something that provides immediate profit. The main issues we observed when migrating from older versions of Spring Data Neo4j to newer ones are roughly in order the following:

First, you must make sure that your application runs against Neo4j in server mode over the Bolt protocol, which means work in two of three cases:

Once you have made sure, that your SDN+OGM application works over Bolt as expected, you can start migrating to SDN.

And finally, add the new dependency, see Chapter 4 for both Gradle and Maven.

You’re then ready to replace annotations:

The following sections are alternatives and roughly sorted by increased effort.

All builds require a local copy of the project:

Before you proceed, verify your locally installed JDK version. The output should be similar:

There is no quality gate in place at the moment to ensure that the code/test ratio stays as is, but please consider adding tests to your contributions.

We have some rather mild checkstyle rules in place, enforcing more or less default Java formatting rules. Your build will break on formatting errors or something like unused imports.