datahub/docs/how/graph-onboarding.md
2020-09-23 14:58:35 -07:00

201 lines
7.4 KiB
Markdown

# How to onboard to GMA graph?
## 1. Define relationship models
Relationship models are used to build edges in the graph.
If you need to define a [relationship] which is not available in the set of [relationship models] provided,
that relationship model should be implemented as a first step for graph onboarding.
Below is an example model for `OwnedBy` relationship:
```
namespace com.linkedin.metadata.relationship
import com.linkedin.common.OwnershipType
/**
* A generic model for the Owned-By relationship
*/
@pairings = [ {
"destination" : "com.linkedin.common.urn.CorpuserUrn",
"source" : "com.linkedin.common.urn.DatasetUrn"
} ]
record OwnedBy includes BaseRelationship {
/** The type of the ownership */
type: OwnershipType
}
```
Fields in this model are translated to properties of the graph edge.
Also, the FQCN of the relationship model, which is `com.linkedin.metadata.relationship.OwnedBy` in this example, is used as the label for edges.
## 2. Define entity models
Entity models are used to build nodes in the graph.
Every GMA [entity] should have its own entity model defined and placed under [entity models] directory.
Below is an example model for `DatasetEntity` relationship.
```
namespace com.linkedin.metadata.entity
import com.linkedin.common.DataPlatformUrn
import com.linkedin.common.DatasetUrn
import com.linkedin.common.FabricType
/**
* Data model for a dataset entity
*/
record DatasetEntity includes BaseEntity {
/**
* Urn for the dataset
*/
urn: DatasetUrn
/**
* Dataset native name e.g. {db}.{table}, /dir/subdir/{name}, or {name}
*/
name: optional string
/**
* Platform urn for the dataset in the form of urn:li:platform:{platform_name}
*/
platform: optional DataPlatformUrn
/**
* Fabric type where dataset belongs to or where it was generated.
*/
origin: optional FabricType
}
```
Fields in this model are translated to properties of the graph node.
Also, the FQCN of the entity model, which is `com.linkedin.metadata.entity.DatasetEntity` in this case, is used as the label for nodes.
## 3. Implement relationship builders
You need to implement relationship builders for your specific [aspect]s and [relationship]s if they are not already defined.
Relationship builders build list of relationships after processing aspects and any relationship builder should implement `BaseRelationshipBuilder` abstract class.
Relationship builders are per aspect and per relationship type.
```java
public abstract class BaseRelationshipBuilder<ASPECT extends RecordTemplate> {
private Class<ASPECT> _aspectClass;
public BaseRelationshipBuilder(Class<ASPECT> aspectClass) {
_aspectClass = aspectClass;
}
/**
* Returns the aspect class this {@link BaseRelationshipBuilder} supports
*/
public Class<ASPECT> supportedAspectClass() {
return _aspectClass;
}
/**
* Returns a list of corresponding relationship updates for the given metadata aspect
*/
public abstract <URN extends Urn> List<GraphBuilder.RelationshipUpdates> buildRelationships(URN urn, ASPECT aspect);
}
```
## 4. Implement graph builders
Graph builders build graph updates by processing [snapshot]s.
They internally use relationship builders to generate edges and nodes of the graph.
All relationship builders for an [entity] should be registered through graph builder.
```java
public abstract class BaseGraphBuilder<SNAPSHOT extends RecordTemplate> implements GraphBuilder<SNAPSHOT> {
private final Class<SNAPSHOT> _snapshotClass;
private final Map<Class<? extends RecordTemplate>, BaseRelationshipBuilder> _relationshipBuildersMap;
public BaseGraphBuilder(@Nonnull Class<SNAPSHOT> snapshotClass,
@Nonnull Collection<BaseRelationshipBuilder> relationshipBuilders) {
_snapshotClass = snapshotClass;
_relationshipBuildersMap = relationshipBuilders.stream()
.collect(Collectors.toMap(builder -> builder.supportedAspectClass(), Function.identity()));
}
@Nonnull
Class<SNAPSHOT> supportedSnapshotClass() {
return _snapshotClass;
}
@Nonnull
@Override
public GraphUpdates build(@Nonnull SNAPSHOT snapshot) {
final Urn urn = RecordUtils.getRecordTemplateField(snapshot, "urn", Urn.class);
final List<? extends RecordTemplate> entities = buildEntities(snapshot);
final List<RelationshipUpdates> relationshipUpdates = new ArrayList<>();
final List<RecordTemplate> aspects = ModelUtils.getAspectsFromSnapshot(snapshot);
for (RecordTemplate aspect : aspects) {
BaseRelationshipBuilder relationshipBuilder = _relationshipBuildersMap.get(aspect.getClass());
if (relationshipBuilder != null) {
relationshipUpdates.addAll(relationshipBuilder.buildRelationships(urn, aspect));
}
}
return new GraphUpdates(Collections.unmodifiableList(entities), Collections.unmodifiableList(relationshipUpdates));
}
@Nonnull
protected abstract List<? extends RecordTemplate> buildEntities(@Nonnull SNAPSHOT snapshot);
}
```
```java
public class DatasetGraphBuilder extends BaseGraphBuilder<DatasetSnapshot> {
private static final Set<BaseRelationshipBuilder> RELATIONSHIP_BUILDERS =
Collections.unmodifiableSet(new HashSet<BaseRelationshipBuilder>() {
{
add(new DownstreamOfBuilderFromUpstreamLineage());
add(new OwnedByBuilderFromOwnership());
}
});
public DatasetGraphBuilder() {
super(DatasetSnapshot.class, RELATIONSHIP_BUILDERS);
}
@Nonnull
@Override
protected List<? extends RecordTemplate> buildEntities(@Nonnull DatasetSnapshot snapshot) {
final DatasetUrn urn = snapshot.getUrn();
final DatasetEntity entity = new DatasetEntity().setUrn(urn)
.setName(urn.getDatasetNameEntity())
.setPlatform(urn.getPlatformEntity())
.setOrigin(urn.getOriginEntity());
setRemovedProperty(snapshot, entity);
return Collections.singletonList(entity);
}
}
```
## 5. Ingestion into graph
The ingestion process for each [entity] is done by graph builders.
The builders will be invoked whenever an [MAE] is received by [MAE Consumer Job].
Graph builders should be extended from BaseGraphBuilder. Check DatasetGraphBuilder as an example above.
For the consumer job to consume those MAEs, you should add your graph builder to the [graph builder registry].
## 6. Graph queries
You can onboard the graph queries which fit to your specific use cases using [Query DAO].
You also need to create [rest.li](https://rest.li) APIs to serve your graph queries.
[BaseQueryDAO] provides an abstract implementation of several graph query APIs.
Refer to [DownstreamLineageResource] rest.li resource implementation to see a use case of graph queries.
[relationship]: ../what/relationship.md
[relationship models]: ../../metadata-models/src/main/pegasus/com/linkedin/metadata/relationship
[entity models]: ../../metadata-models/src/main/pegasus/com/linkedin/metadata/entity
[aspect]: ../what/aspect.md
[snapshot]: ../what/snapshot.md
[entity]: ../what/entity.md
[mae]: ../what/mxe.md#metadata-audit-event-mae
[mae consumer job]: ../architecture/metadata-ingestion.md#mae-consumer-job
[graph builder registry]: ../../metadata-builders/src/main/java/com/linkedin/metadata/builders/graph/RegisteredGraphBuilders.java
[query dao]: ../architecture/metadata-serving.md#query-dao
[BaseQueryDAO]: ../../metadata-dao/src/main/java/com/linkedin/metadata/dao/BaseQueryDAO.java
[DownstreamLineageResource]: ../../gms/impl/src/main/java/com/linkedin/metadata/resources/dataset/DownstreamLineageResource.java