Skip to main content
Version: Next

A Custom Metadata Model

This module hosts a gradle project where you can store your custom metadata model. It contains an example extension for you to follow.

Caveats​

Currently, this project only supports aspects defined in PDL to existing or newly defined entities. You cannot add new aspects to the metadata model directly through yaml configuration yet.

Pre-Requisites​

Before proceeding further, make sure you understand the DataHub Metadata Model concepts defined here and extending the model defined here.

Create your new aspect(s)​

Follow the regular process in creating a new aspect by adding it to the src/main/pegasus folder. e.g. This repository has an Aspect called customDataQualityRules hosted in the DataQualityRules.pdl file that you can follow. Once you've gone through this exercise, feel free to delete the sample aspects that are stored in this module.

Tip: PDL requires that the name of the file must match the name of the class that is defined in it and the package path must also match the directory path, so keep that in mind when you create your aspect pdl file.

Add your aspect(s) to the entity registry​

Add your new aspect(s) to the entity registry by editing the yaml file located under registry/entity-registry.yaml. Note: The registry file must be called entity-registry.yaml or entity-registry.yml for it to be recognized.

Understanding the entity registry​

Here is a sample entity-registry file

id: mycompany-dq-model
entities:
  - name: dataset
    aspects:
      - customDataQualityRules

The entity registry has a few important fields to pay attention to:

  • id: The name of your registry. This drives naming, artifact generation, so make sure you pick a unique name that will not conflict with other names you might create for other registries.
  • entities: A list of entities with aspects attached to them that you are creating additional aspects for as well as any new entities you wish to define. In this example, we are adding the aspect customDataQualityRules to the dataset entity.

Build your new model​

Change your directory to the metadata-models-custom folder and then run this command

../gradlew build

This will create a zip file in the build/dist folder. Then change your directory back to the main datahub folder and run 

./gradlew :metadata-models-custom:install

This will install the zip file as a datahub plugin. It is installed at ~/.datahub/plugins/models/ and if you list the directory you should see the following path if you are following the customDataQualityRules implementation example: ~/.datahub/plugins/models/mycompany-dq-model/0.0.0-dev/

Build a versioned artifact​

../gradlew -PprojVersion=0.0.1 build

This will deposit an artifact called metadata-models-custom-<version>.zip under the build/dist directory.

Deploy your versioned artifact to DataHub​

../gradlew -PprojVersion=0.0.1 install

This will unpack the artifact and deposit it under ~/.datahub/plugins/models/<registry-name>/<registry-version>/.

Deploying to a remote Kubernetes server​

Deploying your customized jar to a remote Kubernetes server requires that you take the output zip (generated from ../gradlew modelArtifact under build/dist) and place the unzipped contents in the volumes mount for the GMS pod on the remote server. First you will need to push the files into a configmap using kubectl:

kubectl create configmap custom-model --from-file=<<path-to-file>> -n <<namespace>>

Then you need to set the volumes for GMS (refer to how jmx exporter configmap is added here: https://github.com/acryldata/datahub-helm/blob/master/charts/datahub/subcharts/datahub-gms/templates/deployment.yaml#L40) This tells GMS that we will be pulling this configmap in. You can do this by setting datahub-gms.extraVolumes in values.yaml which gets appended to the deployment without having to change the helm chart.

Finally you need to mount the volume into the container’s local directory by setting volumeMounts. Refer to how the kafka certs are mounted onto a local path here: https://github.com/acryldata/datahub-helm/blob/master/charts/datahub/subcharts/datahub-gms/templates/deployment.yaml#L182 You can do this by setting the datahub-gms.extraVolumeMounts in values.yaml

at the end your values.yaml should have something like:

datahub-gms:
  ...
  extraVolumes:
    - name: custom-model
      configMap:
        name: custom-model ## should match configmap name above
  extraVolumeMounts:
    - name: custom-model-dir
      mountPath: /etc/plugins/models/<registry-name>/<registry-version>

The mountPath can be configured using ENTITY_REGISTRY_PLUGIN_PATH and defaults to /etc/datahub/plugins/models.

Check if your model got loaded successfully​

Assuming that you are running DataHub on localhost, you can curl the config endpoint to see the model load status. 

curl -s http://localhost:8080/config | jq .
{
  "models": {
    "mycompany-dq-model": {
      "0.0.1": {
        "loadResult": "SUCCESS",
        "registryLocation": "/Users/username/.datahub/plugins/models/mycompany-dq-model/0.0.1",
        "failureCount": 0
      }
    }
  },
  "noCode": "true"
}

Alternatively, you could type in http://localhost:8080/config in your browser.

Add some metadata with your new model​

We have included some sample scripts that you can modify to upload data corresponding to your new data model. The scripts/insert_one.sh script takes the scripts/data/dq_rule.json file and attaches it to the dataset_urn entity using the datahub cli. 

cd scripts
./insert_one.sh

results in 

Update succeeded with status 200

The scripts/insert_custom_aspect.py script shows you how to accomplish the same using the Python SDK. Note that we are just using a raw dictionary here to represent the dq_rule aspect and not a strongly-typed class.

cd scripts
python3 insert_custom_aspect.py

results in

Successfully wrote to DataHub

Advanced Guide​

A few things that you will likely do as you start creating new models and creating metadata that conforms to those models.

Deleting metadata associated with a model​

The datahub cli supports deleting metadata associated with a model as a customization of the delete command.

e.g. datahub delete by-registry --registry-id=mycompany-dq-model:0.0.1 --hard will delete all data written using this registry name and version pair.

Evolve the metadata model​

As you evolve the metadata model, you can publish new versions of the repository and deploy it into DataHub as well using the same steps outlined above. DataHub will check whether your new models are backwards compatible with the previous versioned model and decline loading models that are backwards incompatible.

Custom Plugins​

Adding custom aspects to DataHub's existing data model is a powerful way to extend DataHub without forking the entire repo. Often extending just the data model is not enough and additional custom code might be required. For a few of these use cases a plugin framework was developed to control how instances of custom aspects can be validated, mutated, and generate side effects (additional aspects).

It should be noted that validation, mutation, and generation of the core DataHub aspects can lead to system corruption and should be used by advanced users only.

The /config endpoint documented above has been extended to return information on the instances of the various plugins as well as the classes that were loaded for debugging purposes.

{
  "mycompany-dq-model": {
    "0.0.0-dev": {
      "plugins": {
        "validatorCount": 1,
        "mutationHookCount": 1,
        "mcpSideEffectCount": 1,
        "mclSideEffectCount": 1,
        "validatorClasses": [
          "com.linkedin.metadata.aspect.plugins.validation.CustomDataQualityRulesValidator"
        ],
        "mutationHookClasses": [
          "com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMutator"
        ],
        "mcpSideEffectClasses": [
          "com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMCPSideEffect"
        ],
        "mclSideEffectClasses": [
          "com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMCLSideEffect"
        ]
      }
    }
  }
}

Custom Validators​

Custom aspects might require that instances of those aspects adhere to specific conditions or rules. These conditions could vary wildly depending on the use case however they could be as simple as a null or range check for one or more fields within the custom aspect. Additionally, a lookup can be done on other aspects in order to validate the current aspect using the AspectRetriever.

There are two integration points for validation. The first integration point is on request via the validateProposedAspect method where the aspect is validated independent of the previous value. This validation is performed outside of a database transaction and can perform more intensive checks without introducing added latency within a transaction. Note that added latency from the validation check is still introduced into the request itself.

The second integration point for validation occurs within the database transaction using the validatePreCommitAspect and has access to the new aspect as well as the old aspect. See the included example in CustomDataQualityRulesValidator.java.

Shown below is the interface to be implemented for a custom validator.

public class CustomDataQualityRulesValidator extends AspectPayloadValidator {
    @Override
    protected Stream<AspectValidationException> validateProposedAspects(
            @Nonnull Collection<? extends BatchItem> mcpItems, @Nonnull AspectRetriever aspectRetriever) {
    }

    @Override
    protected Stream<AspectValidationException> validatePreCommitAspects(
            @Nonnull Collection<ChangeMCP> changeMCPs, AspectRetriever aspectRetriever) {
    }
}

In order to register this custom validator add the following to your entity-registry.yml file. This will activate the validator to run on upsert operations for any entity with the custom aspect customDataQualityRules. Alternatively separate validators could be written within the context of specific entities, in this case simply specify the entity name instead of *.


plugins:
  aspectPayloadValidators:
    - className: 'com.linkedin.metadata.aspect.plugins.validation.CustomDataQualityRulesValidator'
      enabled: true
      supportedOperations:
        - UPSERT
      supportedEntityAspectNames:
        - entityName: '*'
          aspectName: customDataQualityRules

Custom Mutator​

Warning: This hook is for advanced users only. It is possible to corrupt data and render your system inoperable.

Mutation hooks have two possible mutation points. The first is the write mutation which can change the data being written to persistent storage. The second mutation hook is a read hook which can modify the data when read from persistent storage.

Write Mutation:

In this example, we want to make sure that the field type is always lowercase regardless of the string being provided by ingestion. The full example can be found in CustomDataQualityMutator.java.

public class CustomDataQualityRulesMutator extends MutationHook {
    @Override
    protected Stream<Pair<ChangeMCP, Boolean>> writeMutation(
            @Nonnull Collection<ChangeMCP> changeMCPS, @Nonnull AspectRetriever aspectRetriever) {
        return changeMCPS.stream()
                .map(
                        changeMCP -> {
                            boolean mutated = false;

                            if (changeMCP.getRecordTemplate() != null) {
                                DataQualityRules newDataQualityRules =
                                        new DataQualityRules(changeMCP.getRecordTemplate().data());

                                for (DataQualityRule rule : newDataQualityRules.getRules()) {
                                    // Ensure uniform lowercase
                                    if (!rule.getType().toLowerCase().equals(rule.getType())) {
                                        mutated = true;
                                        rule.setType(rule.getType().toLowerCase());
                                    }
                                }
                            }

                            return mutated ? changeMCP : null;
                        })
                .filter(Objects::nonNull)
                .map(changeMCP -> Pair.of(changeMCP, true));
    }
}
plugins:
  mutationHooks:
    - className: 'com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMutator'
      enabled: true
      supportedOperations:
        - UPSERT
      supportedEntityAspectNames:
        - entityName: '*'
          aspectName: customDataQualityRules

Read Mutation:

A read mutator would implement the following interface and the following example is a read mutation which hides soft deleted structured properties from being returned on entities. 

public class StructuredPropertiesSoftDelete extends MutationHook {
    @Override
    protected Stream<Pair<ReadItem, Boolean>> readMutation(
            @Nonnull Collection<ReadItem> items, @Nonnull AspectRetriever aspectRetriever) {
        Map<Urn, StructuredProperties> entityStructuredPropertiesMap =
                items.stream()
                        .filter(i -> i.getRecordTemplate() != null)
                        .map(i -> Pair.of(i.getUrn(), i.getAspect(StructuredProperties.class)))
                        .collect(Collectors.toMap(Pair::getKey, Pair::getValue));

        // Apply filter
        Map<Urn, Boolean> mutatedEntityStructuredPropertiesMap =
                StructuredPropertyUtils.filterSoftDelete(entityStructuredPropertiesMap, aspectRetriever);

        return items.stream()
                .map(i -> Pair.of(i, mutatedEntityStructuredPropertiesMap.getOrDefault(i.getUrn(), false)));
    }
}

Note that the supportedOperations is left empty since those operation types only include change types like UPSERT or DELETE

plugins:
  mutationHooks:
    - className: 'com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMutator'
      enabled: true
      supportedEntityAspectNames:
        - entityName: '*'
          aspectName: customDataQualityRules

MetadataChangeProposal (MCP) Side Effects​

Warning: This hook is for advanced users only. It is possible to corrupt data and render your system inoperable.

MCP Side Effects allow for the creation of new aspects based on an input aspect.

Notes:

  • MCPs will write aspects to the primary data store (SQL for example) as well as the search indices.
  • Side effects in general must include a dependency on the metadata-io module since it deals with lower level storage primitives.

The full example can be found in CustomDataQualityRulesMCPSideEffect.java.

public class CustomDataQualityRulesMCPSideEffect extends MCPSideEffect {
    @Override
    protected Stream<ChangeMCP> applyMCPSideEffect(
            Collection<ChangeMCP> changeMCPS, @Nonnull AspectRetriever aspectRetriever) {
        // Mirror aspects to another URN in SQL & Search
        return changeMCPS.stream()
                .map(
                        changeMCP -> {
                            Urn mirror =
                                    UrnUtils.getUrn(changeMCP.getUrn().toString().replace(",PROD)", ",DEV)"));
                            return ChangeItemImpl.builder()
                                    .urn(mirror)
                                    .aspectName(changeMCP.getAspectName())
                                    .recordTemplate(changeMCP.getRecordTemplate())
                                    .auditStamp(changeMCP.getAuditStamp())
                                    .systemMetadata(changeMCP.getSystemMetadata())
                                    .build(aspectRetriever);
                        });
    }
}
plugins:
  mcpSideEffects:
    - className: 'com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMCPSideEffect'
      enabled: true
      supportedOperations:
        - UPSERT
      supportedEntityAspectNames:
        - entityName: '*'
          aspectName: customDataQualityRules

MetadataChangeLog (MCL) Side Effects​

Warning: This hook is for advanced users only. It is possible to corrupt data and render your system inoperable.

MCL Side Effects allow for the creation of new aspects based on an input aspect. In this example, we are generating a timeseries aspect to represent an event. When a DataQualityRule is created or modified we'll record the actor, event type, and timestamp in a timeseries aspect index.

Notes:

  • MCLs are only persisted to the search indices which allows for adding to the search documents only.
  • Dependency on the metadata-io module since it deals with lower level storage primitives.

The full example can be found in CustomDataQualityRulesMCLSideEffect.java.

public class CustomDataQualityRulesMCLSideEffect extends MCLSideEffect {
    @Override
    protected Stream<MCLItem> applyMCLSideEffect(
            @Nonnull Collection<MCLItem> mclItems, @Nonnull AspectRetriever aspectRetriever) {
        return mclItems.stream()
                .map(
                        item -> {
                            // Generate Timeseries event aspect based on non-Timeseries aspect
                            MetadataChangeLog originMCP = item.getMetadataChangeLog();

                            return buildEvent(originMCP)
                                    .map(
                                            event -> {
                                                try {
                                                    MetadataChangeLog eventMCP = originMCP.clone();
                                                    eventMCP.setAspect(GenericRecordUtils.serializeAspect(event));
                                                    eventMCP.setAspectName("customDataQualityRuleEvent");
                                                    return eventMCP;
                                                } catch (CloneNotSupportedException e) {
                                                    throw new RuntimeException(e);
                                                }
                                            })
                                    .map(
                                            eventMCP ->
                                                    MCLItemImpl.builder().metadataChangeLog(eventMCP).build(aspectRetriever));
                        })
                .filter(Optional::isPresent)
                .map(Optional::get);
    }

    private Optional<DataQualityRuleEvent> buildEvent(MetadataChangeLog originMCP) {
        if (originMCP.getAspect() != null) {
            DataQualityRuleEvent event = new DataQualityRuleEvent();
            if (event.getActor() != null) {
                event.setActor(event.getActor());
            }
            event.setEventTimestamp(originMCP.getSystemMetadata().getLastObserved());
            event.setTimestampMillis(originMCP.getSystemMetadata().getLastObserved());
            if (originMCP.getPreviousAspectValue() == null) {
                event.setEventType("RuleCreated");
            } else {
                event.setEventType("RuleUpdated");
            }
            event.setAffectedDataset(originMCP.getEntityUrn());

            return Optional.of(event);
        }
        return Optional.empty();
    }
}
plugins:
  mclSideEffects:
    - className: 'com.linkedin.metadata.aspect.plugins.hooks.CustomDataQualityRulesMCLSideEffect'
      enabled: true
      supportedOperations:
        - UPSERT
      supportedEntityAspectNames:
        - entityName: 'dataset'
          aspectName: customDataQualityRules

The Future​

Hopefully this repository shows you how easily you can extend and customize DataHub's metadata model!

We will be continuing to make the experience less reliant on core changes to DataHub and reducing the need to fork the main repository.