Introduction 

Prior to CDAP 6.0.0, extensions were often added as CDAP applications. Data Prep, Analytics, and Reports were all implemented as CDAP applications, with Data Prep and Analytics running in each namespace they are required in, and Reports running in the system namespace. Running an application in each namespace wastes resources, so it is desirable to move Data Prep and Analytics into the system namespace. However, each of them have namespaced entities. When moved to the system namespace, they both need some sort of namespacing capabilities. The Reports application also is contain part of its logic in the application, but part in the CDAP system itself, due to requirements around accessing system information (run records). In order to cleanly implement these extensions as applications, additional functionality must be provided for system applications. 

Goals

Manage a single Data Prep, Analytics, and Reports application for use across all namespaces. 

User Stories 

1. As a CDAP admin, I want to manage a single system application and not an application per namespace
2. As a CDAP admin, I want to be able to dynamically scale the Services run by system applications
3. As a CDAP admin, I do not want users to be able to create Analytics experiments in a namespace that does not exist
4. As a CDAP admin, I want system applications to use the same backend storage system as the CDAP platform
5. As a CDAP user, I want Analytics experiments and models to be local to a namespace
6. As a CDAP user, I want Analytics experiments and models in a namespace to be deleted when the namespace is deleted
7. As a CDAP system developer, I want to be able to receive notifications of system events (program lifecycle events, entity creation and deletion, etc.)
8. As a CDAP system developer, I want my namespaced resources to be deleted when the namespace is deleted
9. As a CDAP system developer, I want to be able to instantiate user scope plugins in a system application
10. As a CDAP system developer, I want system applications to have the same privileges as the CDAP system
11. As a CDAP system developer, I want to use the same storage SPI as the CDAP system code

Design

This design will focus on the needs that have been brought up by the Data Prep, Analytics, and Reports applications. Future system applications may require additional functionality, but that is out of scope of this design.

All system apps will be moved to run in the system namespace instead of having one application per namespace. The system apps also need to be changed to be namespace-aware, meaning they need to explicitly take a namespace in their requests and store data in such a way that namespace isolation is achieved. Reports is cross namespace by nature and does not need to worry about namespace isolation.

The following entities need to be namespaced:

The introduction of a namespace concept to application specific entities (connections, experiments, etc) is explicitly handled by each system app. The system apps are responsible for managing namespaces themselves, using a data model that meets their needs. CDAP will be extended to ensure apps have the required capabilities to do this.

REST

Namespace needs to be added as a prefix to all Data Prep and Analytics endpoints that manage namespaced entities. We will call it a 'context' to avoid confusion with the application's namespace. For example, Data Prep connection and workspace endpoints will all be prefixed by:

/v2/contexts/<context>/connections
/v2/contexts/<context>/workspaces

Similarly, all Analytics endpoints will be prefixed by:

/v2/contexts/<context>/experiments

The full API would look like:

GET /v3/namespaces/system/apps/dataprep/services/service/methods/v2/contexts/default/connections
GET /v3/namespaces/system/apps/ModelManagementApp/spark/ModelManagerService/methods/v2/contexts/default/experiments

Data Model

System datasets have a usage pattern where they use a Table based dataset as a metadata and entity store for CRUD operations, and a FileSet based dataset as a blob store. The dataset types used before 6.0.0 are:

System applications will require these dataset types, or dataset types that are comparable in functionality. Reports also has a hidden requirement for listening to system events, and more specifically program state changes. It is hidden today because the platform itself listens to those events and writes to a separate Table that the app eventually scans. In order to move everything into the app, the app should have the ability to listen to those events. This can be done through the MessageFetcher interface that is already exposed to apps, with the system namespace passed as an argument.

Accessing the system table SPI

Creating System Tables

System apps need to be able to create system tables in order to use them at runtime. Creation of system tables can be done through a new SystemTableConfigurer interface that the ServiceConfigurer will extend. 

public interface SystemTableConfigurer {
  
  /**
   * Create a system table that conforms to the given table specification when the
   * application is deployed. If the table already exists, nothing happens.
   *
   * @throws UnsupportedOperationException if the application is not a system application
   */
  createTable(StructuredTableSpecification tableSpecification);
}

If an application that is not deployed in the system namespace tries to create a system table, an exception is thrown and app deployment fails.

Note that this approach requires that the SPI is moved into its own module with minimal dependencies, and requires that cdap-api adds the spi as a dependency. If we do not want this, an alternative is to create a new cdap-system-api module that depends on both cdap-api and cdap-spi. The cdap-system-api would contain a special type of service:

public abstract class AbstractSystemService extends AbstractPluginConfigurable<SystemServiceConfigurer> implements Service {

  @Override
  public final void configure(SystemServiceConfigurer serviceConfigurer) {
    this.configurer = serviceConfigurer;
    configure();
  }
}

Only this special type of service would have access to the SystemTableConfigurer. Only applications deployed in the system namespace would be allowed to extend this class. If an application in a user namespace attempts to extend the class, deployment will fail.

Using System Tables

System apps need to be able to perform CRUD operations on the system tables at runtime. In order to do this, the HttpServiceContext will implement the TransactionRunner interface, allowing system apps to write code like:

@GET
@Path("entities/{entity}")
public void getEntity(HttpServiceRequest request, HttpServiceResponder responder,
                      @PathParam("entity") String entity) {
  getContext().execute(tableContext -> {
    StructuredTable table = tableContext.getTable(TABLE_ID);
    table.read(...);
    ...
  });
}

If the application is not running in the system namespace, the execute() call will throw an UnsupportedOperationException. CDAP will also need to know to ship the SPI implementation jars along with the container for any system app. 

Similar considerations about whether the SPI should be available from cdap-api apply here to using system tables at runtime, with an alternative being introducing a special SystemHttpHandler.

Note: Since the system apps will not be running on the same node(s) as the CDAP master, CDAP administrators will need to ensure that any node in their cluster is allowed to talk to the underlying SPI implementation. For example, if Postgres is being used, it must allow incoming connections from any node on the cluster. For HBase, we need to ensure that system apps run as the same user as the CDAP master, using the same auth tokens.

Approach 1

In approach 1, in order to achieve namespace isolation and automatic entity deletion when a namespace is deleted, system apps will create a dataset instance in each namespace it needs. For example, a connection in the 'default' namespace will be stored in a Table in the 'default' namespace. In this way, isolation and automatic deletion is handled by the platform; the app only needs to use the correct dataset.

This approach has the nice property that namespacing does not need to be considered in the application's data model. For example, the entity 'context' (namespace) does not need to be present in the backend table schema. Isolation and automatic deletion are handled by the platform. On the downside, it duplicates every backend table for each namespace, with each table likely containing a very small number of rows. This splits state and makes upgrade more difficult. It also requires the system table SPI to be namespaced, which is not a natural concept for tables within the CDAP system.

Programmatic APIs

In order to implement namespace logic, system apps need to be able to perform several operations that are not supported prior to CDAP 6.0.0.

1. Check namespace existence – System apps need to be able to check that a namespace exists in CDAP before managing any custom entities in that namespace.

2. Dataset admin operations in another namespace – System apps need to be create a dataset in a specific namespace if it doesn't already exist. DatasetManager methods are not namespace aware and currently only operate within the namespace of the application.

3. Plugin operations in another namespace – DataPrep needs to be able to instantiate UDDs (User Defined Directives) in order to execute directive lists. This means system applications need to be able to instantiate plugins whose artifacts are user scoped in some namespace. To give a more concrete example, suppose a 'my-custom-directive' UDD is deployed as a user artifact in namespace 'default'. The Data Prep system application needs to be able to instantiate that directive even though the app is running in the 'system' namespace.

To support this, several existing interfaces are enhanced with namespaced versions of their existing methods. DatasetContext already has a way to get a dataset from another namespace, but DatasetManager cannot check existence of or create a dataset in another namespace. DatasetManager will need to be enhanced with namespaced versions of its current methods:

public interface DatasetManager {
  boolean datasetExists(String name) throws DatasetManagementException;
  boolean datasetExists(String namespace, String name) throws DatasetManagementException;
  ...
  void createDataset(String name, String type, DatasetProperties properties) throws DatasetManagementException;
  void createDataset(String namespace, String name, String type, DatasetProperties properties) throws DatasetManagementException;
  ...
}

Similarly, the ArtifactManager interface available to Service programs must also be modified.

public interface ArtifactManager {
  List<ArtifactInfo> listArtifacts() throws IOException;
  List<ArtifactInfo> listArtifacts(String namespace) throws IOException;  

  CloseableClassLoader createClassLoader(ArtifactInfo artifactInfo,
                                         @Nullable ClassLoader parentClassLoader) throws IOException;
}

The Admin interface will be enhanced to check for namespace existence:

public interface Admin extends DatasetManager, SecureStoreManager, MessagingAdmin {
  boolean namespaceExists(String namespace);


  NamespaceMeta getNamespace(String namespace);
}

This has a side benefit of bringing more consistency to the APIs instead of having mixed APIs that sometimes allow cross namespace operations and sometimes don't.

System app service methods would typically look something like:

@Path("/v2")
public class ModelManagerServiceHandler implements SparkHttpServiceHandler {

  @GET
  @Path("/context/{context}/experiments")
  public void listExperiments(HttpServiceRequest request, HttpServiceResponder responder,
                              @PathParam("context") String context) {
    Admin admin = getContext().getAdmin();
    if (!admin.namespaceExists(context)) {
      responder.sendStatus(404, "Namespace " + namespace + " not found.");
    }
    if (!admin.datasetExists(EXPERIMENTS_DATASET)) {
      admin.createDataset(context, EXPERIMENTS_DATASET, "table", EXPERIMENTS_DATASET_PROPERTIES);
    }
    getContext().execute(datasetContext -> {
      Table experiments = datasetContext.getDataset(context, EXPERIMENTS_DATASET);
      ...
      responder.sendJson(...);
      });
  }
  ...
}

With this approach, the developer is forced into writing quite a lot of repeat code to check if the namespace exists. For convenience, the platform could support some syntactic sugar to check namespace for the user if the endpoint is annotated with a new Namespaced annotation. For example, something like:

@Path("/v2")
public class ModelManagerServiceHandler implements SparkHttpServiceHandler {

  @GET
  @Namespaced(namespace = "context")
  @Path("/contexts/{context}/experiments")
  public void listExperiments(HttpServiceRequest request, HttpServiceResponder responder,
                              @PathParam("context") String context) {
    Admin admin = getContext().getAdmin();
    if (!admin.datasetExists(EXPERIMENTS_DATASET)) {
      admin.createDataset(context, EXPERIMENTS_DATASET, "table", EXPERIMENTS_DATASET_PROPERTIES);
    }
    getContext().execute(datasetContext -> {
      Table experiments = datasetContext.getDataset(context, EXPERIMENTS_DATASET);
      ...
      responder.sendJson(...);
      });
  }
  ...
}

Note that there is still a lot of repeat code around creating the dataset if it doesn't exist.

Authorization

An application running in the system namespace will have the same privileges as the CDAP system. There will need to be some extra work around impersonation and authorization for all methods exposed in the context objects to ensure that the right user is being used. 

For example, when creating a dataset in a namespace, the CDAP system user should be authorized to create it, but the namespace user should actually own the dataset.

Upgrade

When upgrading from CDAP 5.1.x to 6.0.0, no additional work needs to be done regarding datasets, as they will already be in their respective namespaces. Old versions of the DataPrep and Analytics apps will remain in any namespace they were enabled in, so it should be documented that users should delete these apps. CDAP could also provide a tool to do this cleanup. 

For future upgrades (CDAP 6.0.x to 6.1.0 and beyond), every table in every namespace must be upgraded.

Approach 2

In the second approach, the system application maintain a set of backend table in the system namespace and builds namespacing into the table schema. Isolation between namespaces must be implemented by the application. In addition, the application needs to subscribe to system events and delete entities when it sees a namespace deletion event.

This approach has the benefit of matching the CDAP system storage architecture, where there is a single set of tables for all namespaces and not a table per namespace. Upgrade is easier in the future, as only a single set of tables must be upgraded. Authorization is also simpler, as everything happens in system scope and cross namespace authorization never comes into the picture. On the downside, the application needs to listen to namespace deletion events and delete relevant entities itself. This has to be done in a worker program that is always running, which uses one more container than before.

The service endpoints are straightforward and will look something like:

@Path("/v2")
public class ModelManagerServiceHandler implements SparkHttpServiceHandler {

  @GET
  @Path("/contexts/{context}/experiments")
  public void listExperiments(HttpServiceRequest request, HttpServiceResponder responder,
                              @PathParam("context") String context) {
    Admin admin = getContext().getAdmin();
    if (!admin.namespaceExists(context)) {
      responder.sendStatus(404, "Namespace " + namespace + " not found.");
    }
    // use SPI to list experiments in context
  }
  ...
}

Namespacing

Each entity table will contain 'context' (namespace) as a column. The context will be the first part of the primary key. In this way, listing, get, delete, and update will always require context to be specified, ensuring entities don't get mixed across contexts.

Namespace Deletion

The application will need a worker to listen to namespace delete events. The worker will look something like:

public static class SystemEventListener extends AbstractWorker {

  private volatile boolean stopped;

  @Override
  public void run() {
    while (!stopped) {
      String messageId = getLastMessageId();
      try (CloseableIterator<Message> iter = getContext().getMessageFetcher().fetch("system", "topic", 100, messageId)) {
        while (iter.hasNext()) {
          MetadataMessage event = GSON.fromJson(iter.next().getPayloadAsString(), MetadataMessage.class);
          EntityId deletedEntity = event.getEntityId();
          MetadataMessage.Type eventType = event.getType();
          if (eventType == MetadataMessage.Type.ENTITY_DELETION && deletedEntity.getType() == EntityType.NAMESPACE) {
            String context = deletedEntity.getNamespace();
            // delete entities from storage
            ...
          }
          saveMessageId(message.getId());
        }
      }
      TimeUnit.MILLISECONDS.sleep(200);
    }
  }

  @Override
  public void stop() {
    stopped = true;
  }


  private String getLastMessageId() {
    // lookup message ID from storage
  }


  private void saveMessageId() {
    // write message ID to storage
  }
}

In order to prevent race conditions when a namespace is deleted and immediately re-created, each entity will contain a creation time. The namespace metadata will also contain the time that it was created. Based on these times, the application can filter out entities that should not be returned, and actual deletion can be handled in the background.

Scheduling Logic

One main downside to the worker approach is that the worker will be doing nothing almost all of the time. It also needs to manage extra state about the topic offset and understand how to decode the messages.

A better way to handle this would be to allow applications to schedule some logic to be run whenever a system event occurs. Similar to how workflows can be triggered based on program lifecycle events, system applications would be able to trigger lambda functions based on system events. This will simplify the worker code by removing the state logic, removing the polling loop, and ideally removing the need to understand the message format. The code would look more like:

public static class NamespaceDeleteListener extends SystemEventListener {

  @Override
  public void run(MetadataMessage event) {
    String context = event.getEntityId().getNamespace();
    // delete entities from storage
    ...
  }

}

This would involve introducing a new program type and APIs to schedule that type, which could look something like:

addEventListener(new NamespaceDeleteListener());


schedule(buildSchedule("namespaceDelete", MetadataMessage.Type.ENTITY_DELETION, EntityType.NAMESPACE));

API changes

New Programmatic APIs

There will be new programmatic APIs to allow more cross namespace access.

Deprecated Programmatic APIs

None

New REST APIs

There are no new CDAP APIs, but almost every Data Prep and Analytics endpoint will be modified to include namespace in the path.

Deprecated REST API

None

CLI Impact or Changes

• None

UI Impact or Changes

• UI needs to be updated to use the new REST APIs.

Security Impact 

Care needs to be taken to ensure that system applications are authorized to access CDAP entities in other namespaces.

Impact on Infrastructure Outages 

None

Test Scenarios

Releases

Release 6.0.0

Related Work

• Work #1
• Work #2
• Work #3

Future work