CDK Data Reference Guide

About This Guide

This reference guide is the primary source of documentation for the CDK Data module. It covers the high level organization of the APIs, primary classes and interfaces, intended usage, available extension points for customization, and implementation information where helpful and appropriate.

From here on, this guide assumes you are already familiar with the basic design and functionality of HDFS, Hadoop MapReduce, and Java SE 6. Users who are also familiar with Avro, data serialization techniques, common compression algorithms (e.g. gzip, snappy), advanced Hadoop MapReduce topics (e.g. input split calculation), and traditional data management topics (e.g. partitioning schemes, metadata management) will benefit even more.

Overview of the Data Module

The CDK Data module is a set of APIs for interacting with datasets in the Hadoop ecosystem. Specifically built to simplify direct reading and writing of datasets in storage subsystems such as the Hadoop Distributed FileSystem (HDFS), the Data module provides familiar, stream-oriented APIs, that remove the complexity of data serialization, partitioning, organization, and metadata system integration. These APIs do not replace or supersede any of the existing Hadoop APIs. Instead, the Data module acts as a targetted application of those APIs for its stated use case. In other words, many applications will still use the HDFS or Avro APIs directly when the developer has use cases outside of direct dataset create, drop, read, and write operations. On the other hand, for users building applications or systems such as data integration services, the Data module will usually be superior in its default choices, data organization, and metadata system integration, when compared to custom built code.

In keeping with the overarching theme and principles of the CDK, the Data module is prescriptive. Rather than present a do-all Swiss Army knife library, this module makes specific design choices that guide users toward well-known patterns that make sense for many, if not all, cases. It is likely that advanced users with niche use cases or applications will find it difficult, suboptimal, or even impossible to do unusual things. Limiting the user is not a goal, but when revealing an option creates significant opportunity for complexity, or would otherwise require the user to delve into a rathole of additional choices or topics to research, such a tradeoff has been made. The Data module is designed to be immediately useful, obvious, and in line with what most users want, out of the box.

These APIs are designed to easily fit in with dependency injection frameworks like Spring and Google Guice. Users can use constructor injection when using these kinds of systems. Alternatively, users who prefer not to use DI frameworks will almost certainly prefer the builder-style helper classes that come with many of the critical classes. By convention, these builders are always inner static classes named Builder, contained within their constituent classes.

The primary actors in the Data module are entities, dataset repositories, datasets, dataset readers and writers, and metadata providers. Most of these objects are interfaces, permitting multiple implementations, each with different functionality. Today, there exists an implementation of each of these components for the Hadoop FileSystem abstraction, found in the com.cloudera.cdk.data.filesystem package. While, in theory, this means any implementation of Hadoop’s FileSystem abstract class is supported by the Data module, only the local and HDFS filesystem implementations are tested and officially supported. For the remainder of this guide, you can assume the implementation of the Data module interfaces being described is the Hadoop FileSystem implementation.

If you’re not already familiar with Avro schemas, now is a good time to go read a little more about them. You need not concern yourself with the details of how objects are serialized, but the ability to specify the schema to which entities of a dataset must conform is critical. The rest of this guide will assume you have a basic understanding of how to define a schema.

Dataset Repositories and Metadata Providers

A dataset repository is a physical storage location for datasets. In keeping with the relational database analogy, a dataset repository is the equivalent of a database of tables. Developers may organize datasets into different dataset repositories for reasons related to logical grouping, security and access control, backup policies, and so forth. A dataset repository is represented by instances of the com.cloudera.cdk.data.DatasetRepository interface in the Data module. An instance of a DatasetRepository acts as a factory for datasets, supplying methods for creating, loading, and dropping datasets. Each dataset belongs to exactly one dataset repository. There’s no built in support for moving or copying datasets bewteen repositories. MapReduce and other execution engines can easily provide copy functionality, if desired.

DatasetRepository Interface

Dataset create(String, DatasetDescriptor);
Dataset get(String);
boolean drop(String);

The Data module ships with a DatasetRepository implementation com.cloudera.cdk.data.filesystem.FileSystemDatasetRepository built for operating on datasets stored in a filesystem supported by Hadoop’s FileSystem abstraction. This implementation requires an instance of a Hadoop FileSystem, a root directory under which datasets will be (or are) stored, and an optional metadata provider (described later) to be supplied upon instantiation. Once complete, users can freely interact with datasets while the implementation worries about managing files and directories in the underlying filesystem. The supplied metadata provider is used to resolve dataset schemas and other related information.

Instantiating FileSystemDatasetRepository is straight forward.

DatasetRepository repo = new FileSystemDatasetRepository(
  FileSystem.get(new Configuration()),
  new Path("/data")
);

This example uses the currently configured Hadoop FileSystem implementation, typically an HDFS cluster. Since Hadoop’s also supports a “local” implementation of FileSystem, it’s possible to use the Data APIs to interact with data residing on a local OS filesystem. This is especially useful during development and basic functional testing of your code.

Once an instance of a DatasetRepository implementation has been created, new datasets can easily be created, and existing datasets can be loaded or dropped. Here’s a more complete example of creating a dataset to store application event data. You’ll notice a few new classes; don’t worry about them for now. We’ll cover them later.

/*
 * Instantiate a FileSystemDatasetRepository with a Hadoop FileSystem object
 * based on the current configuration. The FileSystem and Configuration
 * classes come from Hadoop. The Path object tells the dataset repository
 * under what path in the configured filesystem datasets reside.
 */
DatasetRepository repo = new FileSystemDatasetRepository(
  FileSystem.get(new Configuration()),
  new Path("/data")
);

/*
 * Create the dataset "users" with the schema defined in the file User.avsc.
 */
Dataset users = repo.create(
  "users",
  new DatasetDescriptor.Builder()
    .schema(new File("User.avsc"))
    .get()
);

Related to the dataset repository, the metadata provider is a service provider interface used to interact with the service that provides dataset metadata information to the rest of the Data APIs. The com.cloudera.cdk.data.MetadataProvider interface defines the contract metadata services must provide to the library, and specifically, the DatasetRepository.

MetadataProvider Interface

void save(String, DatasetDescriptor);
DatasetDescriptor load(String);
boolean delete(String);

MetadataProvider implementations act as a bridge between the Data module and centralized metadata repositories. An obvious example of this (in the Hadoop ecosystem) is HCatalog and the Hive metastore. By providing an implementation that makes the necessary API calls to HCatalog’s REST service, any and all datasets are immediately consumable by systems compatible with HCatalog, the storage system represented by the DatasetRepository implementation, and the format in which the data is written. As it turns out, that’s a pretty tall order and, in keeping with the CDK’s purpose of simplifying rather than presenting additional options, users are encouraged to 1. use HCatalog, 2. allow this library to default to snappy compressed Avro data files, and 3. use systems that also integrate with HCatalog (directly or indirectly). In this way, this library acts as a forth integration point to working with data in HDFS that is HCatalog-aware, in addition to Hive, Pig, and MapReduce input/output formats.

Users aren’t expected to use metadata providers directly. Instead, DatasetRepository implementations accept instances of MetadataProvider plugins, and make whatever calls are needed as users interact with the Data APIs. If you’re paying close attention, you’ll see that we didn’t specify a metadata provider when we instantiated FileSystemDatasetRepository earlier. That’s because FileSystemDatasetRepository uses an implementation of MetadataProvider called FileSystemMetadataProvider by default. Developers are free to explicitly pass a different implementation using the three argument constructor FileSystemDatasetRepository(FileSystem, Path, MetadataProvider) if they want to change this behavior.

The FileSystemMetadataProvider (also in the packge com.cloudera.cdk.data.filesystem) plugin stores dataset metadata information on a Hadoop FileSystem in a hidden directory. As with its sibling FileSystemDatasetRepository, its constructor accepts a Hadoop FileSystem object and a base directory. When metadata needs to be stored, a directory under the supplied base directory with the dataset name is created (if it doesn’t yet exist), and the dataset descriptor information is serialized to a set of files in a directory named .metadata.

Example: Explicitly configuring FileSystemDatasetRepository with FileSystemMetadataProvider

FileSystem fileSystem = FileSystem.get(new Configuration());
Path basePath = new Path("/data");

MetadataProvider metaProvider = new FileSystemMetadataProvider(
  fileSystem, basePath);

DatasetRepository repo = new FileSystemDatasetRepository(
  fileSystem, basePath, metaProvider);

Note how the same base directory of /data is used for both the metadata provider as well as the dataset repository. This is perfectly legal. In fact, this is exactly what happens when you use the two argument form of the FileSystemDatasetRepository constructor. Configured this way, data and metadata will be stored together, side by side, on whatever filesystem Hadoop is currently configured to use. Later, when we create a dataset, we’ll see the resultant file and directory structure created as a result of this configuration.

To use HCatalog, create an instance of HCatalogDatasetRepository (in the package com.cloudera.cdk.data.hcatalog). HCatalogDatasetRepository uses an internal implementation of MetadataProvider called HCatalogMetadataProvider to communicate with HCatalog. When using HCatalog you have two options for specifying the location of the data files. You can let HCatalog manage the location of the data, the so called “managed tables” option, in which case the data is stored in the warehouse directory that is configured by the Hive/HCatalog installation (see the hive.metastore.warehouse.dir setting in hive-site.xml). Alternatively, you can provide an explicit Hadoop FileSystem and root directory for datasets, just like FileSystemDatasetRepository. The latter option is referred to as “external tables” in the context of Hive/HCatalog.

Example: Creating a HCatalogDatasetRepository with managed tables

DatasetRepository repo = new HCatalogDatasetRepository();

Datasets

Summary

• A dataset is a collection of entities.
• A dataset is represented by the interface Dataset.
• The Hadoop FileSystem implementation of a dataset…
  • is stored as Snappy-compressed Avro data files by default.
  • may support pluggable formats such as Parquet in the future.
  • is made up of zero or more files in a directory.

A dataset is a collection of zero or more entities (or records). All datasets have a name and an associated dataset descriptor. The dataset descriptor, as the name implies, describes all aspects of the dataset. Primarily, the descriptor information is the dataset’s required schema and its optional partition strategy. A descriptor must be provided at the time a dataset is created. The schema is defined using the Avro Schema APIs. Entities must all conform to the same schema, however, that schema can evolve based on a set of well-defined rules. The relational database analog of a dataset is a table.

Datasets are represented by the com.cloudera.cdk.data.Dataset interface. Implementations of this interface decide how to physically store the entities within the dataset. Users do not instantiate implementations of the Dataset interface directly. Instead, implementations of the DatasetRepository act as a factory of the appropriate Dataset implementation.

Dataset Interface

String getName();
DatasetDescriptor getDescriptor();

<E> DatasetWriter<E> getWriter();
<E> DatasetReader<E> getReader();

Dataset getPartition(PartitionKey, boolean);
Iterable<Dataset> getPartitions();

The included Hadoop FileSystemDatasetRepository includes a Dataset implementation called FileSystemDataset. This dataset implementation stores data in the configured Hadoop FileSystem as Snappy-compressed Avro data files. Avro data files were selected because all components of CDH support them, they are language agnostic, support block compression, have a compact binary representation, and are natively splittable by Hadoop MapReduce while compressed. While it’s not currently possible to change this behavior, this will suit the needs of almost all users, and should be used whenever possible. Support for different serialization formats is planned for future releases for cases where users know something special about their data.

Upon creation of dataset, a name and a dataset descriptor must be provided to the DatasetRepository#create() method. The descriptor, represented by the com.cloudera.cdk.data.DatasetDescriptor class, holds all metadata associated with the dataset, the most important of which is the schema. Schemas are always represented using Avro’s Schema APIs, regardless of how the data is stored by the underlying dataset implementation. This simplifies the API for users, letting them focus on a single schema definition language for all datasets. In an effort to support different styles of schema definition, the DatasetDescriptor.Builder class supports a number of convenience methods for defining or attaching a schema.

DatasetDescriptor Class

org.apache.avro.Schema getSchema();
PartitionStrategy getPartitionStrategy();
boolean isPartitioned();

DatasetDescriptor.Builder Class

Builder schema(Schema schema);
Builder schema(String json);
Builder schema(File file);
Builder schema(InputStream inputStream);
Builder schema(URL url);

Builder partitionStrategy(PartitionStrategy partitionStrategy);

DatasetDescriptor get();

Note

Some of the less important or specialized methods have been elided here in the interest of simplicity.

From the methods in the DatasetDescriptor.Builder, you can see Avro schemas can be defined in a few different ways. Here, for instance, is an example of creating a dataset with a schema defined in a file on the local filesystem.

DatasetRepository repo = ...
Dataset users = repo.create("users",
  new DatasetDescriptor.Builder()
    .schema(new File("User.avsc"))
    .get()
);

Just as easily, a the schema could be loaded from a Java classpath resource. This example uses Guava’s Resources to simplify resolution, but we could have (almost as) easily used Java’s java.util.ClassLoader directly.

DatasetRepository repo = ...
Dataset users = repo.create("users",
  new DatasetDescriptor.Builder()
    .schema(Resources.getResource("Users.avsc"))
    .get()
);

An instance of Dataset acts as a factory for both reader and writer streams. Each implementation is free to produce stream implementations that make sense for the underlying storage system. The FileSystemDataset implementation, for example, produces streams that read from, or write to, Avro data files on a Hadoop FileSystem implementation.

Reader and writer streams both function similarly to Java’s standard IO streams, but are specialized. As indicated in the Dataset interface earlier, both interfaces are generic. The type parameter indicates the type of entity that they produce or consume, respectively.

DatasetReader Interface

void open();
void close();
boolean isOpen();

boolean hasNext();
E read();

DatasetWriter Interface

void open();
void close();
boolean isOpen();

void write(E);
void flush();

Both readers and writers are single-use objects with a well-defined lifecycle. Instances of both types (or the implementations of each, rather) must be opened prior to invoking any of the IO-generating methods such as DatasetReader’s hasNext() or read(), or DatasetWriter’s write() or flush(). Once a stream has been closed via the close() method, no further IO is permitted, nor may it be reopened.

Writing to a dataset always follows the same sequence of events. A user obtains an instance of a Dataset from a DatasetRepository either by creating a new, or loading an existing dataset. With a reference to a Dataset, you can obtain a writer using its getWriter() method, open it, write any number of entities, flush as necessary, and close it to release resources back to the system. The use of flush() and close() can dramatically affect data durability. Implementations of the DatasetWriter interface are free to define the semantics of data durability as appropriate for their storage subsystem. See the implementation javadoc on either the streams or the dataset for more information.

Example: Writing to a Hadoop FileSystem

DatasetRepository repo = new FileSystemDatasetRepository(
  FileSystem.get(new Configuration()), new Path("/data"));

/*
 * Let's assume MyInteger.avsc is defined as follows:
 * {
 *   "type": "record",
 *   "name": "MyInteger",
 *   "fields": [
 *     { "type": "integer", "name": "i" }
 *   ]
 * }
 */
Dataset integers = repo.create("integers",
  new DatasetDescriptor.Builder()
    .schema("MyInteger.avsc")
    .get()
);

/*
 * Getting a writer never performs IO, so it's safe to do this outside of
 * the try block. Here we're using Avro Generic records, discussed in
 * greater details later. See the Entities section.
 */
DatasetWriter<GenericRecord> writer = integers.getWriter();

try {
  writer.open();

  for (int i = 0; i < Integer.MAX_VALUE; i++) {
    writer.write(
      new GenericRecordBuilder(integers.getDescriptor().getSchema())
        .set("i", i)
        .build()
    );
  }
} finally {
  // Always explicitly close writers!
  writer.close();
}

Reading data from an existing dataset is equally straight forward. The important part to remember is that it is illegal to call read() after the reader has been exhausted (i.e. no more entities remain). Instead, users must use the hasNext() method to test if the reader can produce further data.

Example: Reading from a Hadoop FileSystem

DatasetRepository repo = new FileSystemDatasetRepository(
  FileSystem.get(new Configuration()), new Path("/data"));

// Load the integers dataset.
Dataset integers = repo.get("integers");

DatasetReader<GenericRecord> reader = integers.getReader();

try {
  reader.open();

  while (reader.hasNext()) {
    System.out.println("i: " + reader.read().get("i"));
  }
} finally {
  reader.close();
}

Dropping a dataset - an operation as equally destructive as dropping a table in a relational database - works as expected.

Example: Dropping an existing dataset

DatasetRepository repo = new FileSystemDatasetRepository(
  FileSystem.get(new Configuration()), new Path("/data"));

if (repo.drop("integers")) {
  System.out.println("Dropped dataset integers");
}

As discussed earlier, all operations performed on dataset repositories, datasets, and their associated readers and writers are tightly integrated with the dataset repository’s configured metadata provider. Dropping a dataset like this, for example, removes both the data as well as the associated metadata. All applications that use the Data module APIs will automatically see changes made by one another if they share the same configuration. This is an incredibly powerful concept allowing systems to become immediately aware of data as soon as it’s committed to storage.

Builder identity(String, int);
Builder hash(String, int);

PartitionStrategy get();

{
  "name": "User",
  "type": "record",
  "fields": [
    { "name": "id",           "type": "long"   },
    { "name": "username",     "type": "string" },
    { "name": "emailAddress", "type": "string" },
    { "name": "segment",      "type": "long"   }
  ]
}

DatasetRepository repo = ...

Dataset usersDataset = repo.create(
  "users",
  new DatasetDescriptor.Builder()
    .schema(new File("User.avsc"))
    .partitionStrategy(
      new PartitionStrategy.Builder().identity("segment", 1024).get()
    ).get()
);

id  username  emailAddress  segment
--  --------  ------------  -------
100 A         A@a.com       1
101 B         B@b.com       1
102 C         C@c.com       1
103 D         D@d.com       2
104 E         E@e.com       2

DatasetRepository repo = ...

Dataset users = repo.create(
  "users",
  new DatasetDescriptor.Builder()
    .schema(new File("User.avsc"))
    .partitionStrategy(
      new PartitionStrategy.Builder()
        .identity("segment", 1024)  // Partition first by segment
        .hash("emailAddress", 3)    // and then by hash(email) % 3
        .get()
    ).get()

Entities

Summary

• An entity is a record in a dataset.
• Entities can be POJOs, Avro GenericRecords, or Avro generated (specific) records.
• When in doubt, use GenericRecords.

An entity is a is a single record. The name “entity” is used rather than “record” because the latter caries a connotation of a simple list of primitives, while the former evokes the notion of a POJO (e.g. in JPA). That said, the terms are used interchangably. An entity can take one of three forms, at the user’s option:

1. A plain old Java object

When a POJO is supplied, the library uses reflection to write the object out to storage. While not the fastest, this is the simplest way to get up and running. Users are encouraged to consider Avro GenericRecords for production systems, or after they become familiar with the APIs.

1. An Avro GenericRecord

An Avro GenericRecord instance can be used to easily supply entities that represent a schema without using custom types for each kind of entity. These objects are easy to create and manipulate (see Avro’s GenericRecordBuilder class), especially in code that has no knowledge of specific object types (such as libraries). Serialization of generic records is fast, but requires use of the Avro APIs. This is recommended for most users, in most cases.

1. An Avro specific type

Advanced users may choose to use Avro’s code generation support to create classes that implicitly know how to serialize themselves. While the fastest of the options, this requires specialized knowledge of Avro, code generation, and handling of custom types. Keep in mind that, unlike generic records, the applications that write datasets with specific types must also have the same classes available to the applications that read those datasets.

Note that entities aren’t represented by any particular type in the Data APIs. In each of the above three cases, the entities described are either simple POJOs or are Avro objects. Remember that what has been described here is only the in memory representation of the entity; the Data module may store the data in HDFS in a different serialization format.

Entites may be complex types, representing data structures as simple as a few string attributes, or as complex as necessary. See Example: User entity schema and POJO class for an example of a valid Avro schema, and its associated POJO.

Example: User entity schema and POJO class

Avro schema (User.avsc)
-----------------------
{
  "name": "User",
  "type": "record",
  "fields": [
    // two required fields.
    { "name": "id",          "type": "long" },
    { "name": "username",    "type": "string" },

    // emailAddress is optional; it's value can be a string or a null.
    { "name": "emailAddress", "type": [ "string", "null" ] },

    // friendIds is an array with elements of type long.
    { "name": "friendIds",   "type": { "type": "array", "items": "long" } },
  ]
}

User POJO (User.java)
---------------------
public class User {

  private Long id;
  private String username;
  private String emailAddress;
  private List<Long> friendIds;

  public User() {
    friendIds = new ArrayList<Long>();
  }

  public Long getId() {
    return id;
  }

  public void setId(Long id) {
    this.id = id;
  }

  public String getUsername() {
    return username;
  }

  public void setUsername(String username) {
    this.username = username;
  }

  public String getEmailAddress() {
    return emailAddress;
  }

  public void setEmailAddress(String emailAddress) {
    this.emailAddress = emailAddress;
  }

  public List<Long> getFriendIds() {
    return friendIds;
  }

  public void setFriendIds(List<Long> friendIds) {
    this.friendIds = friendIds;
  }

  /*
   * It's fine to have methods that the schema doesn't know about. They'll
   * just be ignored during serialization.
   */
  public void addFriend(Friend friend) {
    if (!friendIds.contains(friend.getId()) {
      friendIds.add(friend.getId());
    }
  }

}

Instead of defining a POJO, we could also use Avro’s GenericRecordBuilder to create a generic entity that conforms to the User schema we defined earlier.

Example: Using Avro’s GenericRecordBuilder to create a generic entity

/*
 * Load the schema from User.avsc. Later, we'll an easier way to reference
 * Avro schemas when working with the CDK Data APIs.
 */
Schema userSchema = new Schema.Parser().parse(new File("User.avsc"));

/*
 * The GenericRecordBuilder constructs a new record and ensures that we set
 * all the necessary fields with values of an appropriate type.
 */
GenericRecord genericUser = new GenericRecordBuilder(userSchema)
  .set("id", 1L)
  .set("username", "janedoe")
  .set("emailAddress", "jane@doe.com")
  .set("friendIds", Collections.<Long>emptyList())
  .build();

Later, we’ll see how to read and write these entities to a dataset.

Appendix