Intermediate Data Format API

Associated  JIRA ticket : SQOOP-1350 and its sub tickets for more granular discussion points

**The following details are relevant post 1.99.4 release, i,e from 1.99.5 onwards.

Background

Intermediate Data Format (IDF)

• Connectors have FROM and TO parts. A sqoop job represents data transfer between FROM and TO. IDF API represents how the data is represented as it flows between the FROM and TO via sqoop.  
• Connectors basically represent different data sources and each data source can have its custom/ native format that it uses. For instance MongoDb might use JSON as its optimal native format, HDFS can use plain CSV text, S3 can use its own custom format. In simple words, every data source has one thing in common, it is collection of rows and each row is a collection of fields / columns. Most if not all data sources have strict schema that tells what each field type is. IDF encapsulated the native format and the schema associated with each field.
• Before we understand IDF API, it is important to be aware of the two other low level APIs that sqoop defines for data reading and writing between the FROM and TO data sources
        
        DataReader

/**
 * An intermediate layer for passing data from the execution engine
 * to the ETL engine.
 */
public abstract class DataReader {
  /**
   * Read data from the execution engine as an object array.
   * @return - array of objects with each column represented as an object
   * @throws Exception
   */
  public abstract Object[] readArrayRecord() throws Exception;
  /**
   * Read data from execution engine as text - as a CSV record.
   * public abstract Object readContent(int type) throws Exception;
   * @return - CSV formatted data.
   * @throws Exception
   */
  public abstract String readTextRecord() throws Exception;
  /**
   * Read data from execution engine as a native format.
   * @return - the content in the native format of the intermediate data
   * format being used.
   * @throws Exception
   */
  public abstract Object readContent() throws Exception;
}

               DataWriter

/**
 * An intermediate layer for passing data from the ETL framework
 * to the MR framework.
 */
public abstract class DataWriter {
  /**
   * Write an array of objects into the execution framework
   * @param array - data to be written
   */
  public abstract void writeArrayRecord(Object[] array);
  /**
   * Write data into execution framework as text. The Intermediate Data Format
   * may choose to convert the data to another format based on how the data
   * format is implemented
   * @param text - data represented as CSV text.
   */
  public abstract void writeStringRecord(String text);
  /**
   * Write data in the intermediate data format's native format.
   * @param obj - data to be written
   */
  public abstract void writeRecord(Object obj);
}

• IDF API is primarily influenced by the above low level apis to read and write data and hence this API dictates that each custom implementations support the following 3 formats

1. Native format - each row in the data source is a native object, for instance in JSONIDF, an entire row and its fields in sqoop will be represented as a JSON object, in AvroIDF, entire row and its fields will be represented as a Avro record
2. CSV text format - each row and its fields are represented as CSV text
3. Object Array format  - each field in the row is an element in the object array. Hence a row in the data source is represented as a object array

 /**
   * Get one row of data.
   *
   * @return - One row of data, represented in the internal/native format of
   *         the intermediate data format implementation.
   */
  public T getData() {
    return data;
  }
  /**
   * Set one row of data. If validate is set to true, the data is validated
   * against the schema.
   *
   * @param obj - A single row of data to be moved.
   */
  public void setData(T obj) {
    this.data = obj;
  }
  /**
   * Get one row of data as CSV text. Use {@link #SqoopIDFUtils} for reading and writing
   * into the sqoop specified CSV text format for each {@link #ColumnType} field in the row
   * Why a "native" internal format and then return CSV text too?
   * Imagine a connector that moves data from a system that stores data as a
   * serialization format called FooFormat. If I also need the data to be
   * written into HDFS as FooFormat, the additional cycles burnt in converting
   * the FooFormat to text and back is useless - so using the sqoop specified
   * CSV text format saves those extra cycles
   * <p/>
   * Most fast access mechanisms, like mysqldump or pgsqldump write the data
   * out as CSV, and most often the source data is also represented as CSV
   * - so having a minimal CSV support is mandated for all IDF, so we can easily read the
   * data out as text and write as text.
   * <p/>
   * @return - String representing the data in CSV text format.
   */
  public abstract String getCSVTextData();
  /**
   * Set one row of data as CSV.
   */
  public abstract void setCSVTextData(String csvText);
  /**
   * Get one row of data as an Object array.
   * Sqoop uses defined object representation
   * for each column type. For instance org.joda.time to represent date.
   * Use {@link #SqoopIDFUtils} for reading and writing into the sqoop
   * specified object format for each {@link #ColumnType} field in the row
   * </p>
   * @return - String representing the data as an Object array
   * If FROM and TO schema exist, we will use SchemaMatcher to get the data according to "TO" schema
   */
  public abstract Object[] getObjectData();
 /**
  * Set one row of data as an Object array.
  * It also should construct the data representation
  * that the IDF represents so that the object is ready to
  * consume when getData is invoked. Custom implementations
  * will override this method to convert form object array
  * to the data format
  */
  public abstract void setObjectData(Object[] data);
  /**
   * Set the schema for serializing/de-serializing data.
   *
   * @param schema
   *          - the schema used for serializing/de-serializing data
   */
  public void setSchema(Schema schema) {
    if (schema == null) {
      // TODO(SQOOP-1956): throw an exception since working without a schema is dangerous
      return;
    }
    this.schema = schema;
  ..
  }
  /**
   * Serialize the fields of this object to <code>out</code>.
   *
   * @param out <code>DataOuput</code> to serialize this object into.
   * @throws IOException
   */
  public abstract void write(DataOutput out) throws IOException;
  /**
   * Deserialize the fields of this object from <code>in</code>.
   *
   * <p>For efficiency, implementations should attempt to re-use storage in the
   * existing object where possible.</p>
   *
   * @param in <code>DataInput</code> to deseriablize this object from.
   * @throws IOException
   */
  public abstract void read(DataInput in) throws IOException;
  /**
   * Provide the external jars that the IDF depends on
   * @return set of jars
   */
  public Set<String> getJars() {
    return new HashSet<String>();
  }
  @Override
  public int hashCode() {
    final int prime = 31;
    int result = 1;
    result = prime * result + ((data == null) ? 0 : data.hashCode());
    result = prime * result + ((schema == null) ? 0 : schema.hashCode());
    return result;
  }
  

NOTE: The CSV Text format and the Object Array format are custom to Sqoop and the details of this format for every supported column/field type in the schema are described below.

Schema ( a.k.a Row ) 

Schema represents a row of fields that are transferred between FROM and TO. Hence schema holds the list of column/ field types for that row. 

/**
 * Schema represents the data fields that are transferred between {@link #From} and {@link #To}
 */
public class Schema {
  /**
   * Name of the schema, usually a table name.
   */
  private String name;
  /**
   * Optional note.
   */
  private String note;
  /**
   * Creation date.
   */
  private Date creationDate;
  /**
   * Columns associated with the schema.
   */
  private List<Column> columns;
  /**
   * Helper set for quick column name lookups.
   */
  private Set<String> columNames;

Column & ColumnType ( a.k.a Row Fields )

Column is an abstraction to represent a field in a row. There are custom classes for sub type such as String, Number, Date, Map, Array. It has attributes that provide metadata about the column data such as is that field nullable?, if that field is a String, what is its maxsize?, if it is DateTime, does it support timezone?, if it is a Map, what is the type of the key and what the is the type of the value?, if it is Array, what is the type of the elements?, if it is Enum, what are the supported options for the enum?

/**
 * Base class for all the supported types in the Sqoop {@link #Schema}
 */
public abstract class Column {
  /**
   * Name of the column. It is optional
   */
  String name;
  /**
   * Whether the column value can be empty/null
   */
  Boolean nullable;
  /**
   * By default a column is nullable
   */

ColumnType is an handy enum that represents all the field types Sqoop supports. Note there is a umbrella UNKNOWN type for fields that sqoop does not support.

/**
 * All {@link #Column} types supported by Sqoop.
 */
public enum ColumnType {
  ARRAY,
  BINARY,
  BIT,
  DATE,
  DATE_TIME,
  DECIMAL,
  ENUM,
  FIXED_POINT,
  FLOATING_POINT,
  MAP,
  SET,
  TEXT,
  TIME,
  UNKNOWN,
  ;
}

Design goals for IDF

There are a few prior documents that depict the design goals, but it is not crystal clear. Refer to this doc for some context on the research done prior to defining the IDF API. It explains some of the goals of using CSV and Object Array formats. Some of the design influence comes from the its predecessor Sqoop1.

1. Support data transfer across connectors using an "internal" in memory data representation

2. CSV is a common format in many databases and hence sqoop's design goals primarily want to optimize for such data sources. But it is unclear how much of performance gains does CSV text provide. 
   The following is a java doc comment I pulled from the code that explains the choice of CSV.

   Why a "native" internal format and then return CSV text too?
   Imagine a connector that moves data from a system that stores data as a
   serialization format called FooFormat. If I also need the data to be
   written into HDFS as FooFormat, the additional cycles burnt in converting
   the FooFormat to text and back is useless - so using the sqoop specified
   CSV text format saves those extra cycles
   <p/>
   Most fast access mechanisms, like mysqldump or pgsqldump write the data
   out as CSV, and most often the source data is also represented as CSV
   - so having a minimal CSV support is mandated for all IDF, so we can easily read the
   data out as text and write as text.

3. https://issues.apache.org/jira/browse/SQOOP-1811 has discussions that detail the current IDF API design goals 

1.99.5 SQOOP CSV and Object Format in IDF

ARRAY

BINARY

BIT

DATE

DATE_TIME

DECIMAL

ENUM

FIXED_POINT

FLOATING_POINT

MAP

SET

TEXT

TIME

UNKNOWN

Custom Implementations of IDF

CSVIntermediateDataFormat

Relevant JIRA : SQOOP-555 and SQOOP-1350

It is one of the sample implementation of the IDF API.

CSV IDF piggy backs on the the Sqoop CSV Format and its native format is the CSV Format represented as String. 

See the implementation class in the connector-sdk package for more details

NOTE: It may not be obvious but the current IDF design expect every new implementation of it to expose the CSV an ObjectArray formats in addition to its native format.

JSONIntermediateDataFormat

Relevant JIRA: SQOOP-1901

Avro Intermediate Data Format

SqoopIDFUtils 

It is a utility class in sqoop to aid connectors in encoding data into expected CSV format and object format and also parsing the CSV string back to the prescribed object format.

 https://issues.apache.org/jira/browse/SQOOP-1813

Food for Thought.?

(Some of the below are some serious shortcomings of the current design as it exists)

• The choice of using CSVText  as mandated formats for Sqoop IDF are influenced from the Sqoop 1 design, It favors some traditional fast dump databases but there is no real benchmark to prove how optimal it is vs using Avro or other formats for representing the data
• Using  intermediate format might lead to discrepancies in some specific column types, for instance using JODA for representing the date time objects only gives 3 digit precision, where as the sql timestamp from JDBC sources supports 6 digit precision
• More importantly SqoopConnector API has a getIDF..() method, that ties a connector to a  specific intermediate format for all the supported directions ( i.e both FROM and TO at this point) . This means the connector in both FROM and TO side has to provide this format and expect this format respectively. 
• There are 3 different formats as described above in each IDF implementation, so each connector can potentially support one of these formats and that is not obvious at all when a connector proclaims to use a particular implementation of IDF such as CSVIntermediateDataFormat. For instance the GenericJDBCConnector says it uses CSVIntermediateDataFormat but chooses to write objectArray in extractor and readObjectArray in Loader. Hence it is not obvious what is the format underneath that it will read and write to. On the other hand, HDFSConnector also says it uses CSVIntermediateDataFormat but, uses only the CSV text format in the extractor and loader at this point. May change in future. 

• A connector possibly should be able to handle multiple IDFs, and expose the supported IDFs per direction. It is not possible today, For instance a sqoop job should be able to dynamically choose the IDF for HDFSConnector when used in the TO direction. The job might be able to say, use AVRO IDF for the TO side and hence load all my data into HDFS in avro format. This means when doing the load, the HDFS will use the readContent API of the  SqoopOutputFormatDataReader. But today HDFS can only say it uses CSVIntermediateDataFormat and the data loaded into HDFS will need conversion from CSV to Avro as a separate step.