gphdfs Support for Parquet Files (Deprecated)

    You can use the Greenplum Database gphdfs protocol to access Parquet files on a Hadoop file system (HDFS).

    Parent topic: Accessing HDFS Data with gphdfs (Deprecated)

    The Parquet file format is designed to take advantage of compressed, efficient columnar data representation available to projects in the Hadoop ecosystem. Parquet supports complex nested data structures and uses Dremel record shredding and assembly algorithms. Parquet supports very efficient compression and encoding schemes. Parquet allows compression schemes to be specified on a per-column level, and supports adding more encodings as they are invented and implemented.

    For information about the Parquet file format, see the Parquet documentation http://parquet.apache.org/documentation/latest/.

    For an overview of columnar data storage and the Parquet file format, see .

    The gphdfs protocol supports Parquet versions 1.7.0 and later. For each version, the required Parquet jar files are included in a bundled jar file parquet-hadoop-bundle-<version>.jar.

    Earlier Parquet versions not use the Java class names org.apache.parquet and are not supported. The gphdfs protocol expects the Parquet Java class names to be org.apache.parquet.xxx.

    Note: The Cloudera 5.4.x Hadoop distribution includes some Parquet jar files. However, the Java class names in the jar files are parquet.xxx. The jar files with the class name org.apache.parquet can be downloaded and installed on the Greenplum Database hosts.

    For information about downloading the Parquet jar files, see

    On all the Greenplum Database hosts, ensure that the jar files are installed and are on the classpath used by the gphdfs protocol. The classpath is specified by the shell script $GPHOME/lib/hadoop/hadoop_env.sh. As a Hadoop 2 example, you can install the jar files in $HADOOP_HOME/share/hadoop/common/lib. The hadoop_env.sh script file adds the jar files to the classpath.

    As an example, if the directory $HADOOP_HOME/share/hadoop/common/lib does not exist, create it on all Greenplum Database hosts as the gpadmin user. Then, add the add the jar files to the directory on all hosts.

    The hadoop_env.sh script file adds the jar files to classpath for the gphdfs protocol. This fragment in the script file adds the jar files to the classpath.

    The Greenplum Database gphdfs protocol supports the Parquet file format version 1 or 2. Parquet takes advantage of compressed, columnar data representation on HDFS. In a Parquet file, the metadata (Parquet schema definition) contains data structure information is written after the data to allow for single pass writing.

    This is an example of the Parquet schema definition format:

    1. message test {
    2.     repeated byte_array binary_field;
    3.     required int32 int32_field;
    4.     optional int64 int64_field;
    5.     required boolean boolean_field;
    6.     required fixed_len_byte_array(3) flba_field;
    7.     required byte_array someDay (utf8);
    8.     };

    To read from or write to a Parquet file, you create an external table and specify the location of the parquet file in the LOCATION clause and 'PARQUET' in the FORMAT clause. For example, this is the syntax for a readable external table.

    1. CREATE EXTERNAL TABLE <tablename> (<column_spec>) LOCATION ( 'gphdfs://<location>') FORMAT 'PARQUET'

    The location can be an Parquet file or a directory containing a set of Parquet files. For the file name you can specify the wildcard character * to match any number of characters. If the location specifies multiple files when reading Parquet files, Greenplum Database uses the schema in the first file that is read as the schema for the other files.

    The following table identifies how Greenplum database converts the Parquet data type if the Parquet schema definition does not contain an annotation.

    Note: When reading the Parquet int data type as Greenplum Database smallint data type, you must ensure that the Parquet int values do not exceed the Greenplum Database maximum smallint value. If the value is too large, the Greenplum Database value will be incorrect.

    The gphdfs protocol considers Parquet schema annotations for these cases. Otherwise, data conversion is based on the parquet schema primitive type:

    Note: See and the Parquet documentation when specifying decimal, date, interval, ortime* annotations.

    The gphdfs protocol converts the field data to text if the Parquet field type is binary without any annotation, and the data type is defined as text for the corresponding Greenplum Database external table column.

    When reading Parquet type group, the gphdfs protocol converts the group data into an XML document.

    This schema contains a required group with the name inner.

    1. message test {
    2.     required byte_array binary_field;
    3.     required int64 int64_field;
    4. **    required group inner \{
    6.        required boolean test;
    7.        required byte\_array name \(UTF8\);
    8.        \} **

    This how a single row of the group data would be converted to XML.

    This example schema contains a repeated group with the name inner.

    1. message test {
    2.     required byte_array binary_field;
    3.     required int64 int64_field;
    4. **    repeated group inner \{
    5.        int32 age;
    6.        required boolean test;
    7.        required byte\_array name \(UTF8\);
    8.        \} **
    9.     };

    For a repeated group, the Parquet file can contain multiple sets of the group data in a single row. For the example schema, the data for the inner group is converted into XML data.

    This is sample output if the data in the Parquet file contained two sets of data for the inner group.

    1. <inner type="repeated">
    2.   <inner type="group">
    3.     <age type="int">50</age>
    4.     <test type="boolean">true</test>
    5.     <name type="string">fred</name>
    6.   </inner>
    7.   <inner>
    8.     <age type="int">23</age>
    9.     <test type="boolean">false</test>
    10.     <name type="string">sam</name>
    11.   </inner>
    12. </inner>

    The Apache Hive data warehouse software can manage and query large datasets that reside in distributed storage. Apache Hive 0.13.0 and later can store data in Parquet format files. For information about Parquet used by Apache Hive, see .

    Notes on the Hive Generated Parquet Schema

    • When writing data to Parquet files, Hive treats all integer data types tinyint, smallint, int as int32. When you create an external table in Greenplum Database for a Hive generated Parquet file, specify the column data type as int. For example, this Hive CREATE TABLE command stores data in Parquet files.

      1. CREATE TABLE hpSimple(c1 tinyint, c2 smallint, c3 int, c4 bigint, 
      2.     c5 float, c6 double, c7 boolean, c8 string)
      3.   STORED AS PARQUET;

      This is the Hive generated Parquet schema for the hpSimple table data.

      The gphdfs protocol converts the Parquet integer data types to the Greenplum Database data type int.

    • For the Hive char data type, the Greenplum Database column data types can be either text or char. For the Hive varchar data type, the Greenplum Database column data type can be either text or varchar.

    • For the Hive timestamp data type, the Hive generated Parquet schema for the data type specifies that the data is stored as data type int96. The gphdfs protocol converts the int96 data type to the Greenplum Database bytea data type.

    For writable external tables, you can add parameters after the file specified in the location. You add parameters with the http query string syntax that starts with and & between field and value pairs.

    Note:

    1 Creates an internal dictionary. Enabling a dictionary can improve Parquet file compression if text columns contain similar or duplicate data.

    When writing a Parquet file, the gphdfs protocol can generate a Parquet schema based on the table definition.

    • The table name is used as the Parquet message name.
    • The column name is uses as the Parquet field name.

    When creating the Parquet schema from a Greenplum Database table definition, the schema is generated based on the column data type.

    Note: To support Null data, gphdfs protocol specifies the Parquet optional schema annotation when creating a Parquet schema.

    A simple example of a Greenplum Database table definition and the Parquet schema generated by the gphdfs protocol.

    1. CREATE WRITABLE EXTERNAL TABLE films (
    2.    code char(5), 
    3.    title varchar(40),
    4.    id integer,
    5.    date_prod date, 
    6.    subtitle boolean
    7. ) LOCATION ( 'gphdfs://my-films') FORMAT 'PARQUET' ;

    This is the Parquet schema for the Parquet file my-films generated by the gphdfs protocol.

    1. message films {
    2.     optional byte_array code;
    3.     optional byte_array title (utf8);
    4.     optional int32 id;
    5.     optional binary date_prod (utf8);
    6.     optional boolean subtitle;
    7.     };
    • For writable external tables, column definitions in Greenplum Database external table cannot specify NOT NULL to support automatically generating a Parquet schema. When the gphdfs protocol automatically generates a Parquet schema, the gphdfs protocol specifies the field attribute optional to support null in the Parquet schema. Repeated fields can be null in Parquet.
    • The gphdfs protocol supports Parquet nested group structures only for readable external files. The nested structures are converted to an XML document.
    • Greenplum Database does not have an unsigned int data type. Greenplum Database converts the Parquet unsigned int data type to the next largest Greenplum Database int type. For example, Parquet uint_8 is converted to Greenplum Database int (32 bit).
    • Greenplum Database supports any UDT data type or UDT array data type. Greenplum Database attempts to convert the UDT to a sting. If the UDT cannot be converted to a sting, Greenplum Database returns an error.
    • The Date data type in Parquet is starts from 1970.1.1, while Date in Greenplum Database starts from 4173 BC, Greenplum Database cannot convert date data types because largest values are different. A similar situation occurs between Timestamp_millis in Parquet and Timestamp in Greenplum Database.