Use the PXF HDFS connector to read and write Parquet-format data. This section describes how to read and write HDFS files that are stored in Parquet format, including how to create, query, and insert into external tables that reference files in the HDFS data store.

PXF supports reading or writing Parquet files compressed with these codecs: snappy, gzip, and lzo.

PXF currently supports reading and writing primitive Parquet data types only.

Prerequisites

Ensure that you have met the PXF Hadoop Prerequisites before you attempt to read data from or write data to HDFS.

Data Type Mapping

To read and write Parquet primitive data types in Greenplum Database, map Parquet data values to Greenplum Database columns of the same type.

Parquet supports a small set of primitive data types, and uses metadata annotations to extend the data types that it supports. These annotations specify how to interpret the primitive type. For example, Parquet stores both INTEGER and DATE types as the INT32 primitive type. An annotation identifies the original type as a DATE.

Read Mapping

PXF uses the following data type mapping when reading Parquet data:

Note: PXF supports filter predicate pushdown on all parquet data types listed above, except the fixed_len_byte_array and int96 types.

PXF can read a Parquet LIST nested type when it represents a one-dimensional array of certain Parquet types. The supported mappings follow:

Write Mapping

PXF uses the following data type mapping when writing Parquet data:

¹ Because Parquet does not save the field length, a Bpchar that PXF writes to Parquet will be a text of undefined length. ² PXF localizes a Timestamp to the current system time zone and converts it to universal time (UTC) before finally converting to int96. ³ PXF converts a Timestamptz to a UTC timestamp and then converts to int96. PXF loses the time zone information during this conversion.

PXF can write a one-dimensional LIST of certain Parquet data types. The supported mappings follow:

About Parquet Schemas and Data

Parquet is a columnar storage format. A Parquet data file contains a compact binary representation of the data. The schema defines the structure of the data, and is composed of the same primitive and complex types identified in the data type mapping section above.

A Parquet data file includes an embedded schema. You can choose to provide the schema that PXF uses to write the data to HDFS via the SCHEMA custom option in the CREATE WRITABLE EXTERNAL TABLE LOCATION clause (described below):

When you provide the Parquet schema file to PXF, you must specify the absolute path to the file, and the file must reside on the Hadoop file system.

Creating the External Table

The PXF HDFS connector hdfs:parquet profile supports reading and writing HDFS data in Parquet-format. When you insert records into a writable external table, the block(s) of data that you insert are written to one or more files in the directory that you specified.

Use the following syntax to create a Greenplum Database external table that references an HDFS directory:

CREATE [WRITABLE] EXTERNAL TABLE <table_name>
    ( <column_name> <data_type> [, ...] | LIKE <other_table> )
LOCATION ('pxf://<path-to-hdfs-dir>
    ?PROFILE=hdfs:parquet[&SERVER=<server_name>][&<custom-option>=<value>[...]]')
FORMAT 'CUSTOM' (FORMATTER='pxfwritable_import'|'pxfwritable_export')
[DISTRIBUTED BY (<column_name> [, ... ] ) | DISTRIBUTED RANDOMLY];

The specific keywords and values used in the Greenplum Database CREATE EXTERNAL TABLE command are described in the table below.

The PXF hdfs:parquet profile supports the following read option. You specify this option in the CREATE EXTERNAL TABLE LOCATION clause:

The PXF hdfs:parquet profile supports encoding- and compression-related write options. You specify these write options in the CREATE WRITABLE EXTERNAL TABLE LOCATION clause. The hdfs:parquet profile supports the following custom write options:

Note: You must explicitly specify uncompressed if you do not want PXF to compress the data.

Parquet files that you write to HDFS with PXF have the following naming format: <file>.<compress_extension>.parquet, for example 1547061635-0000004417_0.gz.parquet.

Example

This example utilizes the data schema introduced in Example: Reading Text Data on HDFS and adds a new column, item_quantity_per_order, an array with length equal to number_of_orders, that identifies the number of items in each order.

In this example, you create a Parquet-format writable external table that uses the default PXF server to reference Parquet-format data in HDFS, insert some data into the table, and then create a readable external table to read the data.

1. Use the hdfs:parquet profile to create a writable external table. For example:

   postgres=# CREATE WRITABLE EXTERNAL TABLE pxf_tbl_parquet (location text, month text, number_of_orders int, item_quantity_per_order int[], total_sales double precision)
       LOCATION ('pxf://data/pxf_examples/pxf_parquet?PROFILE=hdfs:parquet')
     FORMAT 'CUSTOM' (FORMATTER='pxfwritable_export');
   

2. Write a few records to the pxf_parquet HDFS directory by inserting directly into the pxf_tbl_parquet table. For example:

   postgres=# INSERT INTO pxf_tbl_parquet VALUES ( 'Frankfurt', 'Mar', 3, '{1,11,111}', 3956.98 );
   postgres=# INSERT INTO pxf_tbl_parquet VALUES ( 'Cleveland', 'Oct', 2, '{3333,7777}', 96645.37 );
   

3. Recall that Greenplum Database does not support directly querying a writable external table. To read the data in pxf_parquet, create a readable external Greenplum Database referencing this HDFS directory:

   postgres=# CREATE EXTERNAL TABLE read_pxf_parquet(location text, month text, number_of_orders int, item_quantity_per_order int[], total_sales double precision)
       LOCATION ('pxf://data/pxf_examples/pxf_parquet?PROFILE=hdfs:parquet')
       FORMAT 'CUSTOM' (FORMATTER='pxfwritable_import');
   

4. Query the readable external table read_pxf_parquet:

   postgres=# SELECT * FROM read_pxf_parquet ORDER BY total_sales;
   

    location  | month | number_of_orders | item_quantity_per_order | total_sales
   -----------+-------+------------------+-------------------------+-------------
    Frankfurt | Mar   |              777 | {1,11,111}              |     3956.98
    Cleveland | Oct   |             3812 | {3333,7777}             |    96645.4
   (2 rows)
   

Understanding Overflow Conditions When Writing Numeric Data

PXF uses the HiveDecimal class to write numeric Parquet data. HiveDecimal limits both the precision and the scale of a numeric type to a maximum of 38.

When you define a NUMERIC column in an external table without specifying a precision or scale, PXF internally maps the column to a DECIMAL(38, 18).

PXF handles the following precision overflow conditions:

• You define a NUMERIC column in the external table, and the integer digit count of a value exceeds the maximum supported precision of 38. For example, 1234567890123456789012345678901234567890.12345, which has an integer digit count of 45.
• You define a NUMERIC(<precision>) column with a <precision> greater than 38. For example, NUMERIC(55).
• You define a NUMERIC column in the external table, and the integer digit count of a value is greater than 20 (38-18). For example, 123456789012345678901234567890.12345, which has an integer digit count of 30.

If you define a NUMERIC(<precision>, <scale>) column and the integer digit count of a value is greater than <precision> - <scale>, PXF returns an error. For example, you define a NUMERIC(20,4) column and the value is 12345678901234567.12, which has an integer digit count of 19, which is greater than 20-4=16.

PXF can take one of three actions when it detects an overflow while writing numeric data to a Parquet file: round the value (the default), return an error, or ignore the overflow. The pxf.parquet.write.decimal.overflow property in the pxf-site.xml server configuration governs PXF’s action in this circumstance; valid values for this property follow:

PXF logs a warning when it detects an overflow and the pxf.parquet.write.decimal.overflow property is set to ignore.