Using Apache Parquet Data Files with CDH

For most CDH components, by default Parquet data files are not compressed. Cloudera recommends enabling compression to reduce disk usage and increase read and write performance.

You do not need to specify configuration to read a compressed Parquet file. However, to write a compressed Parquet file, you must specify the compression type. The supported compression types, the compression default, and how you specify compression depends on the CDH component writing the files.

To create a table named PARQUET_TABLE that uses the Parquet format, use a command like the following, substituting your own table name, column names, and data types:

hive> CREATE TABLE parquet_table_name (x INT, y STRING) STORED AS PARQUET;

If the table will be populated with data files generated outside of Impala and Hive, you can create the table as an external table pointing to the location where the files will be created:

hive> create external table parquet_table_name (x INT, y STRING)
  ROW FORMAT SERDE 'parquet.hive.serde.ParquetHiveSerDe'
  STORED AS
    INPUTFORMAT "parquet.hive.DeprecatedParquetInputFormat"
    OUTPUTFORMAT "parquet.hive.DeprecatedParquetOutputFormat"
    LOCATION '/test-warehouse/tinytable';

To populate the table with an INSERT statement, and to read the table with a SELECT statement, see Using the Parquet File Format with Impala Tables.

To set the compression type to use when writing data, configure the parquet.compression property:

set parquet.compression=GZIP;
INSERT OVERWRITE TABLE tinytable SELECT * FROM texttable;

The supported compression types are UNCOMPRESSED, GZIP, and SNAPPY.

Impala can create tables that use Parquet data files, insert data into those tables, convert the data into Parquet format, and query Parquet data files produced by Impala or other components. The only syntax required is the STORED AS PARQUET clause on the CREATE TABLE statement. After that, all SELECT, INSERT, and other statements recognize the Parquet format automatically. For example, a session in the impala-shell interpreter might look as follows:

[localhost:21000] > create table parquet_table (x int, y string) stored as parquet;
[localhost:21000] > insert into parquet_table select x, y from some_other_table;
Inserted 50000000 rows in 33.52s
[localhost:21000] > select y from parquet_table where x between 70 and 100;

Once you create a Parquet table this way in Impala, you can query it or insert into it through either Impala or Hive.

The Parquet format is optimized for working with large data files. In Impala 2.0 and higher, the default size of Parquet files written by Impala is 256 MB; in lower releases, 1 GB. Avoid using the INSERT ... VALUES syntax, or partitioning the table at too granular a level, if that would produce a large number of small files that cannot use Parquet optimizations for large data chunks.

Inserting data into a partitioned Impala table can be a memory-intensive operation, because each data file requires a memory buffer to hold the data before it is written. Such inserts can also exceed HDFS limits on simultaneous open files, because each node could potentially write to a separate data file for each partition, all at the same time. Make sure table and column statistics are in place for any table used as the source for an INSERT ... SELECT operation into a Parquet table. If capacity problems still occur, consider splitting insert operations into one INSERT statement per partition.

Impala can query Parquet files that use the PLAIN, PLAIN_DICTIONARY, BIT_PACKED, and RLE encodings. Currently, Impala does not support RLE_DICTIONARY encoding. When creating files outside of Impala for use by Impala, make sure to use one of the supported encodings. In particular, for MapReduce jobs, parquet.writer.version must not be defined (especially as PARQUET_2_0) for writing the configurations of Parquet MR jobs. Data using the version 2.0 of Parquet writer might not be consumable by Impala, due to use of the RLE_DICTIONARY encoding. Use the default version of the Parquet writer and refrain from overriding the default writer version by setting the parquet.writer.version property or via WriterVersion.PARQUET_2_0 in the Parquet API.

For complete instructions and examples, see Using the Parquet File Format with Impala Tables.

MapReduce requires Thrift in its CLASSPATH and in libjars to access Parquet files. It also requires parquet-format in libjars. Set up the following before running MapReduce jobs that access Parquet data files:

if [ -e /opt/cloudera/parcels/CDH ] ; then
    CDH_BASE=/opt/cloudera/parcels/CDH
else
    CDH_BASE=/usr
fi
THRIFTJAR=`ls -l $CDH_BASE/lib/hive/lib/libthrift*jar | awk '{print $9}' | head -1`
export HADOOP_CLASSPATH=$HADOOP_CLASSPATH:$THRIFTJAR
export LIBJARS=`echo "$CLASSPATH" | awk 'BEGIN { RS = ":" } { print }' | grep parquet-format | tail -1`
export LIBJARS=$LIBJARS,$THRIFTJAR

hadoop jar my-parquet-mr.jar -libjars $LIBJARS

In CDH 5.13 and later, Pig can directly write to Parquet tables using HCatalog, the table management and storage layer for Hadoop. Both partitioned and non-partitioned tables are supported. This process is described below.

Spark SQL supports loading and saving DataFrames from and to a variety of data sources and has native support for Parquet. For information about Parquet, see Using Apache Parquet Data Files with CDH.

To read Parquet files in Spark SQL, use the SQLContext.read.parquet("path") method.

To write Parquet files in Spark SQL, use the DataFrame.write.parquet("path") method.

sqlContext.setConf("spark.sql.parquet.compression.codec","codec") 

Currently, Spark looks up column data from Parquet files by using the names stored within the data files. This is different than the default Parquet lookup behavior of Impala and Hive. If data files are produced with a different physical layout due to added or reordered columns, Spark still decodes the column data correctly. If the logical layout of the table is changed in the metastore database, for example through an ALTER TABLE CHANGE statement that renames a column, Spark still looks for the data using the now-nonexistent column name and returns NULLs when it cannot locate the column values. To avoid behavior differences between Spark and Impala or Hive when modifying Parquet tables, avoid renaming columns, or use Impala, Hive, or a CREATE TABLE AS SELECT statement to produce a new table and new set of Parquet files containing embedded column names that match the new layout.

For an example of writing Parquet files to Amazon S3, see Examples of Accessing S3 Data from Spark. For a similar example for Microsoft ADLS, see Examples of Accessing ADLS Data from Spark.

For general information and examples of Spark working with data in different file formats, see Accessing External Storage from Spark.

Impala has always included Parquet support, using high-performance code written in C++ to read and write Parquet files. The Parquet JARs for use with Hive, Pig, and MapReduce are available with CDH 4.5 and higher. Using the Java-based Parquet implementation on a CDH release lower than CDH 4.5 is not supported.

A Parquet table created by Hive can typically be accessed by Impala 1.1.1 and higher with no changes, and vice versa. Before Impala 1.1.1, when Hive support for Parquet was not available, Impala wrote a dummy SerDe class name into each data file. These older Impala data files require a one-time ALTER TABLE statement to update the metadata for the SerDe class name before they can be used with Hive. See Apache Impala for details.

A Parquet file written by Hive, Impala, Pig, or MapReduce can be read by any of the others. Different defaults for file and block sizes, compression and encoding settings, and so on might cause performance differences depending on which component writes or reads the data files. For example, Impala typically sets the HDFS block size to 256 MB and divides the data files into 256 MB chunks, so that each I/O request reads an entire data file.

In CDH 5.5 and higher, non-Impala components that write Parquet files include extra padding to ensure that the Parquet row groups are aligned with HDFS data blocks. The maximum amount of padding is controlled by the parquet.writer.max-padding setting, specified as a number of bytes. By default, up to 8 MB of padding can be added to the end of each row group. This alignment helps prevent remote reads during Impala queries. The setting does not apply to Parquet files written by Impala, because Impala always writes each Parquet file as a single HDFS data block.

Each release may have limitations. The following are current limitations in CDH:

• Parquet has not been tested with HCatalog. Without HCatalog, Pig cannot correctly read dynamically partitioned tables; this is true for all file formats.
• Impala supports table columns using nested data types or complex data types such as map, struct, or array only in Impala 2.3 (corresponding to CDH 5.5) and higher. Impala 2.2 (corresponding to CDH 5.4) can query only the scalar columns of Parquet files containing such types. Lower releases of Impala cannot query any columns from Parquet data files that include such types.
• Cloudera supports some but not all of the object models from the upstream Parquet-MR project. Currently supported object models are:
  • parquet-avro (recommended for Cloudera users)
  • parquet-thrift
  • parquet-protobuf
  • parquet-pig
  • The Impala and Hive object models built into those components, not available in external libraries. (CDH does not include the parquet-hive module of the parquet-mr project, because recent versions of Hive have Parquet support built in.)

$ # Be careful doing this for a big file! Use parquet-tools head to be safe.
$ parquet-tools cat sample.parq
year = 1992
month = 1
day = 2
dayofweek = 4
dep_time = 748
crs_dep_time = 750
arr_time = 851
crs_arr_time = 846
carrier = US
flight_num = 53
actual_elapsed_time = 63
crs_elapsed_time = 56
arrdelay = 5
depdelay = -2
origin = CMH
dest = IND
distance = 182
cancelled = 0
diverted = 0

year = 1992
month = 1
day = 3
...

$ parquet-tools head -n 2 sample.parq
year = 1992
month = 1
day = 2
dayofweek = 4
dep_time = 748
crs_dep_time = 750
arr_time = 851
crs_arr_time = 846
carrier = US
flight_num = 53
actual_elapsed_time = 63
crs_elapsed_time = 56
arrdelay = 5
depdelay = -2
origin = CMH
dest = IND
distance = 182
cancelled = 0
diverted = 0

year = 1992
month = 1
day = 3
...

$ parquet-tools schema sample.parq
message schema {
  optional int32 year;
  optional int32 month;
  optional int32 day;
  optional int32 dayofweek;
  optional int32 dep_time;
  optional int32 crs_dep_time;
  optional int32 arr_time;
  optional int32 crs_arr_time;
  optional binary carrier;
  optional int32 flight_num;
...

$ parquet-tools meta sample.parq
creator:             impala version 2.2.0-cdh5.4.3 (build 517bb0f71cd604a00369254ac6d88394df83e0f6)

file schema:         schema
-------------------------------------------------------------------
year:                OPTIONAL INT32 R:0 D:1
month:               OPTIONAL INT32 R:0 D:1
day:                 OPTIONAL INT32 R:0 D:1
dayofweek:           OPTIONAL INT32 R:0 D:1
dep_time:            OPTIONAL INT32 R:0 D:1
crs_dep_time:        OPTIONAL INT32 R:0 D:1
arr_time:            OPTIONAL INT32 R:0 D:1
crs_arr_time:        OPTIONAL INT32 R:0 D:1
carrier:             OPTIONAL BINARY R:0 D:1
flight_num:          OPTIONAL INT32 R:0 D:1
...

row group 1:         RC:20636601 TS:265103674
-------------------------------------------------------------------
year:                 INT32 SNAPPY DO:4 FPO:35 SZ:10103/49723/4.92 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
month:                INT32 SNAPPY DO:10147 FPO:10210 SZ:11380/35732/3.14 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
day:                  INT32 SNAPPY DO:21572 FPO:21714 SZ:3071658/9868452/3.21 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
dayofweek:            INT32 SNAPPY DO:3093276 FPO:3093319 SZ:2274375/5941876/2.61 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
dep_time:             INT32 SNAPPY DO:5367705 FPO:5373967 SZ:28281281/28573175/1.01 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
crs_dep_time:         INT32 SNAPPY DO:33649039 FPO:33654262 SZ:10220839/11574964/1.13 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
arr_time:             INT32 SNAPPY DO:43869935 FPO:43876489 SZ:28562410/28797767/1.01 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
crs_arr_time:         INT32 SNAPPY DO:72432398 FPO:72438151 SZ:10908972/12164626/1.12 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
carrier:              BINARY SNAPPY DO:83341427 FPO:83341558 SZ:114916/128611/1.12 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
flight_num:           INT32 SNAPPY DO:83456393 FPO:83488603 SZ:10216514/11474301/1.12 VC:20636601 ENC:PLAIN_DICTIONARY,RLE,PLAIN
...

parquet-tools also provides the functionality to merge Parquet files, but Cloudera strongly discourages its use. Historically, good performance has been associated with large Parquet files. For this reason, many users attempted using parquet-tools merge to achieve good performance. In practice, however, parquet-tools merge doesn't improve Hive performance, and it degrades Impala performance significantly.

Good performance is not a result of large files but large row groups instead up to the HDFS block size. However, parquet-tools merge does not merge row groups, and it just places them one after the other. It was intended to be used for Parquet files that are already arranged in row groups of the desired size. When used to merge many small files, the resulting file will still contain small row groups and you lose the advantage of larger files.

The reason why you may want to combine small files in the first place is frequent ingestion. Because each ingestion will create small files, frequent ingestion will create lots of small files. Periodically, a compaction process is needed to combine these small files into bigger ones. At this moment, the only solution to this problem is rewriting the data. See the Cloudera blog post titled How-to: Ingest and Query “Fast Data” with Impala (Without Kudu) about the best practice for doing compaction while keeping the data continuously available at the same time.

You can find full examples of Java code at the Cloudera Parquet examples GitHub repository.

The TestReadWriteParquet.java example demonstrates the "identity" transform. It reads any Parquet data file and writes a new file with exactly the same content.

The TestReadParquet.java example reads a Parquet data file, and produces a new text file in CSV format with the same content.

The TestWriteParquet.java example demonstrates specifying a schema for writing the first two columns of a CSV input to Parquet output.