Using the Parquet File Format with Impala Tables

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

[impala-host:21000] > create table parquet_table_name (x INT, y STRING) STORED AS PARQUET;

Or, to clone the column names and data types of an existing table:

[impala-host:21000] > create table parquet_table_name LIKE other_table_name STORED AS PARQUET;

In Impala 1.4.0 and higher, you can derive column definitions from a raw Parquet data file, even without an existing Impala table. For example, you can create an external table pointing to an HDFS directory, and base the column definitions on one of the files in that directory:

CREATE EXTERNAL TABLE ingest_existing_files LIKE PARQUET '/user/etl/destination/datafile1.dat'
  STORED AS PARQUET
  LOCATION '/user/etl/destination';

Or, you can refer to an existing data file and create a new empty table with suitable column definitions. Then you can use INSERT to create new data files or LOAD DATA to transfer existing data files into the new table.

CREATE TABLE columns_from_data_file LIKE PARQUET '/user/etl/destination/datafile1.dat'
  STORED AS PARQUET;

The default properties of the newly created table are the same as for any other CREATE TABLE statement. For example, the default file format is text; if you want the new table to use the Parquet file format, include the STORED AS PARQUET file also.

In this example, the new table is partitioned by year, month, and day. These partition key columns are not part of the data file, so you specify them in the CREATE TABLE statement:

CREATE TABLE columns_from_data_file LIKE PARQUET '/user/etl/destination/datafile1.dat'
  PARTITION (year INT, month TINYINT, day TINYINT)
  STORED AS PARQUET;

See CREATE TABLE Statement for more details about the CREATE TABLE LIKE PARQUET syntax.

Once you have created a table, to insert data into that table, use a command similar to the following, again with your own table names:

[impala-host:21000] > insert overwrite table parquet_table_name select * from other_table_name;

If the Parquet table has a different number of columns or different column names than the other table, specify the names of columns from the other table rather than * in the SELECT statement.

Choose from the following techniques for loading data into Parquet tables, depending on whether the original data is already in an Impala table, or exists as raw data files outside Impala.

If you already have data in an Impala or Hive table, perhaps in a different file format or partitioning scheme, you can transfer the data to a Parquet table using the Impala INSERT...SELECT syntax. You can convert, filter, repartition, and do other things to the data as part of this same INSERT statement. See Snappy and GZip Compression for Parquet Data Files for some examples showing how to insert data into Parquet tables.

Any INSERT statement for a Parquet table requires enough free space in the HDFS filesystem to write one block. Because Parquet data files use a block size of 1 GB by default, an INSERT might fail (even for a very small amount of data) if your HDFS is running low on space.

Avoid the INSERT...VALUES syntax for Parquet tables, because INSERT...VALUES produces a separate tiny data file for each INSERT...VALUES statement, and the strength of Parquet is in its handling of data (compressing, parallelizing, and so on) in large chunks.

If you have one or more Parquet data files produced outside of Impala, you can quickly make the data queryable through Impala by one of the following methods:

• The LOAD DATA statement moves a single data file or a directory full of data files into the data directory for an Impala table. It does no validation or conversion of the data. The original data files must be somewhere in HDFS, not the local filesystem.
• The CREATE TABLE statement with the LOCATION clause creates a table where the data continues to reside outside the Impala data directory. The original data files must be somewhere in HDFS, not the local filesystem. For extra safety, if the data is intended to be long-lived and reused by other applications, you can use the CREATE EXTERNAL TABLE syntax so that the data files are not deleted by an Impala DROP TABLE statement.
• If the Parquet table already exists, you can copy Parquet data files directly into it, then use the REFRESH statement to make Impala recognize the newly added data. Remember to preserve the block size of the Parquet data files by using the hadoop distcp -pb command rather than a -put or -cp operation on the Parquet files. See Example of Copying Parquet Data Files for an example of this kind of operation.

Recent versions of Sqoop can produce Parquet output files using the --as-parquetfile option.

If you use Sqoop to convert RDBMS data to Parquet, be careful with interpreting any resulting values from DATE, DATETIME, or TIMESTAMP columns. The underlying values are represented as the Parquet INT64 type, which is represented as BIGINT in the Impala table. The Parquet values represent the time in milliseconds, while Impala interprets BIGINT as the time in seconds. Therefore, if you have a BIGINT column in a Parquet table that was imported this way from Sqoop, divide the values by 1000 when interpreting as the TIMESTAMP type.

If the data exists outside Impala and is in some other format, combine both of the preceding techniques. First, use a LOAD DATA or CREATE EXTERNAL TABLE ... LOCATION statement to bring the data into an Impala table that uses the appropriate file format. Then, use an INSERT...SELECT statement to copy the data to the Parquet table, converting to Parquet format as part of the process.

Loading data into Parquet tables is a memory-intensive operation, because the incoming data is buffered until it reaches one data block in size, then that chunk of data is organized and compressed in memory before being written out. The memory consumption can be larger when inserting data into partitioned Parquet tables, because a separate data file is written for each combination of partition key column values, potentially requiring several large chunks to be manipulated in memory at once.

When inserting into a partitioned Parquet table, Impala redistributes the data among the nodes to reduce memory consumption. You might still need to temporarily increase the memory dedicated to Impala during the insert operation, or break up the load operation into several INSERT statements, or both.

Query performance for Parquet tables depends on the number of columns needed to process the SELECT list and WHERE clauses of the query, the way data is divided into large data files with block size equal to file size, the reduction in I/O by reading the data for each column in compressed format, which data files can be skipped (for partitioned tables), and the CPU overhead of decompressing the data for each column.

select avg(income) from census_data where state = 'CA';

select * from census_data;

Impala can optimize queries on Parquet tables, especially join queries, better when statistics are available for all the tables. Issue the COMPUTE STATS statement for each table after substantial amounts of data are loaded into or appended to it. See COMPUTE STATS Statement for details.

When Impala writes Parquet data files using the INSERT statement, the underlying compression is controlled by the COMPRESSION_CODEC query option. (Prior to Impala 2.0, the query option name was PARQUET_COMPRESSION_CODEC.) The allowed values for this query option are snappy (the default), gzip, and none. The option value is not case-sensitive. If the option is set to an unrecognized value, all kinds of queries will fail due to the invalid option setting, not just queries involving Parquet tables.

In CDH 5.5 / Impala 2.3 and higher, Impala supports the complex types ARRAY, STRUCT, and MAP See Complex Types (CDH 5.5 or higher only) for details. Because these data types are currently supported only for the Parquet file format, if you plan to use them, become familiar with the performance and storage aspects of Parquet first.

Starting in CDH 4.5, you can read and write Parquet data files from Hive, Pig, and MapReduce. See the CDH 4 Installation Guide for details.

Previously, it was not possible to create Parquet data through Impala and reuse that table within Hive. Now that Parquet support is available for Hive in CDH 4.5, reusing existing Impala Parquet data files in Hive requires updating the table metadata. Use the following command if you are already running Impala 1.1.1 or higher:

ALTER TABLE table_name SET FILEFORMAT PARQUET;

If you are running a level of Impala that is older than 1.1.1, do the metadata update through Hive:

ALTER TABLE table_name SET SERDE 'parquet.hive.serde.ParquetHiveSerDe';
ALTER TABLE table_name SET FILEFORMAT
  INPUTFORMAT "parquet.hive.DeprecatedParquetInputFormat"
  OUTPUTFORMAT "parquet.hive.DeprecatedParquetOutputFormat";

Impala 1.1.1 and higher can reuse Parquet data files created by Hive, without any action required.

Impala supports the scalar data types that you can encode in a Parquet data file, but not composite or nested types such as maps or arrays. In Impala 2.2.0 / CDH 5.4.0 and higher, Impala can query Parquet data files that include composite or nested types, as long as the query only refers to columns with scalar types.

If you copy Parquet data files between nodes, or even between different directories on the same node, make sure to preserve the block size by using the command hadoop distcp -pb. To verify that the block size was preserved, issue the command hdfs fsck -blocks HDFS_path_of_impala_table_dir and check that the average block size is at or near 256 MB (or whatever other size is defined by the PARQUET_FILE_SIZE query option).. (The hadoop distcp operation typically leaves some directories behind, with names matching _distcp_logs_*, that you can delete from the destination directory afterward.) Issue the command hadoop distcp for details about distcp command syntax.

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. Use the default version (or format). The default format, 1.0, includes some enhancements that are compatible with older versions. Data using the 2.0 format might not be consumable by Impala, due to use of the RLE_DICTIONARY encoding.

$ # 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
...

Although Parquet is a column-oriented file format, do not expect to find one data file for each column. Parquet keeps all the data for a row within the same data file, to ensure that the columns for a row are always available on the same node for processing. What Parquet does is to set a large HDFS block size and a matching maximum data file size, to ensure that I/O and network transfer requests apply to large batches of data.

Within that data file, the data for a set of rows is rearranged so that all the values from the first column are organized in one contiguous block, then all the values from the second column, and so on. Putting the values from the same column next to each other lets Impala use effective compression techniques on the values in that column.

When Impala retrieves or tests the data for a particular column, it opens all the data files, but only reads the portion of each file containing the values for that column. The column values are stored consecutively, minimizing the I/O required to process the values within a single column. If other columns are named in the SELECT list or WHERE clauses, the data for all columns in the same row is available within that same data file.

If an INSERT statement brings in less than one Parquet block's worth of data, the resulting data file is smaller than ideal. Thus, if you do split up an ETL job to use multiple INSERT statements, try to keep the volume of data for each INSERT statement to approximately 256 MB, or a multiple of 256 MB.

If you reuse existing table structures or ETL processes for Parquet tables, you might encounter a "many small files" situation, which is suboptimal for query efficiency. For example, statements like these might produce inefficiently organized data files:

-- In an N-node cluster, each node produces a data file
-- for the INSERT operation. If you have less than
-- N GB of data to copy, some files are likely to be
-- much smaller than the default Parquet block size.
insert into parquet_table select * from text_table;

-- Even if this operation involves an overall large amount of data,
-- when split up by year/month/day, each partition might only
-- receive a small amount of data. Then the data files for
-- the partition might be divided between the N nodes in the cluster.
-- A multi-gigabyte copy operation might produce files of only
-- a few MB each.
insert into partitioned_parquet_table partition (year, month, day)
  select year, month, day, url, referer, user_agent, http_code, response_time
  from web_stats;

Here are techniques to help you produce large data files in Parquet INSERT operations, and to compact existing too-small data files:

• When inserting into a partitioned Parquet table, use statically partitioned INSERT statements where the partition key values are specified as constant values. Ideally, use a separate INSERT statement for each partition.

• You might set the NUM_NODES option to 1 briefly, during INSERT or CREATE TABLE AS SELECT statements. Normally, those statements produce one or more data files per data node. If the write operation involves small amounts of data, a Parquet table, and/or a partitioned table, the default behavior could produce many small files when intuitively you might expect only a single output file. SET NUM_NODES=1 turns off the "distributed" aspect of the write operation, making it more likely to produce only one or a few data files.

• Be prepared to reduce the number of partition key columns from what you are used to with traditional analytic database systems.

• Do not expect Impala-written Parquet files to fill up the entire Parquet block size. Impala estimates on the conservative side when figuring out how much data to write to each Parquet file. Typically, the of uncompressed data in memory is substantially reduced on disk by the compression and encoding techniques in the Parquet file format. The final data file size varies depending on the compressibility of the data. Therefore, it is not an indication of a problem if 256 MB of text data is turned into 2 Parquet data files, each less than 256 MB.

• If you accidentally end up with a table with many small data files, consider using one or more of the preceding techniques and copying all the data into a new Parquet table, either through CREATE TABLE AS SELECT or INSERT ... SELECT statements.

  To avoid rewriting queries to change table names, you can adopt a convention of always running important queries against a view. Changing the view definition immediately switches any subsequent queries to use the new underlying tables:

  create view production_table as select * from table_with_many_small_files;
  -- CTAS or INSERT...SELECT all the data into a more efficient layout...
  alter view production_table as select * from table_with_few_big_files;
  select * from production_table where c1 = 100 and c2 < 50 and ...;
  

Schema evolution refers to using the statement ALTER TABLE ... REPLACE COLUMNS to change the names, data type, or number of columns in a table. You can perform schema evolution for Parquet tables as follows:

• The Impala ALTER TABLE statement never changes any data files in the tables. From the Impala side, schema evolution involves interpreting the same data files in terms of a new table definition. Some types of schema changes make sense and are represented correctly. Other types of changes cannot be represented in a sensible way, and produce special result values or conversion errors during queries.

• The INSERT statement always creates data using the latest table definition. You might end up with data files with different numbers of columns or internal data representations if you do a sequence of INSERT and ALTER TABLE ... REPLACE COLUMNS statements.

• If you use ALTER TABLE ... REPLACE COLUMNS to define additional columns at the end, when the original data files are used in a query, these final columns are considered to be all NULL values.

• If you use ALTER TABLE ... REPLACE COLUMNS to define fewer columns than before, when the original data files are used in a query, the unused columns still present in the data file are ignored.

• Parquet represents the TINYINT, SMALLINT, and INT types the same internally, all stored in 32-bit integers.

  • That means it is easy to promote a TINYINT column to SMALLINT or INT, or a SMALLINT column to INT. The numbers are represented exactly the same in the data file, and the columns being promoted would not contain any out-of-range values.

  • If you change any of these column types to a smaller type, any values that are out-of-range for the new type are returned incorrectly, typically as negative numbers.

  • You cannot change a TINYINT, SMALLINT, or INT column to BIGINT, or the other way around. Although the ALTER TABLE succeeds, any attempt to query those columns results in conversion errors.

  • Any other type conversion for columns produces a conversion error during queries. For example, INT to STRING, FLOAT to DOUBLE, TIMESTAMP to STRING, DECIMAL(9,0) to DECIMAL(5,2), and so on.

The Parquet format defines a set of data types whose names differ from the names of the corresponding Impala data types. If you are preparing Parquet files using other Hadoop components such as Pig or MapReduce, you might need to work with the type names defined by Parquet. The following figure lists the Parquet-defined types and the equivalent types in Impala.

Primitive types:

BINARY -> STRING
BOOLEAN -> BOOLEAN
DOUBLE -> DOUBLE
FLOAT -> FLOAT
INT32 -> INT
INT64 -> BIGINT
INT96 -> TIMESTAMP

Logical types:

BINARY + OriginalType UTF8 -> STRING
BINARY + OriginalType DECIMAL -> DECIMAL

Complex types:

For the complex types (ARRAY, MAP, and STRUCT) available in CDH 5.5 / Impala 2.3 and higher, Impala only supports queries against those types in Parquet tables.