Using Impala to Query Kudu Tables

The primary key for a Kudu table is a column, or set of columns, that uniquely identifies every row. The primary key value also is used as the natural sort order for the values from the table. The primary key value for each row is based on the combination of values for the columns.

Because all of the primary key columns must have non-null values, specifying a column in the PRIMARY KEY clause implicitly adds the NOT NULL attribute to that column.

The primary key columns must be the first ones specified in the CREATE TABLE statement. For a single-column primary key, you can include a PRIMARY KEY attribute inline with the column definition. For a multi-column primary key, you include a PRIMARY KEY (c1, c2, ...) clause as a separate entry at the end of the column list.

You can specify the PRIMARY KEY attribute either inline in a single column definition, or as a separate clause at the end of the column list:

CREATE TABLE pk_inline
(
  col1 BIGINT PRIMARY KEY,
  col2 STRING,
  col3 BOOLEAN
) PARTITION BY HASH(col1) PARTITIONS 2 STORED AS KUDU;

CREATE TABLE pk_at_end
(
  col1 BIGINT,
  col2 STRING,
  col3 BOOLEAN,
  PRIMARY KEY (col1)
) PARTITION BY HASH(col1) PARTITIONS 2 STORED AS KUDU;

When the primary key is a single column, these two forms are equivalent. If the primary key consists of more than one column, you must specify the primary key using a separate entry in the column list:

CREATE TABLE pk_multiple_columns
(
  col1 BIGINT,
  col2 STRING,
  col3 BOOLEAN,
  PRIMARY KEY (col1, col2)
) PARTITION BY HASH(col2) PARTITIONS 2 STORED AS KUDU;

The SHOW CREATE TABLE statement always represents the PRIMARY KEY specification as a separate item in the column list:

CREATE TABLE inline_pk_rewritten (id BIGINT PRIMARY KEY, s STRING)
  PARTITION BY HASH(id) PARTITIONS 2 STORED AS KUDU;

SHOW CREATE TABLE inline_pk_rewritten;
+------------------------------------------------------------------------------+
| result                                                                       |
+------------------------------------------------------------------------------+
| CREATE TABLE user.inline_pk_rewritten (                                      |
|   id BIGINT NOT NULL ENCODING AUTO_ENCODING COMPRESSION DEFAULT_COMPRESSION, |
|   s STRING NULL ENCODING AUTO_ENCODING COMPRESSION DEFAULT_COMPRESSION,      |
|   PRIMARY KEY (id)                                                           |
| )                                                                            |
| PARTITION BY HASH (id) PARTITIONS 2                                          |
| STORED AS KUDU                                                               |
| TBLPROPERTIES ('kudu.master_addresses'='host.example.com')                   |
+------------------------------------------------------------------------------+

The notion of primary key only applies to Kudu tables. Every Kudu table requires a primary key. The primary key consists of one or more columns. You must specify any primary key columns first in the column list.

The contents of the primary key columns cannot be changed by an UPDATE or UPSERT statement. Including too many columns in the primary key (more than 5 or 6) can also reduce the performance of write operations. Therefore, pick the most selective and most frequently tested non-null columns for the primary key specification. If a column must always have a value, but that value might change later, leave it out of the primary key and use a NOT NULL clause for that column instead. If an existing row has an incorrect or outdated key column value, delete the old row and insert an entirely new row with the correct primary key.

For Kudu tables, you can specify which columns can contain nulls or not. This constraint offers an extra level of consistency enforcement for Kudu tables. If an application requires a field to always be specified, include a NOT NULL clause in the corresponding column definition, and Kudu prevents rows from being inserted with a NULL in that column.

For example, a table containing geographic information might require the latitude and longitude coordinates to always be specified. Other attributes might be allowed to be NULL. For example, a location might not have a designated place name, its altitude might be unimportant, and its population might be initially unknown, to be filled in later.

Because all of the primary key columns must have non-null values, specifying a column in the PRIMARY KEY clause implicitly adds the NOT NULL attribute to that column.

For non-Kudu tables, Impala allows any column to contain NULL values, because it is not practical to enforce a "not null" constraint on HDFS data files that could be prepared using external tools and ETL processes.

CREATE TABLE required_columns
(
  id BIGINT PRIMARY KEY,
  latitude DOUBLE NOT NULL,
  longitude DOUBLE NOT NULL,
  place_name STRING,
  altitude DOUBLE,
  population BIGINT
) PARTITION BY HASH(id) PARTITIONS 2 STORED AS KUDU;

During performance optimization, Kudu can use the knowledge that nulls are not allowed to skip certain checks on each input row, speeding up queries and join operations. Therefore, specify NOT NULL constraints when appropriate.

The NULL clause is the default condition for all columns that are not part of the primary key. You can omit it, or specify it to clarify that you have made a conscious design decision to allow nulls in a column.

Because primary key columns cannot contain any NULL values, the NOT NULL clause is not required for the primary key columns, but you might still specify it to make your code self-describing.

You can specify a default value for columns in Kudu tables. The default value can be any constant expression, for example, a combination of literal values, arithmetic and string operations. It cannot contain references to columns or non-deterministic function calls.

The following example shows different kinds of expressions for the DEFAULT clause. The requirement to use a constant value means that you can fill in a placeholder value such as NULL, empty string, 0, -1, 'N/A' and so on, but you cannot reference functions or column names. Therefore, you cannot use DEFAULT to do things such as automatically making an uppercase copy of a string value, storing Boolean values based on tests of other columns, or add or subtract one from another column representing a sequence number.

CREATE TABLE default_vals
(
  id BIGINT PRIMARY KEY,
  name STRING NOT NULL DEFAULT 'unknown',
  address STRING DEFAULT upper('no fixed address'),
  age INT DEFAULT -1,
  earthling BOOLEAN DEFAULT TRUE,
  planet_of_origin STRING DEFAULT 'Earth',
  optional_col STRING DEFAULT NULL
) PARTITION BY HASH(id) PARTITIONS 2 STORED AS KUDU;

Each column in a Kudu table can optionally use an encoding, a low-overhead form of compression that reduces the size on disk, then requires additional CPU cycles to reconstruct the original values during queries. Typically, highly compressible data benefits from the reduced I/O to read the data back from disk.

The following example shows the Impala keywords representing the encoding types. (The Impala keywords match the symbolic names used within Kudu.) For usage guidelines on the different kinds of encoding, see the Kudu documentation. The DESCRIBE output shows how the encoding is reported after the table is created, and that omitting the encoding (in this case, for the ID column) is the same as specifying DEFAULT_ENCODING.

CREATE TABLE various_encodings
(
  id BIGINT PRIMARY KEY,
  c1 BIGINT ENCODING PLAIN_ENCODING,
  c2 BIGINT ENCODING AUTO_ENCODING,
  c3 TINYINT ENCODING BIT_SHUFFLE,
  c4 DOUBLE ENCODING BIT_SHUFFLE,
  c5 BOOLEAN ENCODING RLE,
  c6 STRING ENCODING DICT_ENCODING,
  c7 STRING ENCODING PREFIX_ENCODING
) PARTITION BY HASH(id) PARTITIONS 2 STORED AS KUDU;

-- Some columns are omitted from the output for readability.
describe various_encodings;
+------+---------+-------------+----------+-----------------+
| name | type    | primary_key | nullable | encoding        |
+------+---------+-------------+----------+-----------------+
| id   | bigint  | true        | false    | AUTO_ENCODING   |
| c1   | bigint  | false       | true     | PLAIN_ENCODING  |
| c2   | bigint  | false       | true     | AUTO_ENCODING   |
| c3   | tinyint | false       | true     | BIT_SHUFFLE     |
| c4   | double  | false       | true     | BIT_SHUFFLE     |
| c5   | boolean | false       | true     | RLE             |
| c6   | string  | false       | true     | DICT_ENCODING   |
| c7   | string  | false       | true     | PREFIX_ENCODING |
+------+---------+-------------+----------+-----------------+

You can specify a compression algorithm to use for each column in a Kudu table. This attribute imposes more CPU overhead when retrieving the values than the ENCODING attribute does. Therefore, use it primarily for columns with long strings that do not benefit much from the less-expensive ENCODING attribute.

The choices for COMPRESSION are LZ4, SNAPPY, and ZLIB.

The following example shows design considerations for several STRING columns with different distribution characteristics, leading to choices for both the ENCODING and COMPRESSION attributes. The country values come from a specific set of strings, therefore this column is a good candidate for dictionary encoding. The post_id column contains an ascending sequence of integers, where several leading bits are likely to be all zeroes, therefore this column is a good candidate for bitshuffle encoding. The body column and the corresponding columns for translated versions tend to be long unique strings that are not practical to use with any of the encoding schemes, therefore they employ the COMPRESSION attribute instead. The ideal compression codec in each case would require some experimentation to determine how much space savings it provided and how much CPU overhead it added, based on real-world data.

CREATE TABLE blog_posts
(
  user_id STRING ENCODING DICT_ENCODING,
  post_id BIGINT ENCODING BIT_SHUFFLE,
  subject STRING ENCODING PLAIN_ENCODING,
  body STRING COMPRESSION LZ4,
  spanish_translation STRING COMPRESSION SNAPPY,
  esperanto_translation STRING COMPRESSION ZLIB,
  PRIMARY KEY (user_id, post_id)
) PARTITION BY HASH(user_id, post_id) PARTITIONS 2 STORED AS KUDU;

Although Kudu does not use HDFS files internally, and thus is not affected by the HDFS block size, it does have an underlying unit of I/O called the block size. The BLOCK_SIZE attribute lets you set the block size for any column.

Hash partitioning is the simplest type of partitioning for Kudu tables. For hash-partitioned Kudu tables, inserted rows are divided up between a fixed number of "buckets" by applying a hash function to the values of the columns specified in the HASH clause. Hashing ensures that rows with similar values are evenly distributed, instead of clumping together all in the same bucket. Spreading new rows across the buckets this way lets insertion operations work in parallel across multiple tablet servers. Separating the hashed values can impose additional overhead on queries, where queries with range-based predicates might have to read multiple tablets to retrieve all the relevant values.

-- 1M rows with 50 hash partitions = approximately 20,000 rows per partition.
-- The values in each partition are not sequential, but rather based on a hash function.
-- Rows 1, 99999, and 123456 might be in the same partition.
CREATE TABLE million_rows (id string primary key, s string)
  PARTITION BY HASH(id) PARTITIONS 50
  STORED AS KUDU;

-- Because the ID values are unique, we expect the rows to be roughly
-- evenly distributed between the buckets in the destination table.
INSERT INTO million_rows SELECT * FROM billion_rows ORDER BY id LIMIT 1e6;

Range partitioning lets you specify partitioning precisely, based on single values or ranges of values within one or more columns. You add one or more RANGE clauses to the CREATE TABLE statement, following the PARTITION BY clause.

Range-partitioned Kudu tables use one or more range clauses, which include a combination of constant expressions, VALUE or VALUES keywords, and comparison operators. (This syntax replaces the SPLIT ROWS clause used with early Kudu versions.) For the full syntax, see CREATE TABLE Statement.

-- 50 buckets, all for IDs beginning with a lowercase letter.
-- Having only a single range enforces the allowed range of values
-- but does not add any extra parallelism.
create table million_rows_one_range (id string primary key, s string)
  partition by hash(id) partitions 50,
  range (partition 'a' <= values < '{')
  stored as kudu;

-- 50 buckets for IDs beginning with a lowercase letter
-- plus 50 buckets for IDs beginning with an uppercase letter.
-- Total number of buckets = number in the PARTITIONS clause x number of ranges.
-- We are still enforcing constraints on the primary key values
-- allowed in the table, and the 2 ranges provide better parallelism
-- as rows are inserted or the table is scanned.
create table million_rows_two_ranges (id string primary key, s string)
  partition by hash(id) partitions 50,
  range (partition 'a' <= values < '{', partition 'A' <= values < '[')
  stored as kudu;

-- Same as previous table, with an extra range covering the single key value '00000'.
create table million_rows_three_ranges (id string primary key, s string)
  partition by hash(id) partitions 50,
  range (partition 'a' <= values < '{', partition 'A' <= values < '[', partition value = '00000')
  stored as kudu;

-- The range partitioning can be displayed with a SHOW command in impala-shell.
show range partitions million_rows_three_ranges;
+---------------------+
| RANGE (id)          |
+---------------------+
| VALUE = "00000"     |
| "A" <= VALUES < "[" |
| "a" <= VALUES < "{" |
+---------------------+

For range-partitioned Kudu tables, an appropriate range must exist before a data value can be created in the table. Any INSERT, UPDATE, or UPSERT statements fail if they try to create column values that fall outside the specified ranges. The error checking for ranges is performed on the Kudu side; Impala passes the specified range information to Kudu, and passes back any error or warning if the ranges are not valid. (A nonsensical range specification causes an error for a DDL statement, but only a warning for a DML statement.)

Ranges can be non-contiguous:

partition by range (year) (partition 1885 <= values <= 1889, partition 1893 <= values <= 1897)

partition by range (letter_grade) (partition value = 'A', partition value = 'B',
  partition value = 'C', partition value = 'D', partition value = 'F')

The ALTER TABLE statement with the ADD PARTITION or DROP PARTITION clauses can be used to add or remove ranges from an existing Kudu table.

ALTER TABLE foo ADD PARTITION 30 <= VALUES < 50;
ALTER TABLE foo DROP PARTITION 1 <= VALUES < 5;

When a range is added, the new range must not overlap with any of the previous ranges; that is, it can only fill in gaps within the previous ranges.

alter table test_scores add range partition value = 'E';

alter table year_ranges add range partition 1890 <= values < 1893;

When a range is removed, all the associated rows in the table are deleted. (This is true whether the table is internal or external.)

alter table test_scores drop range partition value = 'E';

alter table year_ranges drop range partition 1890 <= values < 1893;

Kudu tables can also use a combination of hash and range partitioning.

partition by hash (school) partitions 10,
  range (letter_grade) (partition value = 'A', partition value = 'B',
    partition value = 'C', partition value = 'D', partition value = 'F')

To see the current partitioning scheme for a Kudu table, you can use the SHOW CREATE TABLE statement or the SHOW PARTITIONS statement. The CREATE TABLE syntax displayed by this statement includes all the hash, range, or both clauses that reflect the original table structure plus any subsequent ALTER TABLE statements that changed the table structure.

To see the underlying buckets and partitions for a Kudu table, use the SHOW TABLE STATS or SHOW PARTITIONS statement.