Partitioning requirements

When you use ksqlDB to join streaming data, you must ensure that your streams and tables are co-partitioned, which means that input records on both sides of the join have the same configuration settings for partitions. The only exception is foreign-key table-table joins, which do not have any co-partitioning requirement.

To join two data sources, streams or tables, ksqlDB needs to compare their records based on the joining column. To ensure that records with the same join column are co-located on the same stream task, the join column must coincide with the column that the sources are partitioned by.

Partitioning streams and tables is especially important for stateful or otherwise intensive queries. For more information, see Parallelization.

Keys¶

Tables are always partitioned by their PRIMARY KEY, and ksqlDB doesn't allow repartitioning of tables, meaning you can only use a table's primary key as a join column.

Streams don't have primary keys, but they do have an optional KEY column. A KEY column, when present, defines the partitioning column.

Streams allow joins on expressions other than their key column. When the join criteria differ from the KEY column, ksqlDB internally repartitions the stream, which implicitly defines the correct key and partitioning.

Important

Kafka guarantees the relative order of any two messages from one source partition only if they are both in the same partition after the repartition. Otherwise, Kafka is likely to interleave messages. The use case will determine if these ordering guarantees are acceptable.

The following example shows a users table joined with a clicks stream on the click's userId column. The users table has a correct primary key, id, of the same SQL type. The clicks stream doesn't have a defined key, so ksqlDB internally repartitions it on the joining column (userId) to assign the key before performing the join.

-- clicks stream, with no or unknown key.
-- the schema of stream clicks is: USERID BIGINT | URL STRING
CREATE STREAM clicks (
    userId BIGINT, 
    url STRING
  ) WITH (
    kafka_topic='clickstream', 
    value_format='json',
    partitions=1
  );

-- users table, with userId primary key. 
-- the schema of table users is: USERID BIGINT PRIMARY KEY | FULLNAME STRING
CREATE TABLE users (
    id BIGINT PRIMARY KEY, 
    fullName STRING
  ) WITH (
    kafka_topic='users', 
    value_format='json',
    partitions=1
);

-- join of users table with clicks stream, joining on the table's primary key and the stream's userId column:
-- join will automatically repartition clicks stream:
SELECT 
  c.userId,
  c.url, 
  u.fullName 
FROM clicks c
  JOIN users u ON c.userId = u.id
EMIT CHANGES;

Co-partitioning Requirements¶

When you use ksqlDB to join streaming data, you must ensure that your streams and tables are co-partitioned (except for foreign-key table-table joins), which means that input records on both sides of the join have the same configuration settings for partitions.

The input records for the join must have the same key schema.
The input records must have the same number of partitions on both sides.
Both sides of the join must have the same partitioning strategy.

When your inputs are co-partitioned, records with the same key, from both sides of the join, are delivered to the same stream task during processing.

Records Have the Same Key Schema¶

For a join to work, the keys from both sides must have the same SQL type.

For example, you can join a stream of user clicks that's keyed on a STRING user id with a table of user profiles that's also keyed on a STRING user id. Records with the exact same user id on both sides will be joined.

If the schema of the columns you wish to join on don't match, it may be possible to CAST one side to match the other. For example, if one side of the join had a INT userId column, and the other a LONG, then you may choose to cast the INT side to a LONG:

-- stream with INT userId
CREATE STREAM clicks (
    userId INT KEY, 
    url STRING
  ) WITH (
    kafka_topic='clickstream', 
    value_format='json'
  );

-- table with BIGINT id stored in the key:
CREATE TABLE users (
    id BIGINT PRIMARY KEY, 
    fullName STRING
  ) WITH (
    kafka_topic='users', 
    value_format='json'
  );

-- Join utilising a CAST to convert the left sides join column to match the rights type.
SELECT 
  clicks.url, 
  users.fullName 
FROM clicks 
  JOIN users ON CAST(clicks.userId AS BIGINT) = users.id
EMIT CHANGES;

Tables created on top of existing Kafka topics, for example those created with a CREATE TABLE statement, are keyed on the data held in the key of the records in the Kafka topic. ksqlDB presents this data in the PRIMARY KEY column.

Tables created inside ksqlDB from other sources, for example those created with a CREATE TABLE AS SELECT statement, will copy the key from their source(s) unless there is an explicit GROUP BY or JOIN clause, which can change what the table is keyed on.

Note

ksqlDB automatically repartitions a stream if a join requires it, but for stream-table and table-table joins, ksqlDB rejects a join on a (right) table's column that is not the primary key. This is because repartitioning a table's topic has the potential to reorder events and misinterpret tombstones, which can lead to unintended or undesired side effects.

If you are using the same sources in more than one join that requires the data to be repartitioned, you may prefer to repartition manually to avoid ksqlDB repartitioning multiple times.

To repartition a stream, use the PARTITION BY clause. Be aware that Kafka guarantees the relative order of any two messages from one source partition only if they are also both in the same partition after the repartition. Otherwise, Kafka is likely to interleave messages. The use case will determine if these ordering guarantees are acceptable.

The PARTITION BY clause moves the columns into the key. If you want them in the value also, you must copy them by using the AS_VALUE function.

Important

If the PARTITION BY expression evaluates to NULL, the resulting row is produced to a random partition. You may want to use COALESCE to wrap the expression and convert any NULL values to a default value, for example, PARTITION BY COALESCE(MY_UDF_THAT_MAY_FAIL(Col0), 0).

For example, if you need to re-partition a stream to be keyed by a product_id field, and keys need to be distributed over 6 partitions to make a join work, use the following SQL statement:

CREATE STREAM products_rekeyed 
  WITH (PARTITIONS=6) AS 
  SELECT * 
   FROM products
   PARTITION BY product_id;

For more information, see How to rekey a stream with a value.

Records Have the Same Number of Partitions¶

The input records for joins must have the same number of partitions on both sides, except for foreign-key table-table joins.

ksqlDB checks this part of the co-partitioning requirement and rejects any join where the partition counts differ.

Use the DESCRIBE <source name> EXTENDED command in the CLI to determine the Kafka topic under a source, and use the SHOW TOPICS command in the CLI to list topics and their partition counts.

If the sides of the join have different partition counts, you may want to change the partition counts of the source topics, or repartition one side to match the partition count of the other.

The following example creates a repartitioned stream, maintaining the existing key, with the specified number of partitions.

CREATE STREAM products_rekeyed 
  WITH (PARTITIONS=6) AS 
  SELECT * FROM products;

Records Have the Same Partitioning Strategy¶

Records on both sides of the join must have the same partitioning strategy. If you use the default partitioner settings across all applications, and your producers don't specify an explicit partition, you don't need to worry about the partitioning strategy.

But if the producer applications for your records have custom partitioners specified in configuration, the same custom partitioner logic must be used for records on both sides of the join. The applications that write to the join inputs must have the same partitioning strategy, so that records with the same key are delivered to the same partition.

This means that the input records must be in the same partition on both sides of the join. For example, in a stream-table join, if a userId key with the value alice123 is in Partition 1 for the stream, but alice123 is in Partition 2 for the table, the join won't match, even though both sides are keyed by userId.

ksqlDB can't verify whether the partitioning strategies are the same for both join inputs, so you must ensure this.

The DefaultPartitioner class implements the following partitioning strategy:

If the producer specifies a partition in the record, use it.
If the producer specifies a key instead of a partition, choose a partition based on a hash of the key.
If the producer doesn't specify a partition or a key, choose a partition in a round-robin fashion.

Custom partitioner classes implement the Partitioner interface and are assigned in the producer configuration property, partitioner.class.

For example implementations of a custom partitioner, see Built for realtime: Big data messaging with Apache Kafka, Part2 and Apache Kafka Foundation Course - Custom Partitioner.