Kafka storage handler and table properties

Kafka storage handler

CREATE EXTERNAL TABLE kafka_table
   (`timestamp` timestamp , `page` string,  `newPage` boolean, 
   added int, deleted bigint, delta double)
   STORED BY 'org.apache.hadoop.hive.kafka.KafkaStorageHandler'
   TBLPROPERTIES
   ("kafka.topic" = "test-topic", "kafka.bootstrap.servers"="localhost:9092");

Storage handler-based optimizations

The storage handler can optimize reads using a filter push-down when you execute a query such as the following time-based lookup supported on Kafka 0.11 or later:

SELECT COUNT(*) FROM kafka_table 
   WHERE `__timestamp` >  1000 * to_unix_timestamp(CURRENT_TIMESTAMP - interval '10' MINUTES) ; 

The Kafka consumer supports seeking on the stream based on an offset, which the storage handler leverages to push down filters over metadata columns. The storage handler in the example above performs seeks based on the Kafka record __timestamp to read only recently arrived data.

The following logical operators and predicate operators are supported in the WHERE clause:

Logical operators: OR, AND

Predicate operators: <, <=, >=, >, =

The storage handler reader optimizes seeks by performing partition pruning to go directly to a particular partition offset used in the WHERE clause:

SELECT COUNT(*)  FROM kafka_table 
  WHERE (`__offset` < 10 AND `__offset` > 3 AND `__partition` = 0) 
  OR (`__partition` = 0 AND `__offset` < 105 AND `__offset` > 99) 
  OR (`__offset` = 109);

The storage handler scans partition 0 only, and then read only records between offset 4 and 109.

Kafka metadata

In addition to the user-defined payload schema, the Kafka storage handler appends to the table some additional columns, which you can use to query the Kafka metadata fields:

__key

Kafka record key (byte array)

__partition

Kafka record partition identifier (int 32)

__offset

Kafka record offset (int 64)

__timestamp

Kafka record timestamp (int 64)

The partition identifier, record offset, and record timestamp plus a key-value pair constitute a Kafka record. Because the key-value is a 2-byte array, you must use SerDe classes to transform the array into a set of columns.

Table Properties