Stream join example with Apache Kafka?

Question

I was looking for an example using Kafka Streams on how to do this sort of thing, ie join a customers table with a addresses table and sink the data to ES:-

Customers

+------+------------+----------------+-----------------------+
| id   | first_name | last_name      | email                 |
+------+------------+----------------+-----------------------+
| 1001 | Sally      | Thomas         | sally.thomas@acme.com |
| 1002 | George     | Bailey         | gbailey@foobar.com    |
| 1003 | Edward     | Davidson       | ed@walker.com         |
| 1004 | Anne       | Kim            | annek@noanswer.org    |
+------+------------+----------------+-----------------------+

Addresses

+----+-------------+---------------------------+------------+--------------+-------+----------+
| id | customer_id | street                    | city       | state        | zip   | type     |
+----+-------------+---------------------------+------------+--------------+-------+----------+
| 10 |        1001 | 3183 Moore Avenue         | Euless     | Texas        | 76036 | SHIPPING |
| 11 |        1001 | 2389 Hidden Valley Road   | Harrisburg | Pennsylvania | 17116 | BILLING  |
| 12 |        1002 | 281 Riverside Drive       | Augusta    | Georgia      | 30901 | BILLING  |
| 13 |        1003 | 3787 Brownton Road        | Columbus   | Mississippi  | 39701 | SHIPPING |
| 14 |        1003 | 2458 Lost Creek Road      | Bethlehem  | Pennsylvania | 18018 | SHIPPING |
| 15 |        1003 | 4800 Simpson Square       | Hillsdale  | Oklahoma     | 73743 | BILLING  |
| 16 |        1004 | 1289 University Hill Road | Canehill   | Arkansas     | 72717 | LIVING   |
+----+-------------+---------------------------+------------+--------------+-------+----------+

Output Elasticsearch index

"hits": [
  {
    "_index": "customers_with_addresses",
    "_type": "_doc",
    "_id": "1",
    "_score": 1.3278645,
    "_source": {
      "first_name": "Sally",
      "last_name": "Thomas",
      "email": "sally.thomas@acme.com",
      "addresses": [{
        "street": "3183 Moore Avenue",
        "city": "Euless",
        "state": "Texas",
        "zip": "76036",
        "type": "SHIPPING"
      }, {
        "street": "2389 Hidden Valley Road",
        "city": "Harrisburg",
        "state": "Pennsylvania",
        "zip": "17116",
        "type": "BILLING"
      }],
    }
  }, ….

Table data is coming from Debezium topics, am I correct in thinking I need some Java in the middle to join the streams, output it to a new topic which then sinks that into ES?

Would anyone have any example code of this?

Thanks.

Answer 1

Depending on how strict your requirement is to nest multiple addresses in one customer node, you can do this in KSQL (which is built on top of Kafka Streams).

Populate some test data into Kafka (which in your case is done already through Debezium):

$ curl -s "https://api.mockaroo.com/api/ffa9ff20?count=10&key=ff7856d0" | kafkacat -b localhost:9092 -t addresses -P

$ curl -s "https://api.mockaroo.com/api/9b868890?count=4&key=ff7856d0" | kafkacat -b localhost:9092 -t customers -P

Fire up KSQL and to start with just inspect the data:

ksql> PRINT 'addresses' FROM BEGINNING ;
Format:JSON
{"ROWTIME":1558519823351,"ROWKEY":"null","id":1,"customer_id":1004,"street":"8 Moulton Center","city":"Bronx","state":"New York","zip":"10474","type":"BILLING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":2,"customer_id":1001,"street":"5 Hollow Ridge Alley","city":"Washington","state":"District of Columbia","zip":"20016","type":"LIVING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":3,"customer_id":1000,"street":"58 Maryland Point","city":"Greensboro","state":"North Carolina","zip":"27404","type":"LIVING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":4,"customer_id":1002,"street":"55795 Derek Avenue","city":"Temple","state":"Texas","zip":"76505","type":"LIVING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":5,"customer_id":1002,"street":"164 Continental Plaza","city":"Modesto","state":"California","zip":"95354","type":"SHIPPING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":6,"customer_id":1004,"street":"6 Miller Road","city":"Louisville","state":"Kentucky","zip":"40205","type":"BILLING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":7,"customer_id":1003,"street":"97 Shasta Place","city":"Pittsburgh","state":"Pennsylvania","zip":"15286","type":"BILLING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":8,"customer_id":1000,"street":"36 Warbler Circle","city":"Memphis","state":"Tennessee","zip":"38109","type":"SHIPPING"}
{"ROWTIME":1558519823351,"ROWKEY":"null","id":9,"customer_id":1001,"street":"890 Eagan Circle","city":"Saint Paul","state":"Minnesota","zip":"55103","type":"SHIPPING"}
{"ROWTIME":1558519823354,"ROWKEY":"null","id":10,"customer_id":1000,"street":"8 Judy Terrace","city":"Washington","state":"District of Columbia","zip":"20456","type":"SHIPPING"}
^C
Topic printing ceased

ksql>
ksql> PRINT 'customers' FROM BEGINNING;
Format:JSON
{"ROWTIME":1558519852363,"ROWKEY":"null","id":1001,"first_name":"Jolee","last_name":"Handasyde","email":"jhandasyde0@nhs.uk"}
{"ROWTIME":1558519852363,"ROWKEY":"null","id":1002,"first_name":"Rebeca","last_name":"Kerrod","email":"rkerrod1@sourceforge.net"}
{"ROWTIME":1558519852363,"ROWKEY":"null","id":1003,"first_name":"Bobette","last_name":"Brumble","email":"bbrumble2@cdc.gov"}
{"ROWTIME":1558519852368,"ROWKEY":"null","id":1004,"first_name":"Royal","last_name":"De Biaggi","email":"rdebiaggi3@opera.com"}

Now we declare a STREAM (Kafka topic + schema) on the data so that we can manipulate it further:

ksql> CREATE STREAM addresses_RAW (ID INT, CUSTOMER_ID INT, STREET VARCHAR, CITY VARCHAR, STATE VARCHAR, ZIP VARCHAR, TYPE VARCHAR) WITH (KAFKA_TOPIC='addresses', VALUE_FORMAT='JSON');

 Message
----------------
 Stream created
----------------

ksql> CREATE STREAM customers_RAW (ID INT, FIRST_NAME VARCHAR, LAST_NAME VARCHAR, EMAIL VARCHAR) WITH (KAFKA_TOPIC='customers', VALUE_FORMAT='JSON');

 Message
----------------
 Stream created
----------------

We're going to model the customers as a TABLE , and to do that the Kafka messages need to be keyed correctly (and the moment they have null keys, as can be seen from the "ROWKEY":"null" in the PRINT output above). You can configure Debezium to set the message key so this step may not be necessary for you in KSQL:

ksql> CREATE STREAM CUSTOMERS_KEYED WITH (PARTITIONS=1) AS SELECT * FROM CUSTOMERS_RAW PARTITION BY ID;

 Message
----------------------------
 Stream created and running
----------------------------

Now we declare a TABLE ( state for a given key, instantiated from a Kafka topic + schema):

ksql> CREATE TABLE CUSTOMER (ID INT, FIRST_NAME VARCHAR, LAST_NAME VARCHAR, EMAIL VARCHAR) WITH (KAFKA_TOPIC='CUSTOMERS_KEYED', VALUE_FORMAT='JSON', KEY='ID');

 Message
---------------
 Table created
---------------

Now we can join the data:

ksql> CREATE STREAM customers_with_addresses AS 
      SELECT CUSTOMER_ID, 
             FIRST_NAME + ' ' + LAST_NAME AS FULL_NAME, 
             FIRST_NAME, 
             LAST_NAME, 
             TYPE AS ADDRESS_TYPE, 
             STREET, 
             CITY, 
             STATE, 
             ZIP 
        FROM ADDRESSES_RAW A 
             INNER JOIN CUSTOMER C 
             ON A.CUSTOMER_ID = C.ID;

 Message
----------------------------
 Stream created and running
----------------------------

This creates a new KSQL STREAM which in turn populates a new Kafka topic.

ksql> SHOW STREAMS;

 Stream Name                              | Kafka Topic                          | Format
------------------------------------------------------------------------------------------
 CUSTOMERS_KEYED                          | CUSTOMERS_KEYED                      | JSON
 ADDRESSES_RAW                            | addresses                            | JSON
 CUSTOMERS_RAW                            | customers                            | JSON
 CUSTOMERS_WITH_ADDRESSES                 | CUSTOMERS_WITH_ADDRESSES             | JSON

The stream has a schema:

ksql> DESCRIBE CUSTOMERS_WITH_ADDRESSES;

Name                 : CUSTOMERS_WITH_ADDRESSES
 Field        | Type
------------------------------------------
 ROWTIME      | BIGINT           (system)
 ROWKEY       | VARCHAR(STRING)  (system)
 CUSTOMER_ID  | INTEGER          (key)
 FULL_NAME    | VARCHAR(STRING)
 FIRST_NAME   | VARCHAR(STRING)
 ADDRESS_TYPE | VARCHAR(STRING)
 LAST_NAME    | VARCHAR(STRING)
 STREET       | VARCHAR(STRING)
 CITY         | VARCHAR(STRING)
 STATE        | VARCHAR(STRING)
 ZIP          | VARCHAR(STRING)
------------------------------------------
For runtime statistics and query details run: DESCRIBE EXTENDED <Stream,Table>;

We can query the stream:

ksql> SELECT * FROM CUSTOMERS_WITH_ADDRESSES WHERE CUSTOMER_ID=1002;
1558519823351 | 1002 | 1002 | Rebeca Kerrod | Rebeca | LIVING | Kerrod | 55795 Derek Avenue | Temple | Texas | 76505
1558519823351 | 1002 | 1002 | Rebeca Kerrod | Rebeca | SHIPPING | Kerrod | 164 Continental Plaza | Modesto | California | 95354

We can also stream it to Elasticsearch using Kafka Connect:

curl -i -X POST -H "Accept:application/json" \
    -H  "Content-Type:application/json" http://localhost:8083/connectors/ \
    -d '{
      "name": "sink-elastic-customers_with_addresses-00",
      "config": {
        "connector.class": "io.confluent.connect.elasticsearch.ElasticsearchSinkConnector",
        "topics": "CUSTOMERS_WITH_ADDRESSES",
        "connection.url": "http://elasticsearch:9200",
        "type.name": "type.name=kafkaconnect",
        "key.ignore": "true",
        "schema.ignore": "true",
        "key.converter": "org.apache.kafka.connect.storage.StringConverter",
        "value.converter": "org.apache.kafka.connect.json.JsonConverter",
        "value.converter.schemas.enable": "false"
      }
    }'

Result:

$ curl -s http://localhost:9200/customers_with_addresses/_search | jq '.hits.hits[0]'
{
  "_index": "customers_with_addresses",
  "_type": "type.name=kafkaconnect",
  "_id": "CUSTOMERS_WITH_ADDRESSES+0+2",
  "_score": 1,
  "_source": {
    "ZIP": "76505",
    "CITY": "Temple",
    "ADDRESS_TYPE": "LIVING",
    "CUSTOMER_ID": 1002,
    "FULL_NAME": "Rebeca Kerrod",
    "STATE": "Texas",
    "STREET": "55795 Derek Avenue",
    "LAST_NAME": "Kerrod",
    "FIRST_NAME": "Rebeca"
  }
}

Answer 2

Yes, You can implement the solution using Kafka streams API in java in following way.

1. Consume the topics as stream.
2. Aggregate the address stream in a list using customer ID and convert the stream into table.
3. Join Customer stream with address table

Below is the example(considering data is consumed in json format) :

KStream<String,JsonNode> customers = builder.stream("customer", Consumed.with(stringSerde, jsonNodeSerde));
KStream<String,JsonNode> addresses = builder.stream("address", Consumed.with(stringSerde, jsonNodeSerde));

// Select the customer ID as key in order to join with address. 
KStream<String,JsonNode> customerRekeyed = customers.selectKey(value-> value.get("id").asText());

ObjectMapper mapper = new ObjectMapper();    
// Select Customer_id as key to aggregate the addresses and join with customer
KTable<String,JsonNode> addressTable = addresses
        .selectKey(value-> value.get("customer_id").asText())
        .groupByKey()
        .aggregate(() ->mapper::createObjectNode,  //initializer
                   (key,value,aggregate) -> aggregate.add(value),
                 Materialized.with(stringSerde, jsonNodeSerde)
         );  //adder

// Join Customer Stream with Address Table
KStream<String,JsonNode> customerAddressStream = customerRekeyed.leftJoin(addressTable,
               (left,right) -> {
                      ObjectNode finalNode = mapper.createObjectNode();
                      ArrayList addressList = new ArrayList<JsonNode>();
                      // Considering the address is arrayNode
                      ((ArrayNode)right).elements().forEachRemaining(addressList ::add);
                      left.putArray("addresses").allAll(addressList);              
                      return left;
               },Joined.keySerde(stringSerde).withValueSerde(jsonNodeSerde));

You can refer the details about all type of joins here :

https://docs.confluent.io/current/streams/developer-guide/dsl-api.html#joining

Answer 3

We built a demo and blog post on this very use case (streaming aggregates to Elasticsearch) a while ago on the Debezium blog.

One issue to keep in mind is that this solution (based on Kafka Streams, but I reckon it's the same for KSQL) is prone to exposing intermediary join results. Eg assume you insert a customer and 10 addresses in one transaction. The stream join approach might first produce an aggregate of the customer and their first five addresses and shortly thereafter the complete aggregate with all the 10 addresses. This might or might not be desirable for your specific use case. I also remember that handling deletions isn't trivial (eg if you delete one of the 10 addresses, so you'll have to produce the aggregate again with the remaining 9 addresses with might have been untouched, though).

An alternative to consider can be the outbox pattern where you'd essentially produce an explicit event with the precomputed aggregated from within your application itself. Ie it requires a little help of the application, but then it avoids the subtleties of producing that join result after the fact.