This is an example of a Kafka Streams based microservice (packaged in form of an Uber JAR). The scenario is simple

Basics

• A producer application continuously emits CPU usage metrics into a Kafka topic (cpu-metrics-topic)
• The consumer is a Kafka Streams application which uses the Processor (low level) Kafka Streams API to calculate the Cumulative Moving Average of the CPU metrics of each machine
• Consumers can be horizontally scaled - the processing work is distributed amongst many nodes and the process is elastic and flexible thanks to Kafka Streams (and the fact that it leverages Kafka for fault tolerance etc.)
• Each instance has its own (local) state for the calculated average. A custom REST API has been (using Jersey JAX-RS implementation) to tap into this state and provide a unified view of the entire system (moving averages of CPU usage of all machines)

Setup

This project has two modules

• Producer application
• Consumer application (uses Kafka Streams)

To try things out...

• Start Kafka broker. Configure num.partitions in Kafka broker server.properties file to 5 (to experiment with this application)
• Build producer & consumer application - browse to the respective directory and execute mvn clean install
• Trigger producer application - java -jar kafka-cpu-metrics-producer.jar. It will start emitting records to Kafka
• Start one instance of consumer application - java -jar kafka-cpu-metrics-consumer-1.0.jar -DKAFKA_CLUSTER=<kafka host:port> (defaults to localhost:9092 if not provided). Note the auto-selected port. It will start calculating the moving average of machine CPU metrics
• Access the metrics on this instance - http://localhost:<inst1_port>/metrics
• Start another instance of consumer application - java -jar kafka-cpu-metrics-consumer-1.0.jar -DKAFKA_CLUSTER=<kafka host:port> (note the auto-selected port), wait for a few seconds - the load will now be distributed amongst the two instances. Access the metrics http://localhost:<inst2_port>/metrics - you will see the metrics (JSON/XML payload) for all the machines as well as the instance on which the Cumulative Moving Average has been calculated

sample output - https://gist.github.com/abhirockzz/48e89873ae23c93d0a5cc721c87cc536

• You can also search for metrics for a specific machine ID http://localhost:<any_app_port>/metrics/<machineID>

sample output - https://gist.github.com/abhirockzz/2ca2297fbc9aec269d31707f61b4c45e

You can keep increasing the number of instances such that they are less than or equal to the number of partitions of your Kafka topic

Having more instances than number of partitions is not going to have any effect on parallelism and that instance will be inactive until any of the existing instance is stopped