A practical guide to the quirks of Kafka Streams

Posted on 2018-02-26 by

Kafka Streams is a lightweight library that reads from a Kafka topic, does some processing and writes back to a Kafka topic. It processes messages one by one but is also able to keep some state. Because it leverages the mechanics of Kafka consumer groups, scaling out is easy. Each instance of your Kafka Streams application will take responsibility for a subset of topic partitions and will process a subset of the keys that go through your stream.

Kafka Streams at first glance

One of the first things we need to keep in mind when we start developing with Kafka Streams is that there are 2 different API’s that we can use. There is the high-level DSL (using the KStreamsBuilder) and the low-level Processor API (using the TopologyBuilder). The DSL is easier to use than the Processor API at the cost of flexibility and control over details.

We will begin our journey by learning more about the DSL, knowing that sooner or later we will have to rewrite our application using the Processor API. It doesn’t take long before we stumble upon the word count example.

val source: KStream[String, String] = KStreamBuilderS.stream[String, String]("streams-file-input")
val counts: KTableS[String, Long] = source
  .flatMapValues { value => value.toLowerCase(Locale.getDefault).split(" ") }
  .map { (_, value) => (value, value) }

Easy! Let’s just skip the documentation and finish our application!

Our first Kafka Streams application

Unfortunately, we soon return to the documentation because each aggregation on our KStream seems to return a KTable and we want to learn more about this stream/table duality. Aggregations also allow for windowing, so we continue reading about windowing. Now that we know something about the theoretical context, we return to our code. For our use case, we need 1 aggregate result for each window. However, we soon discover that each input message results in an output message on our output topic. This means that all the intermediate aggregates are spammed on the output topic. This is our first disappointment.

A simple join

The DSL has all the usual suspects like filter, map, and flatMap. But even though join is also one of those usual suspects that we have done a thousand times, it would be best to read the documentation on join semantics before trying out some code. For in Kafka Streams, there are a bit more choices involved in joining, due to the stream/table duality. But whatever our join looks like, we should know something about Kafka’s partitioning. Joins can be done most efficiently when the value that you want to join on, is the key of both streams and both source topics have the same number of partitions. If this is the case, the streams are co-partitioned, which means that each task that is created by the application instances, can be assigned to one (co-)partition where it will find all the data it needs for doing the join in one place.


Wherever there are aggregations, windows or joins, there is state involved. KS will create a new RocksBD StateStore for each window in order to store the messages that fall within that window. Unfortunately, we would like to have a lot of windows in our application, and since we also have about 500,000 different keys, we soon discover that this quickly grows out of hands. After having turned the documentation inside out, we learn that each one of those stores has a cache size of 100 MB by default. But even after we change this to 0, our KS application is too state-heavy.

Interactive queries

The same StateStores that allow us to do joining and aggregating also allows us to keep a materialized view of a Kafka topic. The data in the topic will be stored by the application. If the topic is not already compacted, the local key-value-store will compact your data locally. The data will constantly be updated from the topic it is built from.

The store can be scaled out by running an extra instance of our application. Partitions will be divided among tasks, and tasks will be divided among instances, which results in each instance holding a subset of the data. This can lead to the situation where an instance is queried for a key that is contained in another instance. The queried instance may not be able to provide the value corresponding to the key, it knows, however, which other instance does hold the key. So we can relay the query. The downside is that we have to implement this API by ourselves.

In order for this to work, we need an extra configuration element

    val p = new Properties()
    p.put(StreamsConfig.APPLICATION_SERVER_CONFIG, s"${settings.Streams.host}:${settings.Streams.port}")

Testing our app

A popular library for unit-testing Kafka Streams apps seems to be MockedStreams. However, not all topologies can be successfully tested with MockedStreams. But we can skip unit testing; integration tests are more important, anyway. Should we try using some EmbeddedKafka or spin up docker containers with docker-it? In the future, testing Kafka Streams apps will hopefully be easier (https://cwiki.apache.org/confluence/display/KAFKA/KIP-247%3A+Add+public+test+utils+for+Kafka+Streams).

When testing our topology, we wonder why our stream is not behaving like we would expect. After a while, we start banging our heads against a wall. Then we turn the documentation inside out again and we find some settings that may be helpful.

    p.put(StreamsConfig.COMMIT_INTERVAL_MS_CONFIG, "100")
    p.put(ProducerConfig.BATCH_SIZE_CONFIG, "1")

Our second Kafka Streams application

Then the inevitable happens: we need to do something that is not covered by the DSL; we need to resort to the Processor API. This would look a bit like the code below.

    val builder: TopologyBuilder = new KStreamBuilder()
      .addSource("in", keySerde.deserializer(), valueSerde.deserializer(), settings.Streams.inputTopic)
      .addProcessor("myProcessor", new MyProcessorSupplier(), "in")
      .addSink("out", settings.Streams.outputTopic, keySerde.serializer(), valueSerde.serializer(), "myProcessor")

The Processor interface lets us implement a process method that is called for each message, as well as a punctuate method that can be scheduled to run periodically. Inside these methods, we can use ProcessorContext.forward to forward messsages down the topology graph.

Periodically, in Kafka 0.11, means stream-time which means punctuate will be triggered by the first message that comes along after the method is scheduled to run. In our case we want to use wallclock-time, so we use the ScheduledThreadPoolExecutor to do our own scheduling. But if we do this, the ProcessorContext might have moved to a different node in our topology and the forward method will have unexpected behavior. The workaround for this is to get hold of the current node object and pass it along to our scheduled code.

val currentNode = context.asInstanceOf[ProcessorContextImpl].currentNode().asInstanceOf[ProcessorNode[MyKey, MyValue]]

In Kafka 1.0 a PunctuationType was introduced to make it possible to choose between wallclock-time and stream-time.


By the time we had finished our application, we had re-written it a few times, seen all parts of the documentation more often than we would like and searched through the KS source code. We were unfamiliar with all the settings and unaware of all the places where messages could be cached, delayed, postponed or dropped, and at certain moments we started doubting our notion of time.

In retrospect, we should have kept things simple. Don’t use too much state. Don’t think a clever Processor will bend KS’s quirks in our favor. Don’t waste too much code on workarounds, because before you know it there will be a new version released that will break the API or obsolete our workarounds.

And for those of you wanting to write an article or blog-post on Kafka Streams, make sure to finish it before the new release gets out.


  • https://docs.confluent.io/current/streams/
  • http://www.bigendiandata.com/2016-12-05-Data-Types-Compared/
  • https://www.confluent.io/blog/avro-kafka-data/
  • https://github.com/confluentinc/kafka-streams-examples
  • https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Streams+Join+Semantics

Leave a Reply

Your email address will not be published. Required fields are marked *