Kafka Transactions

Problem statement

We want to create a streaming application that will make use of some simple account transaction messages.

{
  "customer": "Bob",
  "amount": 64,
  "category": "CG01",
  "occurred": 1567148940000  
}

Broadly we will perform the following tasks:

Calculate the total for each customer.
Calculate the total for a month for each customer
Enrich the transaction data with another source.

To do this we will use 5 topics, 2 input and 3 output:

Topic	Type	Key	Value
transactions-topic	Input	String	json (String)
category-topic	Input	String	String
customer-total-topic	Output	String	Long
customer-rolling-total-topic	Output	String	Long
enhanced-transactions-topic	OUtput	String	json (String)

We do not need to tell Kafka what it will be storing when we create the topics, but we do need to tell our Producer and Streaming Application how to serialize and deserialize these values.

The Producer

Ultimately we will be writing transactions to a topic transactions-topic in json format. Have a bit of a think about what key we should use!

We will be following kafka's exactly-once semantics, read about it here: https://www.confluent.io/blog/exactly-once-semantics-are-possible-heres-how-apache-kafka-does-it/

The producer will have access to a LegacyBankingSystem that will provide you with transactions to be published, so don't worry about generating this data yourself.

Calculating a customers balance

We will then create a streaming topology that can handle these events. We will then aggregate a customers balance and write this back to another topic - for this we will use a compacted topic: https://kafka.apache.org/documentation/#compaction

If we were to see the following events to the transactions-topic:

[{
  "customer": "Bob",
  "amount": 64,
  "category": "CG01",
  "occurred": 1567148940000  
},
{
  "customer": "Bob",
  "amount": 64,
  "category": "CG01",
  "occurred": 1567148940000  
},
{
  "customer": "Alice",
  "amount": 64,
  "category": "CG01",
  "occurred": 1567148940000  
}]

How many records should be in the customer-total-topic? And what should their values be?

Calculating the last 30 day spend.

We will use the same topology to create another view of the data - the total spend over the last 30 days (so the balance will reset to 0 after 30 days)

Again we will look to use a Compacted topic here.

Enriching transaction data.

You might notice that there is a category code embedded into each transaction record. Not very useful huh? Let's use a GlobalKTable to allow us to reference data sourced from the category-topic to enrich out transactions. If there is no corresponding value for a category, we will retain the code value.

https://docs.confluent.io/current/streams/concepts.html#streams-concepts-ktable

Scripts to help

Go have a look at create-topics.sh file for more details...

Create the topics

Create the input topic:

kafka-topics --bootstrap-server kafka-1:19092 --create --topic transaction-topic --partitions 1 --replication-factor 1

Create the category topic:

kafka-topics --bootstrap-server kafka-1:19092 --create --topic category-topic --partitions 1 --replication-factor 1

Create the enhanced transaction topic

kafka-topics --bootstrap-server kafka-1:19092 --create --topic enhanced-transaction-topic --partitions 1 --replication-factor 1

Create the running total topic:

kafka-topics --bootstrap-server kafka-1:19092 --create \
            --topic customer-total-topic \
            --partitions 1 \
            --replication-factor 1 \
            --config cleanup.policy=compact \
            --config min.cleanable.dirty.ratio=0.01 \
            --config segment.ms=100

You will notice here are are using a cleanup.policy of compact, but also making the segment super small (100ms)! This is so we can see the log compaction happening without waiting the default hours to see compaction in action.

Create the windowed topic

kafka-topics --bootstrap-server kafka-1:19092 --create \
            --topic customer-rolling-total-topic \
            --partitions 1 \
            --replication-factor 1 \
            --config cleanup.policy=compact \
            --config min.cleanable.dirty.ratio=0.01 \
            --config segment.ms=100

Consume a topic

We will use the console consumer to keep track of how things are going:

 kafka-console-consumer.sh --bootstrap-server kafka-1:19092 \
             --topic customer-total-topic \
             --from-beginning \
             --formatter kafka.tools.DefaultMessageFormatter \
             --property print.key=true \
             --property print.value=true \
             --property key.deserializer=org.apache.kafka.common.serialization.StringDeserializer \
             --property value.deserializer=org.apache.kafka.common.serialization.LongDeserializer

Other useful commands

# list topics
kafka-topics --bootstrap-server kafka-1:19092 --list
 
 
 # Read the raw transactions
 kafka-console-consumer --bootstrap-server kafka-1:19092 \
             --topic transaction-topic \
             --from-beginning \
             --formatter kafka.tools.DefaultMessageFormatter \
             --property print.key=true \
             --property print.value=true \ 
             
# Produce directly to the queue             
kafka-console-producer --broker-list kafka-1:19092 --topic category-topic --property "parse.key=true" --property "key.separator=:"

Branching for your convenience

We will be working through problems as we go - to help us along there are branches created that can step us through the solution.

master This is the first branch to look at. Nothing is solved!
producer-complete This branch solves the problem of putting messages onto the transaction-topic
totals-complete This branch solves the account total aggregation
rolling-totals-complete This branch solves the rolling window problem
transaction-enhancement-complete This branch solves the data enhancement problem
complete This branch sinks the various streams back to topics

deanbaker / kafka-bank-transactions