Skip to content

Latest commit

 

History

History

1.11

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 

Fast Data: Financial Transaction Processing with Apache Flink

During this demo we use Apache Flink and Apache Kafka to setup a high-volume financial transactions pipeline.

  • Estimated time for completion:
  • Manual install: 15min
  • Target audience: Anyone interested in stream data processing and analytics with Apache Kafka and Apache Flink.

A video of this demo can be found here.

Table of Contents:

Architecture

Financial transaction processing demo architecture

This demo implements a data processing infrastructure that is able to spot money laundering. In the context of money laundering, we want to detect amounts larger than $10.000 transferred between two accounts, even if that amount is split into many small batches. See also US and EU legislation and regulations on this topic for more information.

The architecture follows more or less the SMACK stack architecture:

  • Events: Event are being generated by a small GOLANG generator. The events are in the form 'Sunday, 23-Jul-17 01:06:47 UTC;66;26;7810', where the first field '23-Jul-17 01:06:47 UTC' represents the (increasing) timestamp of transactions; the second field '66' represent the sender account; the third field the receiver account; and the fourth field represent the dollar amount transferred during that transaction.
  • Ingestion: The generated events are being ingested and buffered by a Kafka queue with the default topic 'transactions'.
  • Stream Processing: As we require fast response times, we use Apache Flink as a Stream processor running the FinancialTransactionJob.
  • Storage: Here we diverge a bit from the typical SMACK stack setup and don't write the results into a Datastore such as Apache Cassandra. Instead we write the results again into a Kafka Stream (default: 'fraud'). Note, that Kafka also offers data persistence for all unprocessed events.
  • Actor: In order to view the results we use again a small Golang viewer which simply reads and displays the results from the output Kafka stream.

Prerequisites

  • A running DC/OS 1.11 or higher cluster with at least 3 private agents and 1 public agent. Each agent should have 2 CPUs and 5 GB of RAM available. The DC/OS CLI also needs to be installed.
  • The dcos/demo Git repo must be available locally, use: git clone https://github.com/dcos/demos/ if you haven't done so yet.
  • SSH cluster access must be set up.

The DC/OS services used in the demo are as follows:

  • Apache Kafka
  • Apache Flink

Install

Kafka

Install the Apache Kafka package :

$ dcos package install kafka

Note that if you are unfamiliar with Kafka and its terminology, you can check out the respective 101 example.

Next, figure out where the broker is:

$ dcos kafka endpoints broker
{
  "address": [
    "10.0.2.64:1025",
    "10.0.2.83:1025",
    "10.0.0.161:1025"
  ],
  "dns": [
    "kafka-0-broker.kafka.autoip.dcos.thisdcos.directory:1025",
    "kafka-1-broker.kafka.autoip.dcos.thisdcos.directory:1025",
    "kafka-2-broker.kafka.autoip.dcos.thisdcos.directory:1025"
  ],
  "vip": "broker.kafka.l4lb.thisdcos.directory:9092"
}

Note the FQDN for the vip, in our case broker.kafka.l4lb.thisdcos.directory:9092, which is independent of the actual broker locations. It is possible to use the FQDN of any of the brokers, but using the VIP FQDN will give us load balancing.

Create Kafka Topics

Fortunately, creating topic is very simple using the DC/OS Kafka CLI. If you have installed Kafka from the UI you might have to install the cli extensions using `dcos pacakge install kafka --cli'. If you installed Kafka as above using the CLI then it will automatically install the CLI extensions.

We need two Kafka topics, one with the generated transactions and one for fraudulent transactions, which we can create with:

dcos kafka topic create transactions and dcos kafka topic create fraud

Generator

After the Kafka queues are created, we can now deploy the data generator with this marathon app definition. As the GoLang generator contains all its dependencies, the app defintion doesn't require a container image and is running without Docker.

dcos marathon app add https://raw.githubusercontent.com/dcos/demos/master/flink/1.11/generator/generator.json

Flink

Finally, we can deploy Apache Flink :

$ dcos package install flink

At this point we have all of the required elements installed - Kafka, Flink and the data generator. We are now ready to start the demo.

Final View

After install your DC/OS UI should look as follows:

All services of the fintrans demo in the DC/OS UI

Use

The core piece of this demo is the FinancialTransactionJob which we will submit to Flink.

First we need to upload the jar file into Flink. Please note that the jar file is too large to be included in this github repo, but can be downloaded here.

In the Services tab of the DCOS UI, hover over the name of the flink service, and click on the link which appears to the right of it. This will open the Flink web UI in a new tab.

Flink UI

In the Flink web UI, click on Submit New Job, then click the Add New button. This will allow you to select the jar file from $DEMO_HOME and upload it.

Jar file uploaded

Once we hit Submit, we should see the job begin to run in the Flink web UI.

Running Flink job

Viewing Output

Now once the Flink job is running, we only need a way to visualize the results. We do that with another simple GoLang program.

dcos marathon app add https://raw.githubusercontent.com/dcos/demos/master/flink/1.11/actor/fraudDisplay.json

We can easily check the output by checking the task output from the UI.

Task Output in the UI

Should you have any questions or suggestions concerning the demo, please raise an issue in Jira or let us know via the [email protected] mailing list.