Flink Getting started¶

Update

Created 2018
Updated 12/21/2024 - review done.

This chapter reviews the different environments for deploying Flink jobs on a developer's workstation. Options include using downloading product tar file, Docker Compose, Minikube ot Colima -k3s, or a hybrid approach that combines a Confluent Cloud Kafka cluster with a local Flink instance. For Confluent Cloud for Flink see this note.

Install locally¶

The tar file can be downloaded from the open-source web site as a tar file and untar. The script in 'deployment/product-tar` folder does this download and untar.

Once done start flink using the start-cluster.sh script in flink-1.19.1/bin. See product documentation.

 ./bin/start-cluster.sh

With docker images¶

Pre-requisites¶

Get docker cli, helm, and kubectl
Clone this repository.
For docker container execution, you need a docker engine, with docker compose CLIs, Colima or Minikube and docker-ce engine.

Three choices: Colima with Kubernetes, Minikube or docker compose, for each of those environment see the next section and for Flink Kubernetes operator deployment and configuratuin see the dedicated chapter.

Colima with Kubernetes¶

An alternate to use Docker Desktop, Colima is an open source to run container on Linux or Mac.

Start a k3s cluster:

colima start --kubernetes

Deploy the Flink and Confluent Platform operators (see Makefile in deployment/k8s and its readme).

Define a Flink cluster, optional a Kafka Cluster.

Minikube¶

Install Minikube, and review some best practices
Start with enough memory and cpu

minikube start --cpus='3' --memory='4096'

Only to newly created minikube profile, install Flink Operator for kubernetes

If we want integration with Kafka and Schema registry select one platform:

Install Confluent Plaform Operator

kubectl create namespace confluent
kubectl config set-context --current --namespace confluent
helm repo add confluentinc https://packages.confluent.io/helm
helm repo update
helm upgrade --install confluent-operator confluentinc/confluent-for-kubernetes

Or Strimzi Operator in the kafka namespace:

kubectl create namespace kafka
kubectl config set-context --current --namespace kafka
kubectl create -f 'https://strimzi.io/install/latest?namespace=kafka' -n kafka

with Apicu.io for Operator for schema management.

Docker Desktop and Compose¶

During development, we can use docker-compose to start a simple Flink session cluster or a standalone job manager to execute one unique job, which has the application jar mounted inside the docker image. We can use this same environment to do SQL based Flink apps.

As Task manager will execute the job, it is important that the container running the flink has access to jars needed to connect to external sources like Kafka or other tools like FlinkFaker. Therefore there is a Dockerfile to get some important jars to build a custom Flink image that we will use for Taskmanager and SQL client.

If specific integrations are needed, get the needed jar references, update the dockerfile and then build the Custom Flink image, under deployment/custom-flink-image folder

docker build -t jbcodeforce/myflink .

Start Flink session cluster using the following command:

# under this repository and deployment/local folder
docker compose up -d

The docker compose starts one job manager and one task manager server:

services:
  jobmanager:
    image: flink:latest
    hostname: jobmanager
    ports:
      - "8081:8081"
    command: jobmanager
    user: "flink:flink"
    environment:
      FLINK_PROPERTIES: "jobmanager.rpc.address: jobmanager"
    volumes:  
        - .:/home
  taskmanager:
    image: flink:latest 
    hostname: taskmanager
    depends_on:
      - jobmanager
    command: taskmanager
    user: "flink:flink"
    scale: 1
    volumes:
        - .:/home
    environment:
      - |
        FLINK_PROPERTIES=
        jobmanager.rpc.address: jobmanager
        taskmanager.numberOfTaskSlots: 4

The docker compose mounts the local folder to /home in both the job manager and task manager containers so that, we can submit jobs from the job manager (accessing the compiled jar) and also access the input data files in the task manager container.

See this section to deploy an application with flink

Docker compose with Kafka and Flink¶

In the deployment/local folder the docker compose start a one node kafka broker, one zookeeper, one job manager and one task manager.

docker compose -f kafka-docker-compose.yaml up -d

The SQL client can be used to compute some aggregation on the sale events created by the E-commerce simulator. To start the simulator using a Python virtual environment do:

pip install -r requirements.txt
python simulator.py

The application sends events like the following:

{'event_type': 'user_action', 
 'timestamp': '2024-09-04T15:24:59.450582', 
 'user_id': 'user5', 
 'action': 'add_to_cart', 
 'page': 'category', 
 'product': 'headphones'
}

Use the Kafdrop interface to verify the messages in the topic
Connect to SQL client container

docker exec -ti sql-client bash
# in the shell
./sql-client.sh

User page view on kafka stream

CREATE TABLE user_page_views (
    event_type STRING,
    user_id STRING,
    action STRING,
    page STRING,
    product STRING,
    timestamp_str STRING,        # (1)
    timestamp_sec TIMESTAMP(3),  # derived field
    WATERMARK FOR timestamp_sec AS TO_TIMESTAMP(timestamp_str, 'yyyy-MM-dd HH:mm:ss') - INTERVAL '5' SECOND
) WITH (
    'connector' = 'kafka',
    'topic' = 'ecommerce_events',
    'properties.bootstrap.servers' = 'kafka:29092',
    'properties.group.id' = 'sql-flink-grp-1',
    'properties.auto.offset.reset' = 'earliest',
    'format' = 'json'  
);

The event timestamp is a string created by the Kafka producer

WATERMARK statement is used to define a watermark strategy for handling event time in streaming applications. Watermarks are crucial for dealing with out-of-order events, allowing Flink to manage late arrivals and trigger processing based on event time rather than processing time. A watermark is a timestamp that indicates that no events with a timestamp earlier than the watermark will arrive.

It is important to set the consumer properties like consumer group id, the offset reset strategy...

The next SQL statement is to count the number of page per user

SELECT 
    user_id, 
    page,
    COUNT(page) AS page_views 
FROM 
    user_page_views 
GROUP BY 
    user_id,
    page;

The results

Confluent Cloud¶

See getting started product documentation and this summary.

To use the Confluent client flink sql client see this note