Hey there folks 👋 I was using Flink with my custom helm templating and I decided to migrate to the Flink K8s Operator. I’m building my own docker image and publishing it. I override the image following this example. I’m using the mode standalone mode (

mode: standalone

) to be able to use Flink in the reactive mode (

scheduler-mode: reactive

). My task manager pod has my modifications (

/opt/flink/usrlib/

jars) but my job manager pod seems to have the same content as the default flink image. Do you know what could lead to this? My deployment is

apiVersion: <http://flink.apache.org/v1beta1|flink.apache.org/v1beta1>
kind: FlinkDeployment
metadata:
  name: {{ template "flink-application-mode.name" . }}
spec:
  image: "{{ .Values.imageRepository }}/{{ .Values.imageName }}:{{ .Values.imageTag }}"
  imagePullPolicy: IfNotPresent
  serviceAccount: {{ .Values.serviceAccountName }}
  flinkVersion: v1_16
  podTemplate:
    spec:
      containers:
        - name: flink-main-container
          env:
            - name: APPLICATION_CREDENTIALS
              value: /secrets/credentials.json
          volumeMounts:
            ...
      volumes:
        ...
  flinkConfiguration:
    scheduler-mode: reactive
  jobManager:
    replicas: 1
    resource:
      {{- toYaml .Values.resources.jobmanager.requests | nindent 6 }}
  taskManager:
    replicas: {{ .Values.replicaCount }}
    resource:
      {{- toYaml .Values.resources.taskmanager.requests | nindent 6 }}
  mode: standalone
  job:
    jarURI: local:///opt/flink/usrlib/custom.jar
    entryClass: {{ required "A valid .Values.jobmanager.entryClass entry required!" .Values.jobmanager.entryClass }}
    parallelism: 2
    upgradeMode: savepoint # last-state OR stateless

Do you get the same problem with

native

mode?

Hey @Gyula Fóra Thank you for the reply. I did not know about the

native

mode. What is the difference between it and

standalone

? I’m sticking to

standalone

because it is what I had before. But if it does sort my issues I’m happy to move 😛 Does the

native

mode enables manual replica changes as well? I’m sticking with

reactive

to be sure if I upscale/downscale taskmanagers, they do rebalance without a partition reattribution (I’m mainly using Kafka).

and in general I recommend that for prod over standalone + reactive

it means they works “the same”, am I getting this right?

due to the limitations of standalone setup

Which limitations are we talking about? Thnak you again for the support 🤝

Native is the default mode for the operator. It uses the Flink native Kubernetes integration which automatically creates taskmanager pods individually on-demand.

In my case I was setting the replicas manually. and the instruction on the docs is to use standalone when using Flink in application mode. And reactive to be able to upscale/downscale tasknmanagers on demand (manually tho)

What docs are you referring to?

Here and

reactive

is only supported by

standalone

(otherwise the deployment breaks saying it only is supported by it)

okay, let me put it this way. I am 99% sure that you don’t need the reactive mode

Fair. I’ll try with the

native

mode. Do you mind if I ping you in this thread later on? Thank you for chiming in

The reactive mode was created at a time before the Flink Kubernetes Operator. It allowed the job to adjust to the cluster available resources. But this is not needed with the operator. You change you deployment yaml (

spec.job.parallelism

) and the entire purpose of the operator that it knows how to create more reasources and upgrade your job

So the

native

mode made the trick indeed. I’m setting

taskManager.replicas

to

1

and

job.parallelism

to

32

(the kafka topic I’m reading has 32 partitions). I was expecting the operator to spin up a single task manager, but it is spinning 16 instances don’t matter how I change the

job.parallelism

or

taskManager.replicas

. Do you guys know if this is the expected behaviour indeed? And if not

taskManager.replicas

, how should I limit the replicas of the task manager using Flink k8s operator? PS: I’m also posting outside of the thread because it is another problem 😄

This may be due to the scheduler-mode: reactive setting you have specified: https://nightlies.apache.org/flink/flink-docs-master/docs/deployment/elastic_scaling/#configuration

Task manager replicas takes no effect in native mode because it's automatically computed based on parallelism and number of task slots config of the tm

I changed back to

standalone

mode. The trick with the JM was this piece of code here. To prevent this behaviour, I mounted a configmap volume using a

subpath

inside

$FLINK_HOME/usrlib/

to prevent a new volume to be mounted there. Thank you @Gyula Fóra for the support on the

native

mode.