Flink on Kubernetes - Session vs Job cluster mode and storage

[M Singh]

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Yang Wang]

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[M Singh]

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Yang Wang]

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Arvid Heise-3]

Creds on AWS are typically resolved through roles assigned to K8s pods (for example with KIAM [1]).

[1] https://github.com/uswitch/kiam

On Tue, Feb 25, 2020 at 3:36 AM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[M Singh]

Thanks Yang and Arvid for your advice and pointers.  Mans

On Wednesday, February 26, 2020, 09:52:26 AM EST, Arvid Heise <[hidden email]> wrote:

Creds on AWS are typically resolved through roles assigned to K8s pods (for example with KIAM [1]).

[1] https://github.com/uswitch/kiam

On Tue, Feb 25, 2020 at 3:36 AM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[M Singh]

BTW - Is there any limit to the amount of data that can be stored on emptyDir in K8 ?  

On Wednesday, February 26, 2020, 07:33:54 PM EST, M Singh <[hidden email]> wrote:

Thanks Yang and Arvid for your advice and pointers.  Mans

On Wednesday, February 26, 2020, 09:52:26 AM EST, Arvid Heise <[hidden email]> wrote:

Creds on AWS are typically resolved through roles assigned to K8s pods (for example with KIAM [1]).

[1] https://github.com/uswitch/kiam

On Tue, Feb 25, 2020 at 3:36 AM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Yang Wang]

I think the only limitation is the disk size of your kubelet machine. Please remember

to set the "sizeLimit" of your empty dir. Otherwise, your pod may be killed due to 

ephemeral storage is full.

Best,

Yang

M Singh <[hidden email]> 于2020年2月27日周四 上午8:34写道：

BTW - Is there any limit to the amount of data that can be stored on emptyDir in K8 ?  

On Wednesday, February 26, 2020, 07:33:54 PM EST, M Singh <[hidden email]> wrote:

Thanks Yang and Arvid for your advice and pointers.  Mans

On Wednesday, February 26, 2020, 09:52:26 AM EST, Arvid Heise <[hidden email]> wrote:

Creds on AWS are typically resolved through roles assigned to K8s pods (for example with KIAM [1]).

[1] https://github.com/uswitch/kiam

On Tue, Feb 25, 2020 at 3:36 AM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Eleanore Jin]

Hi Yang,

regarding your statement below:

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

Does it also apply to Flink Job Cluster? When the JM pod restarted by Kubernetes, the image contains the application jar also, so if the statement also applies to the Flink Job Cluster mode, can you please elaborate why? 

Thanks a lot! 

Eleanore

On Mon, Feb 24, 2020 at 6:36 PM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Yang Wang]

Hi Jin Yi,

For standalone per-job cluster, it is a little different about the recovery. 

Just as you say, the user jar has built in the image, when the JobManager failed

and relaunched by the K8s, the user `main()` will be executed again to get the

job graph, not like session cluster to get the job graph from high-availability storage.

Then the job will be submitted again and the job could recover from the latest

checkpoint(assume that you have configured the high-availability).

Best,

Yang

Jin Yi <[hidden email]> 于2020年2月27日周四 下午2:50写道：

Hi Yang,

regarding your statement below:

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

Does it also apply to Flink Job Cluster? When the JM pod restarted by Kubernetes, the image contains the application jar also, so if the statement also applies to the Flink Job Cluster mode, can you please elaborate why? 

Thanks a lot! 

Eleanore

On Mon, Feb 24, 2020 at 6:36 PM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Yang Wang]

Hi Hao Sun,

I just post the explanation to the user ML so that others could also have the same problem.

Gven the job graph is fetched from the jar, do we still need Zookeeper for HA? Maybe we still need it for checkpoint locations?

Yes, we still need the zookeeper(maybe in the future we will have a native K8s HA based on etcd) for the complete recovery. You

are right. We still need it for finding the checkpoint locations. Also the Zookeeper will be used for leader election and leader retriever.

Best,

Yang

Hao Sun <[hidden email]> 于2020年2月28日周五 上午1:41写道：

Hi Yang, given the job graph is fetched from the jar, do we still need Zookeeper for HA? Maybe we still need it for checkpoint locations?

Hao Sun

On Thu, Feb 27, 2020 at 5:13 AM Yang Wang <[hidden email]> wrote:

Hi Jin Yi,

For standalone per-job cluster, it is a little different about the recovery. 

Just as you say, the user jar has built in the image, when the JobManager failed

and relaunched by the K8s, the user `main()` will be executed again to get the

job graph, not like session cluster to get the job graph from high-availability storage.

Then the job will be submitted again and the job could recover from the latest

checkpoint(assume that you have configured the high-availability).

Best,

Yang

Jin Yi <[hidden email]> 于2020年2月27日周四 下午2:50写道：

Hi Yang,

regarding your statement below:

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

Does it also apply to Flink Job Cluster? When the JM pod restarted by Kubernetes, the image contains the application jar also, so if the statement also applies to the Flink Job Cluster mode, can you please elaborate why? 

Thanks a lot! 

Eleanore

On Mon, Feb 24, 2020 at 6:36 PM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.

[Hao Sun]

Sounds good. Thank you!

On Thu, Feb 27, 2020 at 6:52 PM Yang Wang <[hidden email]> wrote:

Hi Hao Sun,

I just post the explanation to the user ML so that others could also have the same problem.

Gven the job graph is fetched from the jar, do we still need Zookeeper for HA? Maybe we still need it for checkpoint locations?

Yes, we still need the zookeeper(maybe in the future we will have a native K8s HA based on etcd) for the complete recovery. You

are right. We still need it for finding the checkpoint locations. Also the Zookeeper will be used for leader election and leader retriever.

Best,

Yang

Hao Sun <[hidden email]> 于2020年2月28日周五 上午1:41写道：

Hi Yang, given the job graph is fetched from the jar, do we still need Zookeeper for HA? Maybe we still need it for checkpoint locations?

Hao Sun

On Thu, Feb 27, 2020 at 5:13 AM Yang Wang <[hidden email]> wrote:

Hi Jin Yi,

For standalone per-job cluster, it is a little different about the recovery. 

Just as you say, the user jar has built in the image, when the JobManager failed

and relaunched by the K8s, the user `main()` will be executed again to get the

job graph, not like session cluster to get the job graph from high-availability storage.

Then the job will be submitted again and the job could recover from the latest

checkpoint(assume that you have configured the high-availability).

Best,

Yang

Jin Yi <[hidden email]> 于2020年2月27日周四 下午2:50写道：

Hi Yang,

regarding your statement below:

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

Does it also apply to Flink Job Cluster? When the JM pod restarted by Kubernetes, the image contains the application jar also, so if the statement also applies to the Flink Job Cluster mode, can you please elaborate why? 

Thanks a lot! 

Eleanore

On Mon, Feb 24, 2020 at 6:36 PM Yang Wang <[hidden email]> wrote:

Hi M Singh,

> Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 
In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.
Let me know if I have misunderstood anything.

Since you are starting JM/TM with K8s deployment, when they failed new JM/TM will be created. If you do not set the high

availability configuration, your jobs could recover when TM failed. However, they could not recover when JM failed. With HA

configured, the jobs could always be recovered and you do not need to re-submit again.

> Mans - Is there any safe way of a passing creds ?

Yes, you are right, Using configmap to pass the credentials is not safe. On K8s, i think you could use secrets instead[1].

> Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

What i mean is the job failed, and it could be restarted by your configured restart strategy[2].

> Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

 

Yes, if you are saving the checkpoint in S3 and also set the `high-availability.storageDir` to S3. Then you do not need persistent volume. Since

the local directory is only used for local cache, so you could directly use the overlay filesystem or empryDir(better io performance).

[1]. https://kubernetes.io/docs/tasks/inject-data-application/distribute-credentials-secure/

[2]. https://ci.apache.org/projects/flink/flink-docs-master/ops/config.html#fault-tolerance

M Singh <[hidden email]> 于2020年2月25日周二 上午5:53写道：

Thanks Wang for your detailed answers.

From what I understand the native_kubernetes also leans towards creating a session and submitting a job to it.  

Regarding other recommendations, please my inline comments and advice.

On Sunday, February 23, 2020, 10:01:10 PM EST, Yang Wang <[hidden email]> wrote:

Hi Singh,

Glad to hear that you are looking to run Flink on the Kubernetes. I am

trying to answer your question based on my limited knowledge and

others could correct me and add some more supplements.

I think the biggest difference between session cluster and per-job cluster

on Kubernetesis the isolation. Since for per-job, a dedicated Flink cluster

will be started for the only one job and no any other jobs could be submitted.

Once the job is finished, then the Flink cluster will be destroyed immediately.

The second point is one-step submission. You do not need to start a Flink

cluster first and then submit a job to the existing session.

> Are there any benefits with regards to

1. Configuring the jobs

No matter you are using the per-job cluster or submitting to the existing

session cluster, they share the configuration mechanism. You do not have

to change any codes and configurations.

2. Scaling the taskmanager

Since you are using the Standalone cluster on Kubernetes, it do not provide

an active resourcemanager. You need to use external tools to monitor and

scale up the taskmanagers. The active integration is still evolving and you

could have a taste[1].

Mans - If we use the session based deployment option for K8 - I thought K8 will automatically restarts any failed TM or JM. 

In the case of failed TM - the job will probably recover, but in the case of failed JM - perhaps we need to resubmit all jobs.

Let me know if I have misunderstood anything.

3. Restarting jobs

For the session cluster, you could directly cancel the job and re-submit. And

for per-job cluster, when the job is canceled, you need to start a new per-job

cluster from the latest savepoint.

4. Managing the flink jobs

The rest api and flink command line could be used to managing the jobs(e.g.

flink cancel, etc.). I think there is no difference for session and per-job here.

5. Passing credentials (in case of AWS, etc)

I am not sure how do you provide your credentials. If you put them in the 

config map and then mount into the jobmanager/taskmanager pod, then both

session and per-job could support this way.

Mans - Is there any safe way of a passing creds ?

6. Fault tolerence and recovery of jobs from failure

For session cluster, if one taskmanager crashed, then all the jobs which have tasks

on this taskmanager will failed. 

Both session and per-job could be configured with high availability and recover

from the latest checkpoint. 

Mans - Does a task manager failure cause the job to fail ?  My understanding is the JM failure are catastrophic while TM failures are recoverable.

> Is there any need for specifying volume for the pods?

No, you do not need to specify a volume for pod. All the data in the pod 

local directory is temporary. When a pod crashed and relaunched, the

taskmanager will retrieve the checkpoint from zookeeper + S3 and resume

from the latest checkpoint.

Mans - So if we are saving checkpoint in S3 then there is no need for disks - should we use emptyDir ?

[1]. https://ci.apache.org/projects/flink/flink-docs-master/ops/deployment/native_kubernetes.html

M Singh <[hidden email]> 于2020年2月23日周日 上午2:28写道：

Hey Folks:

I am trying to figure out the options for running Flink on Kubernetes and am trying to find out the pros and cons of running in Flink Session vs Flink Cluster mode (https://ci.apache.org/projects/flink/flink-docs-stable/ops/deployment/kubernetes.html#flink-session-cluster-on-kubernetes).

I understand that in job mode there is no need to submit the job since it is part of the job image.  But what are other the pros and cons of this approach vs session mode where a job manager is deployed and flink jobs can be submitted it ?  Are there any benefits with regards to:

1. Configuring the jobs 

2. Scaling the taskmanager

3. Restarting jobs

4. Managing the flink jobs

5. Passing credentials (in case of AWS, etc)

6. Fault tolerence and recovery of jobs from failure

Also, we will be keeping the checkpoints for the jobs on S3.  Is there any need for specifying volume for the pods ?  If volume is required do we need provisioned volume and what are the recommended alternatives/considerations especially with AWS.

If there are any other considerations, please let me know.

Thanks for your advice.