flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jain, Ankit" <ankit.j...@here.com>
Subject Re: High Availability on Yarn
Date Thu, 11 May 2017 20:19:50 GMT
Got the answer on #2, looks like that will work, still looking for suggestions on #1.


From: "Jain, Ankit" <ankit.jain@here.com>
Date: Thursday, May 11, 2017 at 8:26 AM
To: Stephan Ewen <sewen@apache.org>, "user@flink.apache.org" <user@flink.apache.org>
Subject: Re: High Availability on Yarn

Following up further on this.

1)      We are using a long running EMR cluster to submit jobs right now and as you know EMR
hasn’t made Yarn ResourceManager HA.

Is there any way we can use the information put in Zookeeper by Flink Job Manager to bring
the jobs back up on a new EMR cluster if RM goes down?

We are not looking for completely automated option but maybe write a script which reads Zookeeper
and re-starts all jobs on a fresh EMR cluster?

I am assuming if  Yarn ResouceManager goes down, there is no way to just bring it back up
– you have to start a new EMR cluster?

2)      Regarding elasticity, I know for now a running flink cluster can’t make use of new
hosts added to EMR but can I am guessing Yarn will still see the new hosts and new flink jobs
can make use it, is that right?


From: "Jain, Ankit" <ankit.jain@here.com>
Date: Monday, May 8, 2017 at 9:09 AM
To: Stephan Ewen <sewen@apache.org>, "user@flink.apache.org" <user@flink.apache.org>
Subject: Re: High Availability on Yarn

Thanks Stephan – we will go with a central ZooKeeper Instance and hopefully have it started
through a cloudformation script as part of EMR startup.

Is Zk also used to keep track of checkpoint metadata and the execution graph of the running
job to recover from ApplicationMaster failure as Aljoscha was guessing below or only for leader
election in case of accidently running multiple Application Masters ?


From: Stephan Ewen <sewen@apache.org>
Date: Monday, May 8, 2017 at 9:00 AM
To: "user@flink.apache.org" <user@flink.apache.org>, "Jain, Ankit" <ankit.jain@here.com>
Subject: Re: High Availability on Yarn


ZooKeeper is required in YARN setups still. Even if there is only one JobManager in the normal
case, Yarn can accidentally create a second one when there is a network partition.
To prevent that this leads to inconsistencies, we use ZooKeeper.

Flink uses ZooKeeper very little, so you can just let Flink attach to any existing ZooKeeper,
or user one ZooKeeper cluster for very many Flink clusters/jobs.


On Mon, May 8, 2017 at 2:11 PM, Aljoscha Krettek <aljoscha@apache.org<mailto:aljoscha@apache.org>>
Yes, it’s recommended to use one ZooKeeper cluster for all Flink clusters.


On 5. May 2017, at 16:56, Jain, Ankit <ankit.jain@here.com<mailto:ankit.jain@here.com>>

Thanks for the update Aljoscha.

@Till Rohrmann<mailto:trohrmann@apache.org>,
Can you please chim in?

Also, we currently have a long running EMR cluster where we create one flink cluster per job
– can we just choose to install Zookeeper when creating the EMR cluster and use one Zookeeper
instance for ALL of flink jobs?
Recommendation is to have a dedicated Zookeeper instance per flink job?


From: Aljoscha Krettek <aljoscha@apache.org<mailto:aljoscha@apache.org>>
Date: Thursday, May 4, 2017 at 1:19 AM
To: "Jain, Ankit" <ankit.jain@here.com<mailto:ankit.jain@here.com>>
Cc: "user@flink.apache.org<mailto:user@flink.apache.org>" <user@flink.apache.org<mailto:user@flink.apache.org>>,
Till Rohrmann <trohrmann@apache.org<mailto:trohrmann@apache.org>>
Subject: Re: High Availability on Yarn

Yes, for YARN there is only one running JobManager. As far as I Know, In this case ZooKeeper
is only used to keep track of checkpoint metadata and the execution graph of the running job.
Such that a restoring JobManager can pick up the data again.

I’m not 100 % sure on this, though, so maybe Till can shed some light on this.

On 3. May 2017, at 16:58, Jain, Ankit <ankit.jain@here.com<mailto:ankit.jain@here.com>>

Thanks for your reply Aljoscha.

After building better understanding of Yarn and spending copious amount of time on Flink codebase,
I think I now get how Flink & Yarn interact – I plan to document this soon in case it
could help somebody starting afresh with Flink-Yarn.

Regarding Zookeper, in YARN mode there is only one JobManager running, do we still need leader

If the ApplicationMaster goes down (where JM runs) it is restarted by Yarn RM and while restarting,
Flink AM will bring back previous running containers.  So, where does Zookeeper sit in this


From: Aljoscha Krettek <aljoscha@apache.org<mailto:aljoscha@apache.org>>
Date: Wednesday, May 3, 2017 at 2:05 AM
To: "Jain, Ankit" <ankit.jain@here.com<mailto:ankit.jain@here.com>>
Cc: "user@flink.apache.org<mailto:user@flink.apache.org>" <user@flink.apache.org<mailto:user@flink.apache.org>>,
Till Rohrmann <trohrmann@apache.org<mailto:trohrmann@apache.org>>
Subject: Re: High Availability on Yarn

As a first comment, the work mentioned in the FLIP-6 doc you linked is still work-in-progress.
You cannot use these abstractions yet without going into the code and setting up a cluster
“by hand”.

The documentation for one-step deployment of a Job to YARN is available here: https://ci.apache.org/projects/flink/flink-docs-release-1.2/setup/yarn_setup.html#run-a-single-flink-job-on-yarn<https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fci.apache.org%2Fprojects%2Fflink%2Fflink-docs-release-1.2%2Fsetup%2Fyarn_setup.html%23run-a-single-flink-job-on-yarn&data=01%7C01%7C%7Cfb13823970ba4476ebbf08d49203846c%7C6d4034cd72254f72b85391feaea64919%7C1&sdata=fzBiQtv7MR2%2Fehg6GepwPa1uWxpqEgPJakto2B8k0Zk%3D&reserved=0>

Regarding your third question, ZooKeeper is mostly used for discovery and leader election.
That is, JobManagers use it to decide who is the main JM and who are standby JMs. TaskManagers
use it to discover the leading JobManager that they should connect to.

I’m also cc’ing Till, who should know this stuff better and can maybe explain it in a
bit more detail.

On 1. May 2017, at 18:59, Jain, Ankit <ankit.jain@here.com<mailto:ankit.jain@here.com>>

Hi fellow users,
We are trying to straighten out high availability story for flink.

Our setup includes a long running EMR cluster, job submission is a two-step process – 1)
Flink cluster is first created using flink yarn client on the EMR cluster already running
2) Flink job is submitted.

I also saw references that with 1.2, these two steps have been combined into 1 – is that
change in FlinkYarnSessionCli.java? Can somebody point to documentation please?

W/o worrying about Yarn RM (not Flink Yarn RM that seems to be newly introduced) failure for
now, I want to understand first how task manager & job manager failures are handled.

My questions-
1)       https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=65147077<https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fcwiki.apache.org%2Fconfluence%2Fpages%2Fviewpage.action%3FpageId%3D65147077&data=01%7C01%7C%7Cfb13823970ba4476ebbf08d49203846c%7C6d4034cd72254f72b85391feaea64919%7C1&sdata=29Se7mWQZ09ukF3rkQNmSRPXY4RkA8RCNO4ec4Glj8I%3D&reserved=0>
suggests a new RM has been added and now there is one JobManager for each job. Since Yarn
RM will now talk to Flink RM( instead of JobManager previously), will Yarn automatically restart
failing Flink RM?
2)       Is there any documentation on behavior of new Flink RM that will come up? How will
previously running JobManagers & TaskManagers find out about new RM?
3)       https://ci.apache.org/projects/flink/flink-docs-release-1.3/setup/jobmanager_high_availability.html#configuration<https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fci.apache.org%2Fprojects%2Fflink%2Fflink-docs-release-1.3%2Fsetup%2Fjobmanager_high_availability.html%23configuration&data=01%7C01%7C%7Cfb13823970ba4476ebbf08d49203846c%7C6d4034cd72254f72b85391feaea64919%7C1&sdata=nYTYWaaWA4T1D7EwvL%2B7mwhrVcqn6xTzCv8SS6x%2FqLM%3D&reserved=0>
requires configuring Zookeeper even for Yarn – Is this needed for handling Task Manager
failures or JM or both? Will Yarn not take care of JM failures?

It may sound like I am little confused between role of Yarn and Flink components– who has
the most burden of HA? Documentation in current state is lacking clarity – I know it is
still evolving.

Please let me know if somebody can help clear the confusion.


View raw message