You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@spark.apache.org by "张志强(旺轩)" <zz...@alibaba-inc.com> on 2015/12/15 10:23:27 UTC

spark with label nodes in yarn

Hi all,

 

Has anyone tried label based scheduling via spark on yarn? I've tried that,
it didn't work, spark 1.4.1 + apache hadoop 2.6.0

 

Any feedbacks are welcome.

 

Thanks

Allen

Re: spark with label nodes in yarn

Posted by Ted Yu <yu...@gmail.com>.
SPARK-6470 was integrated to 1.5.0 release.

Please use 1.5.0 or newer release.

SPARK-7173 <https://issues.apache.org/jira/browse/SPARK-7173> adds support
for setting "spark.yarn.am.nodeLabelExpression"

Cheers

On Tue, Dec 15, 2015 at 1:55 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

> Hi Ted,
>
>
>
> Thanks for your quick response, but I think the link you gave it to me is
> more advanced feature.
>
> Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470)
>
>
> And I just tried for this feature with spark 1.5.0, what happened to me
> was I was blocked to get the YARN containers by setting
> spark.yarn.executor.nodeLabelExpression property. My question,
> https://issues.apache.org/jira/browse/SPARK-7173 will fix this?
>
>
>
> Thanks
>
> Allen
>
>
>
>
>
> *发件人:* Ted Yu [mailto:yuzhihong@gmail.com]
> *发送时间:* 2015年12月15日 17:39
> *收件人:* 张志强(旺轩)
> *抄送:* dev@spark.apache.org
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Please take a look at:
>
> https://issues.apache.org/jira/browse/SPARK-7173
>
>
>
> Cheers
>
>
> On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi all,
>
>
>
> Has anyone tried label based scheduling via spark on yarn? I’ve tried
> that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0
>
>
>
> Any feedbacks are welcome.
>
>
>
> Thanks
>
> Allen
>
>

Re: spark with label nodes in yarn

Posted by "张志强(旺轩)" <zz...@alibaba-inc.com>.
Oops...

 

I do use spark 1.5.0 and apache hadoop 2.6.0 (spark 1.4.1 + apache hadoop 2.6.0 is a typo), sorry

 

Thanks,

Allen

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 22:59
收件人: 张志强(旺轩)
抄送: Saisai Shao; dev
主题: Re: spark with label nodes in yarn

 

Please upgrade to Spark 1.5.x

 

1.4.1 didn't support node label feature.

 

Cheers

 

On Tue, Dec 15, 2015 at 2:20 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi SaiSai,

 

OK, it make sense to me , what I need is just to schedule the executors, AND I leave one nodemanager at least with no any labels.

 

It’s weird to me that YARN page shows my application is running, but actually it is still waiting for its executor

 

See the attached.

 

Thanks,

Allen

 

发件人: Saisai Shao [mailto:sai.sai.shao@gmail.com] 
发送时间: 2015年12月15日 18:07
收件人: 张志强(旺轩)
抄送: Ted Yu; dev

主题: Re: spark with label nodes in yarn

 

SPARK-6470 only supports node label expression for executors.

SPARK-7173 supports node label expression for AM (will be in 1.6).

 

If you want to schedule your whole application through label expression, you have to configure both am and executor label expression. If you only want to schedule executors through label expression, the executor configuration is enough, but you have to make sure your cluster has some nodes with no label.

 

You can refer to this document (http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html).

 

Thanks

Saisai

 

 

On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi Ted,

 

Thanks for your quick response, but I think the link you gave it to me is more advanced feature.

Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470) 

And I just tried for this feature with spark 1.5.0, what happened to me was I was blocked to get the YARN containers by setting spark.yarn.executor.nodeLabelExpression property. My question, https://issues.apache.org/jira/browse/SPARK-7173 will fix this?

 

Thanks

Allen

 

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 17:39
收件人: 张志强(旺轩)
抄送: dev@spark.apache.org
主题: Re: spark with label nodes in yarn

 

Please take a look at:

https://issues.apache.org/jira/browse/SPARK-7173

 

Cheers


On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi all,

 

Has anyone tried label based scheduling via spark on yarn? I’ve tried that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0

 

Any feedbacks are welcome.

 

Thanks

Allen

Re: spark with label nodes in yarn

Posted by Saisai Shao <sa...@gmail.com>.
Yes, of course, capacity scheduler also needs to be configured.

On Wed, Dec 16, 2015 at 10:41 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

> one more question , do I have to configure label for my capacity
> scheduler? is this mandatory?
>
>
>
> *发件人:* AllenZ [mailto:zzq98736@alibaba-inc.com]
> *发送时间:* 2015年12月16日 9:21
> *收件人:* 'Ted Yu'
> *抄送:* 'Saisai Shao'; 'dev'
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Oops...
>
>
>
> I do use spark 1.5.0 and apache hadoop 2.6.0 (spark 1.4.1 + apache hadoop
> 2.6.0 is a typo), sorry
>
>
>
> Thanks,
>
> Allen
>
>
>
> *发件人:* Ted Yu [mailto:yuzhihong@gmail.com]
> *发送时间:* 2015年12月15日 22:59
> *收件人:* 张志强(旺轩)
> *抄送:* Saisai Shao; dev
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Please upgrade to Spark 1.5.x
>
>
>
> 1.4.1 didn't support node label feature.
>
>
>
> Cheers
>
>
>
> On Tue, Dec 15, 2015 at 2:20 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi SaiSai,
>
>
>
> OK, it make sense to me , what I need is just to schedule the executors,
> AND I leave one nodemanager at least with no any labels.
>
>
>
> It’s weird to me that YARN page shows my application is running, but
> actually it is still waiting for its executor
>
>
>
> See the attached.
>
>
>
> Thanks,
>
> Allen
>
>
>
> *发件人:* Saisai Shao [mailto:sai.sai.shao@gmail.com]
> *发送时间:* 2015年12月15日 18:07
> *收件人:* 张志强(旺轩)
> *抄送:* Ted Yu; dev
>
> *主题:* Re: spark with label nodes in yarn
>
>
>
> SPARK-6470 only supports node label expression for executors.
>
> SPARK-7173 supports node label expression for AM (will be in 1.6).
>
>
>
> If you want to schedule your whole application through label expression,
> you have to configure both am and executor label expression. If you only
> want to schedule executors through label expression, the executor
> configuration is enough, but you have to make sure your cluster has some
> nodes with no label.
>
>
>
> You can refer to this document (
> http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html
> ).
>
>
>
> Thanks
>
> Saisai
>
>
>
>
>
> On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi Ted,
>
>
>
> Thanks for your quick response, but I think the link you gave it to me is
> more advanced feature.
>
> Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470)
>
>
> And I just tried for this feature with spark 1.5.0, what happened to me
> was I was blocked to get the YARN containers by setting
> spark.yarn.executor.nodeLabelExpression property. My question,
> https://issues.apache.org/jira/browse/SPARK-7173 will fix this?
>
>
>
> Thanks
>
> Allen
>
>
>
>
>
> *发件人:* Ted Yu [mailto:yuzhihong@gmail.com]
> *发送时间:* 2015年12月15日 17:39
> *收件人:* 张志强(旺轩)
> *抄送:* dev@spark.apache.org
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Please take a look at:
>
> https://issues.apache.org/jira/browse/SPARK-7173
>
>
>
> Cheers
>
>
> On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi all,
>
>
>
> Has anyone tried label based scheduling via spark on yarn? I’ve tried
> that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0
>
>
>
> Any feedbacks are welcome.
>
>
>
> Thanks
>
> Allen
>
>
>
>
>

Re: spark with label nodes in yarn

Posted by "张志强(旺轩)" <zz...@alibaba-inc.com>.
one more question , do I have to configure label for my capacity scheduler? is this mandatory?

 

发件人: AllenZ [mailto:zzq98736@alibaba-inc.com] 
发送时间: 2015年12月16日 9:21
收件人: 'Ted Yu'
抄送: 'Saisai Shao'; 'dev'
主题: Re: spark with label nodes in yarn

 

Oops...

 

I do use spark 1.5.0 and apache hadoop 2.6.0 (spark 1.4.1 + apache hadoop 2.6.0 is a typo), sorry

 

Thanks,

Allen

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 22:59
收件人: 张志强(旺轩)
抄送: Saisai Shao; dev
主题: Re: spark with label nodes in yarn

 

Please upgrade to Spark 1.5.x

 

1.4.1 didn't support node label feature.

 

Cheers

 

On Tue, Dec 15, 2015 at 2:20 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi SaiSai,

 

OK, it make sense to me , what I need is just to schedule the executors, AND I leave one nodemanager at least with no any labels.

 

It’s weird to me that YARN page shows my application is running, but actually it is still waiting for its executor

 

See the attached.

 

Thanks,

Allen

 

发件人: Saisai Shao [mailto:sai.sai.shao@gmail.com] 
发送时间: 2015年12月15日 18:07
收件人: 张志强(旺轩)
抄送: Ted Yu; dev

主题: Re: spark with label nodes in yarn

 

SPARK-6470 only supports node label expression for executors.

SPARK-7173 supports node label expression for AM (will be in 1.6).

 

If you want to schedule your whole application through label expression, you have to configure both am and executor label expression. If you only want to schedule executors through label expression, the executor configuration is enough, but you have to make sure your cluster has some nodes with no label.

 

You can refer to this document (http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html).

 

Thanks

Saisai

 

 

On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi Ted,

 

Thanks for your quick response, but I think the link you gave it to me is more advanced feature.

Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470) 

And I just tried for this feature with spark 1.5.0, what happened to me was I was blocked to get the YARN containers by setting spark.yarn.executor.nodeLabelExpression property. My question, https://issues.apache.org/jira/browse/SPARK-7173 will fix this?

 

Thanks

Allen

 

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 17:39
收件人: 张志强(旺轩)
抄送: dev@spark.apache.org
主题: Re: spark with label nodes in yarn

 

Please take a look at:

https://issues.apache.org/jira/browse/SPARK-7173

 

Cheers


On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi all,

 

Has anyone tried label based scheduling via spark on yarn? I’ve tried that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0

 

Any feedbacks are welcome.

 

Thanks

Allen

Re: spark with label nodes in yarn

Posted by Ted Yu <yu...@gmail.com>.
Please upgrade to Spark 1.5.x

1.4.1 didn't support node label feature.

Cheers

On Tue, Dec 15, 2015 at 2:20 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

> Hi SaiSai,
>
>
>
> OK, it make sense to me , what I need is just to schedule the executors,
> AND I leave one nodemanager at least with no any labels.
>
>
>
> It’s weird to me that YARN page shows my application is running, but
> actually it is still waiting for its executor
>
>
>
> See the attached.
>
>
>
> Thanks,
>
> Allen
>
>
>
> *发件人:* Saisai Shao [mailto:sai.sai.shao@gmail.com]
> *发送时间:* 2015年12月15日 18:07
> *收件人:* 张志强(旺轩)
> *抄送:* Ted Yu; dev
> *主题:* Re: spark with label nodes in yarn
>
>
>
> SPARK-6470 only supports node label expression for executors.
>
> SPARK-7173 supports node label expression for AM (will be in 1.6).
>
>
>
> If you want to schedule your whole application through label expression,
> you have to configure both am and executor label expression. If you only
> want to schedule executors through label expression, the executor
> configuration is enough, but you have to make sure your cluster has some
> nodes with no label.
>
>
>
> You can refer to this document (
> http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html
> ).
>
>
>
> Thanks
>
> Saisai
>
>
>
>
>
> On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi Ted,
>
>
>
> Thanks for your quick response, but I think the link you gave it to me is
> more advanced feature.
>
> Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470)
>
>
> And I just tried for this feature with spark 1.5.0, what happened to me
> was I was blocked to get the YARN containers by setting
> spark.yarn.executor.nodeLabelExpression property. My question,
> https://issues.apache.org/jira/browse/SPARK-7173 will fix this?
>
>
>
> Thanks
>
> Allen
>
>
>
>
>
> *发件人:* Ted Yu [mailto:yuzhihong@gmail.com]
> *发送时间:* 2015年12月15日 17:39
> *收件人:* 张志强(旺轩)
> *抄送:* dev@spark.apache.org
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Please take a look at:
>
> https://issues.apache.org/jira/browse/SPARK-7173
>
>
>
> Cheers
>
>
> On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi all,
>
>
>
> Has anyone tried label based scheduling via spark on yarn? I’ve tried
> that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0
>
>
>
> Any feedbacks are welcome.
>
>
>
> Thanks
>
> Allen
>
>
>

Re: spark with label nodes in yarn

Posted by "张志强(旺轩)" <zz...@alibaba-inc.com>.
Hi SaiSai,

 

OK, it make sense to me , what I need is just to schedule the executors, AND I leave one nodemanager at least with no any labels.

 

It’s weird to me that YARN page shows my application is running, but actually it is still waiting for its executor

 

See the attached.

 

Thanks,

Allen

 

发件人: Saisai Shao [mailto:sai.sai.shao@gmail.com] 
发送时间: 2015年12月15日 18:07
收件人: 张志强(旺轩)
抄送: Ted Yu; dev
主题: Re: spark with label nodes in yarn

 

SPARK-6470 only supports node label expression for executors.

SPARK-7173 supports node label expression for AM (will be in 1.6).

 

If you want to schedule your whole application through label expression, you have to configure both am and executor label expression. If you only want to schedule executors through label expression, the executor configuration is enough, but you have to make sure your cluster has some nodes with no label.

 

You can refer to this document (http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html).

 

Thanks

Saisai

 

 

On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi Ted,

 

Thanks for your quick response, but I think the link you gave it to me is more advanced feature.

Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470) 

And I just tried for this feature with spark 1.5.0, what happened to me was I was blocked to get the YARN containers by setting spark.yarn.executor.nodeLabelExpression property. My question, https://issues.apache.org/jira/browse/SPARK-7173 will fix this?

 

Thanks

Allen

 

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 17:39
收件人: 张志强(旺轩)
抄送: dev@spark.apache.org
主题: Re: spark with label nodes in yarn

 

Please take a look at:

https://issues.apache.org/jira/browse/SPARK-7173

 

Cheers


On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi all,

 

Has anyone tried label based scheduling via spark on yarn? I’ve tried that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0

 

Any feedbacks are welcome.

 

Thanks

Allen

Re: spark with label nodes in yarn

Posted by Saisai Shao <sa...@gmail.com>.
SPARK-6470 only supports node label expression for executors.
SPARK-7173 supports node label expression for AM (will be in 1.6).

If you want to schedule your whole application through label expression,
you have to configure both am and executor label expression. If you only
want to schedule executors through label expression, the executor
configuration is enough, but you have to make sure your cluster has some
nodes with no label.

You can refer to this document (
http://docs.hortonworks.com/HDPDocuments/HDP2/HDP-2.3.2/bk_yarn_resource_mgt/content/configuring_node_labels.html
).

Thanks
Saisai


On Tue, Dec 15, 2015 at 5:55 PM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

> Hi Ted,
>
>
>
> Thanks for your quick response, but I think the link you gave it to me is
> more advanced feature.
>
> Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470)
>
>
> And I just tried for this feature with spark 1.5.0, what happened to me
> was I was blocked to get the YARN containers by setting
> spark.yarn.executor.nodeLabelExpression property. My question,
> https://issues.apache.org/jira/browse/SPARK-7173 will fix this?
>
>
>
> Thanks
>
> Allen
>
>
>
>
>
> *发件人:* Ted Yu [mailto:yuzhihong@gmail.com]
> *发送时间:* 2015年12月15日 17:39
> *收件人:* 张志强(旺轩)
> *抄送:* dev@spark.apache.org
> *主题:* Re: spark with label nodes in yarn
>
>
>
> Please take a look at:
>
> https://issues.apache.org/jira/browse/SPARK-7173
>
>
>
> Cheers
>
>
> On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
>
> Hi all,
>
>
>
> Has anyone tried label based scheduling via spark on yarn? I’ve tried
> that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0
>
>
>
> Any feedbacks are welcome.
>
>
>
> Thanks
>
> Allen
>
>

Re: spark with label nodes in yarn

Posted by "张志强(旺轩)" <zz...@alibaba-inc.com>.
Hi Ted,

 

Thanks for your quick response, but I think the link you gave it to me is more advanced feature.

Yes, I noticed SPARK-6470(https://issues.apache.org/jira/browse/SPARK-6470) 

And I just tried for this feature with spark 1.5.0, what happened to me was I was blocked to get the YARN containers by setting spark.yarn.executor.nodeLabelExpression property. My question, https://issues.apache.org/jira/browse/SPARK-7173 will fix this?

 

Thanks

Allen

 

 

发件人: Ted Yu [mailto:yuzhihong@gmail.com] 
发送时间: 2015年12月15日 17:39
收件人: 张志强(旺轩)
抄送: dev@spark.apache.org
主题: Re: spark with label nodes in yarn

 

Please take a look at:

https://issues.apache.org/jira/browse/SPARK-7173

 

Cheers


On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:

Hi all,

 

Has anyone tried label based scheduling via spark on yarn? I’ve tried that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0

 

Any feedbacks are welcome.

 

Thanks

Allen

Re: spark with label nodes in yarn

Posted by Ted Yu <yu...@gmail.com>.
Please take a look at:
https://issues.apache.org/jira/browse/SPARK-7173

Cheers

> On Dec 15, 2015, at 1:23 AM, 张志强(旺轩) <zz...@alibaba-inc.com> wrote:
> 
> Hi all,
>  
> Has anyone tried label based scheduling via spark on yarn? I’ve tried that, it didn’t work, spark 1.4.1 + apache hadoop 2.6.0
>  
> Any feedbacks are welcome.
>  
> Thanks
> Allen