spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tathagata Das <t...@databricks.com>
Subject Re: Is it this a BUG?: Why Spark Flume Streaming job is not deploying the Receiver to the specified host?
Date Wed, 19 Aug 2015 00:37:56 GMT
I dont think there is a super clean way for doing this. Here is an idea.
Run a dummy job with large number of partitions/tasks, which will access
SparkEnv.get.blockManager().blockManagerId().host() and return it.
sc.makeRDD(1 to 100, 100).map { _ =>
SparkEnv.get.blockManager().blockManagerId().host()
}.collect().distinct()

But that said, I do recommend using the pull based model. With the push
based model it become realy hard to deal with scenarios where the whole
node (where the receiver is supposed to run) goes down and the receiver
cannot run anywhere else.


On Tue, Aug 18, 2015 at 5:25 PM, diplomatic Guru <diplomaticguru@gmail.com>
wrote:

> Thank you Tathagata for your response.  Yes, I'm using push model on Spark
> 1.2. For my scenario I do prefer the push model. Is this the case on the
> later version 1.4 too?
>
> I think I can find a workaround for this issue but only if I know how to
> obtain the worker(executor) ID. I can get the detail of the driver like
> this:
>
> *ss.ssc().env().blockManager().blockManagerId().host()*
>
> *But not sure how I could the executor Id from the driver.*
>
> *When the job is submitted, I can see that blockmanager being registered
> with the Driver and Executor IP address:*
>
>
>
>
>
>
>
>
> *15/08/18 23:31:40 INFO YarnClientSchedulerBackend: Registered executor:
> Actor[akka.tcp://sparkExecutor@05151113997207:41630/user/Executor#1210147506]
> with ID 115/08/18 23:31:40 INFO RackResolver: Resolved 05151113997207 to
> /0513_R-0050/RJ0515/08/18 23:31:41 INFO BlockManagerMasterActor:
> Registering block manager 05151113997207:56921 with 530.3 MB RAM,
> BlockManagerId(1, 05151113997207, 56921)The BlockManagerMasterActor appears
> to be doing the registering. Is there anyway I can access this from the
> SparkContext?Thanks.*
>
>
>
> On 18 August 2015 at 22:40, Tathagata Das <tdas@databricks.com> wrote:
>
>> Are you using the Flume polling stream or the older stream?
>>
>> Such problems of binding used to occur in the older push-based approach,
>> hence we built the polling stream (pull-based).
>>
>>
>> On Tue, Aug 18, 2015 at 4:45 AM, diplomatic Guru <
>> diplomaticguru@gmail.com> wrote:
>>
>>> I'm testing the Flume + Spark integration example (flume count).
>>>
>>> I'm deploying the job using yarn cluster mode.
>>>
>>> I first logged into the Yarn cluster, then submitted the job and passed
>>> in a specific worker node's IP to deploy the job. But when I checked the
>>> WebUI, it failed to bind to the specified IP because the receiver was
>>> deployed to a different host, not the one I asked it to. Do you know?
>>>
>>> For your information,  I've also tried passing the IP address used by
>>> the resource manager to find resources but no joy. But when I set the host
>>> to 'localhost' and deploy to the cluster it is binding a worker node that
>>> is selected by the resource manager.
>>>
>>>
>>>
>>
>

Mime
View raw message