Re: Influence about removing RequiresInstanceClassLoading from AbstractHadoopProcessor processor

[abellnotring]

Hi,jeff,

    There is no Kerberos authentication  in my HADOOP clusters , but I find UGI is initialized with an ExtendConfiguration(extend from hadoop Configruation) when those processor's instance was first scheduled .I would use nifi to connect different HADOOP cluster, will it  run into any issues?(I’m running tests for this)

Thanks 

 By Hai Luo

On 11/13/2019 07:01，Jeff<jtsw...@gmail.com> wrote：

If you remove the @RequiresInstanceClassloading, the UserGroupInformation class from Hadoop (hadoop-common, if I remember correctly) will be shared across all instances that come from a particular NAR (such as PutHDFS, ListHDFS, FetchHDFS, etc, from nifi-hadoop-nar-x.y.z.nar).  If you are using Kerberos in those processors and configured different principals across the various processors, you could run into issues when the processors attempt to acquire new TGTs, most likely the first time a relogin is attempted.  UGI has some static state and @RequiresInstanceClassloading makes sure each instance of a processor with that annotation has its own classloader to keep that kind of state from being shared across instances.

On Mon, Nov 11, 2019 at 9:41 PM abellnotring <abellnotr...@sina.com> wrote:

Hi,Peter & All

     I’m using kylo to manage the nifi flow(called feed in kylo),and there are 4200 instances(600+ instances extended from  AbstractHadoopProcessor) in my nifi canvas,The NIFI Non-Heap Memory has increased more than  6GB after some days running ,which is extremely abnormal . I have analyzed the class loaded into Compressed Class Space ,and found most of the CCS was used by classes related by AbstractHadoopProcessor.

   So  I think removing RequiresInstanceClassLoading from AbstractHadoopProcessor processor may be a Solution for reducing the CCS used.

   Do you have any ideas for this ?

    

    Thanks

    By Hai Luo

On 11/12/2019 02:17，Shawn Weeks<swe...@weeksconsulting.us> wrote：

I’m assuming your talking about the snappy problem. If you use compress content prior to puthdfs you can compress with Snappy as it uses the Java Native Snappy Lib. The HDFS processors are limited to the actual Hadoop Libraries so they’d have to change from Native to get around this. I’m pretty sure we need instance loading to handle the other issues mentioned.

 

Thanks

Shawn

 

From: Joe Witt <joe.w...@gmail.com>
Reply-To: "users@nifi.apache.org" <users@nifi.apache.org>
Date: Monday, November 11, 2019 at 8:56 AM
To: "users@nifi.apache.org" <users@nifi.apache.org>
Subject: Re: Influence about removing RequiresInstanceClassLoading from AbstractHadoopProcessor processor

 

Peter

 

The most common challenge is if two isolated instances both want to use a native lib.  No two native libs with the same name can be in the same jvm.  We need to solve that for sure.

 

Thanks

 

On Mon, Nov 11, 2019 at 9:53 AM Peter Turcsanyi <turcsa...@apache.org> wrote:

Hi Hai Luo,

 

@RequiresInstanceClassLoading makes possible to configure separate / isolated "Additional Classpath Resources" settings on your HDFS processors (eg. S3 storage driver on one of your PutHDFS and Azure Blob on the other).

 

Is there any specific reason / use case why you are considering to remove it?

 

Regards,

Peter Turcsanyi

 

On Mon, Nov 11, 2019 at 3:30 PM abellnotring <abellnotr...@sina.com> wrote:

Hi,all

     I’m considering removing the RequiresInstanceClassLoading annotation from class AbstractHadoopProcessor,

     Does anybody know the potential Influence?

 

    Thanks

    By Hai Luo