Re: does flink support yarn/hdfs HA?

Posted by Till Rohrmann on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/does-flink-support-yarn-hdfs-HA-tp16448p16507.html

Quick question,

did you configure the application.classpath in your Yarn configuration? The reason I’m asking is that the AbstractYarnDescriptor sets up the Yarn classpath and if application.classpath is not configured, then it adds the HADOOP_CONF_DIR path to the classpath. If you check the classpath which you’ve sent me, then you’ll see that the expanded HADOOP_CONF_DIR is not part of it. That either means that HADOOP_CONF_DIR has not been set for the user starting the Yarn session or that there is maybe the application.classpath configured.

Cheers,
Till

​

On Wed, Nov 1, 2017 at 9:58 AM, 邓俊华 <[hidden email]> wrote:

      hi,

      

     When I use "new YarnConfiguration(HadoopUtils.getHadoopConfiguration())" in flink-1.3.2,I can't get yarn-site.xml property. So I get it from HadoopUtils.getHadoopConfiguration().

     The attached file is the bin/yarn-session.sh classpath. And the HADOOP_CONF_DIR/YARN_CONF_DIR as below, I am sure there are *-site.xml.

     

    

     

------------------------------------------------------------------

发件人：Till Rohrmann <[hidden email]>

发送时间：2017年11月1日(星期三) 16:32

收件人：邓俊华 <[hidden email]>

抄　送：Stephan Ewen <[hidden email]>; user <[hidden email]>; 黄亮 <[hidden email]>

主　题：Re: does flink support yarn/hdfs HA?

Hi,

thanks for looking into the problem. I actually opened a PR [1] for fixing this but it does not include the changes to HadoopUtils and, thus, does not solve the problem entirely.

I’m wondering why you have to add explicitly the yarn-site.xml. The YarnConfiguration should add the yarn-site.xml as a default resource which is extracted from the classpath. Could you share the classpath of the launched YarnJobManager (you find it a the top of the YarnJobManager logs). Also it would be interesting to see what HADOOP_CONF_DIR is actually pointing to. The reason I’m asking is that HADOOP_CONF_DIR should be included in the classpath of the launched application master.

[1] https://github.com/apache/flink/pull/4926

Cheers,
Till

On Tue, Oct 31, 2017 at 7:11 AM, 邓俊华 <[hidden email]> wrote:

    hi Till,

    The issue can be fixed by the below solution.

     

    In org.apache.flink.yarnYarnApplicationMasterRunner.java:

        replace "final YarnConfiguration yarnConfig = new YarnConfiguration();" with "final YarnConfiguration yarnConfig = new YarnConfiguration(HadoopUtils.getHadoopConfiguration());"

   In org.apache.flink.runtime.util.HadoopUtils.java:

        add the below code in getHadoopConfiguration()

             if (new File(possibleHadoopConfPath + "/yarn-site.xml").exists()) {
      retConf.addResource(new org.apache.hadoop.fs.Path(possibleHadoopConfPath + "/yarn-site.xml"));

      if (LOG.isDebugEnabled()) {
       LOG.debug("Adding " + possibleHadoopConfPath + "/yarn-site.xml to hadoop configuration");
      }
     }

     I am using flink-1.3.2 and I also find it happened in flink-1.4-snapshot. Hope it won't happened in new realease.

      

------------------------------------------------------------

发件人：邓俊华 <[hidden email]>

发送时间：2017年10月31日(星期二) 11:16

收件人：Till Rohrmann <[hidden email]>

抄　送：Stephan Ewen <[hidden email]>; user <[hidden email]>; 黄亮 <[hidden email]>

主　题：回复：does flink support yarn/hdfs HA?

      hi Till,

     

     I have a test with the solution you provide. But it didn't fix the issue.

    1>     There is no 'org.apache.flink.runtime.util.HadoopUtils' in flink-1.3.2,it's in flink-1.4.0-snapshot.

              It can't change to  final YarnConfiguration yarnConfig = new YarnConfiguration(org.apache.flink.runtime.util.HadoopUtils.HadoopUtils.getHadoopConfiguration(flinkConfiguration)); to fix the issue.     

    2>     There is another api ''org.apache.flink.api.java.hadoop.mapred.utils.HadoopUtils" to get hdfs-site.xml and core-site.xml. So I changed to as below:

    final YarnConfiguration yarnConfig = new YarnConfiguration(org.apache.flink.api.java.hadoop.mapred.utils.HadoopUtils.getHadoopConfiguration());

 It truelly can read hdfs/core file,but it can't read yarn-site.xml now.

    So can I think flink doesn't support yarn/hdfs HA for now?  It support  the default configuration for yarn and can't read the yarn conf with custom configuration?

    

2017-10-31 11:14:12,860 DEBUG org.apache.hadoop.ipc.Client                                  - IPC Client <a href="tel:(209)%20645-0263" target="_blank">(2096450263) connection to spark2/174.18.4.205:9000 from admin: stopped, remaining connections 12017-10-31 11:14:13,259 INFO  org.apache.flink.yarn.YarnJobManager                          - Attempting to recover all jobs.2017-10-31 11:14:13,260 DEBUG org.apache.flink.runtime.jobmanager.ZooKeeperSubmittedJobGraphStore  - Retrieving all stored job ids from ZooKeeper under flink/application_1508845594022_0020/jobgraphs.2017-10-31 11:14:13,261 DEBUG org.apache.zookeeper.ClientCnxn                               - Reading reply sessionid:0x15cf415d94c00ee, packet:: clientPath:null serverPath:null finished:false header:: 55,3  replyHeader:: 55,8592068434,0  request:: '/flink/application_1508845594022_0020/jobgraphs,F  response:: s{8592068425,8592068425,1509419643087,1509419643087,0,0,0,0,0,0,8592068425} 2017-10-31 11:14:13,263 DEBUG org.apache.zookeeper.ClientCnxn                               - Reading reply sessionid:0x15cf415d94c00ee, packet:: clientPath:null serverPath:null finished:false header:: 56,12  replyHeader:: 56,8592068434,0  request:: '/flink/application_1508845594022_0020/jobgraphs,F  response:: v{},s{8592068425,8592068425,1509419643087,1509419643087,0,0,0,0,0,0,8592068425} 2017-10-31 11:14:13,264 INFO  org.apache.flink.yarn.YarnJobManager                          - There are no jobs to recover.2017-10-31 11:14:13,487 INFO  org.apache.hadoop.ipc.Client                                  - Retrying connect to server: 0.0.0.0/0.0.0.0:8030. Already tried 9 time(s); retry policy is RetryUpToMaximumCountWithFixedSleep(maxRetries=10, sleepTime=1000 MILLISECONDS)2017-10-31 11:14:13,489 DEBUG org.apache.hadoop.ipc.Client                                  - closing ipc connection to 0.0.0.0/0.0.0.0:8030: Connection refusedjava.net.ConnectException: Connection refused	at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)	at sun.nio.ch.SocketChannelImpl.finishConnect(SocketChannelImpl.java:744)	at org.apache.hadoop.net.SocketIOWithTimeout.connect(SocketIOWithTimeout.java:206)	at org.apache.hadoop.net.NetUtils.connect(NetUtils.java:531)	at org.apache.hadoop.net.NetUtils.connect(NetUtils.java:495)	at org.apache.hadoop.ipc.Client$Connection.setupConnection(Client.java:614)	at org.apache.hadoop.ipc.Client$Connection.setupIOstreams(Client.java:712)	at org.apache.hadoop.ipc.Client$Connection.access$2900(Client.java:375)	at org.apache.hadoop.ipc.Client.getConnection(Client.java:1528)	at org.apache.hadoop.ipc.Client.call(Client.java:1451)	at org.apache.hadoop.ipc.Client.call(Client.java:1412)	at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.invoke(ProtobufRpcEngine.java:229)	at com.sun.proxy.$Proxy15.registerApplicationMaster(Unknown Source)	at org.apache.hadoop.yarn.api.impl.pb.client.ApplicationMasterProtocolPBClientImpl.registerApplicationMaster(ApplicationMasterProtocolPBClientImpl.java:107)

     

    

------------------------------------------------------------------

发件人：Till Rohrmann <[hidden email]>

发送时间：2017年10月27日(星期五) 20:24

收件人：邓俊华 <[hidden email]>

抄　送：Stephan Ewen <[hidden email]>; user <[hidden email]>

主　题：Re: does flink support yarn/hdfs HA?

Hi,

looking at the code, the YarnConfiguration only loads the core-site.xml and the yarn-site.xml and does not load the hdfs-site.xml. This seems to be on purpose and, thus, we should try to add it. I think the solution could be as easy as final YarnConfiguration yarnConfig = new YarnConfiguration(HadoopUtils.getHadoopConfiguration(flinkConfiguration));

Do you want to open a JIRA issue and a PR for this issue?

Cheers,
Till

On Fri, Oct 27, 2017 at 1:59 PM, 邓俊华 <[hidden email]> wrote:

     it pass the "final YarnConfiguration yarnConfig = new YarnConfiguration();" in https://github.com/apache/flink/blob/master/flink-yarn/src/main/java/org/apache/flink/yarn/YarnApplicationMasterRunner.java not the conf in https://github.com/apache/flink/blob/master/flink-filesystems/flink-hadoop-fs/src/main/java/org/apache/flink/runtime/util/HadoopUtils.java;

See

    

     

------------------------------------------------------------------

发件人：Stephan Ewen <[hidden email]>

发送时间：2017年10月27日(星期五) 19:50

收件人：邓俊华 <[hidden email]>

抄　送：user <[hidden email]>; Stephan Ewen <[hidden email]>; Till Rohrmann <[hidden email]>

主　题：Re: does flink support yarn/hdfs HA?

The it should probably work to apply a similar logic to the YarnConfiguration loading as to the Hadoop Filesystem configuration loading in https://github.com/apache/flink/blob/master/flink-filesystems/flink-hadoop-fs/src/main/java/org/apache/flink/runtime/util/HadoopUtils.java

On Fri, Oct 27, 2017 at 1:42 PM, 邓俊华 <[hidden email]> wrote:

Yes, every yarn cluster node has *-site.xml and  YARN_CONF_DIR and HADOOP_CONF_DIR be  configured  in every yarn cluster node.

------------------------------------------------------------------

发件人：Stephan Ewen <[hidden email]>

发送时间：2017年10月27日(星期五) 19:38

收件人：user <[hidden email]>

抄　送：邓俊华 <[hidden email]>; Till Rohrmann <[hidden email]>

主　题：Re: does flink support yarn/hdfs HA?

Hi!

The HDFS config resolution logic is in [1]. It should take the environment variable into account.

The problem (bug) seems to be that in the Yarn Utils, the configuration used for Yarn (AppMaster) loads purely from classpath, not from the environment variables.

And that classpath does not include the proper hdfs-site.xml.

Now, there are two ways I could imagine fixing that:

(1) When instantiating the Yarn Configuration, also add resources from the environment variable as in [1].

(2) The client (which puts together all the resources that should be in the classpath of the processes running on Yarn needs to add hadoop these config files to the resources.

Are the environment variables YARN_CONF_DIR and HADOOP_CONF_DIR typically defined on the nodes of a Yarn cluster?

Stephan

[1]: https://github.com/apache/flink/blob/master/flink-filesystems/flink-hadoop-fs/src/main/java/org/apache/flink/runtime/util/HadoopUtils.java

On Fri, Oct 27, 2017 at 12:38 PM, 邓俊华 <[hidden email]> wrote:

hi,

Now I run spark streming on yarn with yarn/HA. I want to migrate spark streaming to flink. But it can't work. It can't conncet to hdfs logical namenode url.

It always throws exception as fllow. I tried to trace the flink code and I found  there maybe  something wrong when read the hadoop conf.

As fllow, code the failoverProxyProvider is null , so it think  it's a No-HA case. The YarnConfiguration didn't read hdfs-site.xml to get "dfs.client.failover.proxy.provider.startdt" value,only yarn-site.xml and core-site.xml be read.

I have configured and  export YARN_CONF_DIR and HADOOP_CONF_DIR in /etc/profile and I am sure there are *-site.xml in this diretory. I  can  also get the value in env.

Can anybody  provide me some advice? 

 

2017-10-25 21:02:14,721 DEBUG org.apache.hadoop.hdfs.BlockReaderLocal                       - dfs.domain.socket.path = 
2017-10-25 21:02:14,755 ERROR org.apache.flink.yarn.YarnApplicationMasterRunner             - YARN Application Master initialization failed
java.lang.IllegalArgumentException: java.net.UnknownHostException: startdt
        at org.apache.hadoop.security.SecurityUtil.buildTokenService(SecurityUtil.java:378)
        at org.apache.hadoop.hdfs.NameNodeProxies.createNonHAProxy(NameNodeProxies.java:310)
        at org.apache.hadoop.hdfs.NameNodeProxies.createProxy(NameNodeProxies.java:176)
        at org.apache.hadoop.hdfs.DFSClient.<init>(DFSClient.java:678)
        at org.apache.hadoop.hdfs.DFSClient.<init>(DFSClient.java:619)
        at org.apache.hadoop.hdfs.DistributedFileSystem.initialize(DistributedFileSystem.java:149)
        at org.apache.hadoop.fs.FileSystem.createFileSystem(FileSystem.java:2669)
        at org.apache.hadoop.fs.FileSystem.access$200(FileSystem.java:94)
        at org.apache.hadoop.fs.FileSystem$Cache.getInternal(FileSystem.java:2703)
        at org.apache.hadoop.fs.FileSystem$Cache.get(FileSystem.java:2685)
        at org.apache.hadoop.fs.FileSystem.get(FileSystem.java:373)
        at org.apache.hadoop.fs.Path.getFileSystem(Path.java:295)
        at org.apache.flink.yarn.Utils.createTaskExecutorContext(Utils.java:385)
        at org.apache.flink.yarn.YarnApplicationMasterRunner.runApplicationMaster(YarnApplicationMasterRunner.java:324)
        at org.apache.flink.yarn.YarnApplicationMasterRunner$1.call(YarnApplicationMasterRunner.java:195)
        at org.apache.flink.yarn.YarnApplicationMasterRunner$1.call(YarnApplicationMasterRunner.java:192)
        at org.apache.flink.runtime.security.HadoopSecurityContext$1.run(HadoopSecurityContext.java:43)
        at java.security.AccessController.doPrivileged(Native Method)
        at javax.security.auth.Subject.doAs(Subject.java:415)
        at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:<a href="tel:1698" style="color:#000000;font-family:Tahoma,Arial;font-size:14.0px;font-style:normal;font-variant:normal;font-weight:normal;text-transform:none" target="_blank">1698)
        at org.apache.flink.runtime.security.HadoopSecurityContext.runSecured(HadoopSecurityContext.java:40)
        at org.apache.flink.yarn.YarnApplicationMasterRunner.run(YarnApplicationMasterRunner.java:192)
        at org.apache.flink.yarn.YarnApplicationMasterRunner.main(YarnApplicationMasterRunner.java:116)
Caused by: java.net.UnknownHostException: startdt
        ... 23 more