azure - 如何通过动态资源分配来运行spark + cassandra + mesos (dcos)？

Question

通过 Marathon，我们在每个从属节点上运行 MesosExternalShuffleService。当我们在粗粒度模式下通过 dcos CLI 提交 Spark 作业而无需动态分配时，一切都会按预期工作。但是，当我们通过动态分配提交相同的作业时，它会失败。

16/12/08 19:20:42 ERROR OneForOneBlockFetcher: Failed while starting block fetches
java.lang.RuntimeException: java.lang.RuntimeException: Failed to open file:/tmp/blockmgr-d4df5df4-24c9-41a3-9f26-4c1aba096814/30/shuffle_0_0_0.index
at   org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.getSortBasedShuffleBlockData(ExternalShuffleBlockResolver.java:234)
...
at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:354)
...
Caused by: java.io.FileNotFoundException: /tmp/blockmgr-d4df5df4-24c9-41a3-9f26-4c1aba096814/30/shuffle_0_0_0.index (No such file or directory)

完整描述:

• 我们使用 Azure 门户将 Mesos (DCOS) 与 Marathon 一起安装。
• 通过我们安装的 Universe 软件包:Cassandra、Spark 和 Marathon-lb
• 我们在 Cassandra 中生成了测试数据。
• 在笔记本电脑上我安装了 dcos CLI

当我提交如下作业时，一切都按预期工作:

./dcos spark run --submit-args="--properties-file coarse-grained.conf --class portal.spark.cassandra.app.ProductModelPerNrOfAlerts http://marathon-lb-default.marathon.mesos:10018/jars/spark-cassandra-assembly-1.0.jar"
Run job succeeded. Submission id: driver-20161208185927-0043

cqlsh:sp> select count(*) from product_model_per_alerts_by_date ;

count
-------
476

粗粒度.conf:

spark.cassandra.connection.host 10.32.0.17
spark.serializer org.apache.spark.serializer.KryoSerializer
spark.executor.cores 1
spark.executor.memory 1g
spark.executor.instances 2
spark.submit.deployMode cluster
spark.cores.max 4

portal.spark.cassandra.app.ProductModelPerNrOfAlerts:

package portal.spark.cassandra.app

import org.apache.spark.sql.{SQLContext, SaveMode}
import org.apache.spark.{SparkConf, SparkContext}

object ProductModelPerNrOfAlerts {
  def main(args: Array[String]): Unit = {

     val conf = new SparkConf(true)
                    .setAppName("cassandraSpark-ProductModelPerNrOfAlerts")

     val sc = new SparkContext(conf)

     val sqlContext = new SQLContext(sc)

     import sqlContext.implicits._

     val df = sqlContext
             .read
             .format("org.apache.spark.sql.cassandra")
             .options(Map("table" -> "asset_history", "keyspace" -> "sp"))
            .load()
            .select("datestamp","product_model","nr_of_alerts")

     val dr = df
           .groupBy("datestamp","product_model")
           .avg("nr_of_alerts")
           .toDF("datestamp","product_model","nr_of_alerts")

     dr.write
          .mode(SaveMode.Overwrite)
          .format("org.apache.spark.sql.cassandra")
          .options(Map("table" -> "product_model_per_alerts_by_date", "keyspace" -> "sp"))
          .save()


     sc.stop()
 }
}

动态分配

通过 Marathon 我们运行 Mesos 外部随机播放服务:

{
  "id": "spark-mesos-external-shuffle-service-tt",
  "container": {
     "type": "DOCKER",
     "docker": {
        "image": "jpavt/mesos-spark-hadoop:mesos-external-shuffle-service-1.0.4-2.0.1",
        "network": "BRIDGE",
        "portMappings": [
           { "hostPort": 7337, "containerPort": 7337, "servicePort": 7337 }
         ],
       "forcePullImage":true,
       "volumes": [
         {
           "containerPath": "/tmp",
           "hostPath": "/tmp",
           "mode": "RW"
         }
       ]
     }
   },
   "instances": 9,
   "cpus": 0.2,
   "mem": 512,
   "constraints": [["hostname", "UNIQUE"]]
 }

jpavt/mesos-spark-hadoop 的 Dockerfile:mesos-external-shuffle-service-1.0.4-2.0.1:

FROM mesosphere/spark:1.0.4-2.0.1
WORKDIR /opt/spark/dist
ENTRYPOINT ["./bin/spark-class", "org.apache.spark.deploy.mesos.MesosExternalShuffleService"]

现在，当我提交动态分配作业时，它失败了:

./dcos spark run --submit-args="--properties-file dynamic-allocation.conf --class portal.spark.cassandra.app.ProductModelPerNrOfAlerts http://marathon-lb-default.marathon.mesos:10018/jars/spark-cassandra-assembly-1.0.jar"
 Run job succeeded. Submission id: driver-20161208191958-0047

select count(*) from product_model_per_alerts_by_date ;

count
-------
 5

动态分配.conf:

spark.cassandra.connection.host 10.32.0.17
spark.serializer org.apache.spark.serializer.KryoSerializer
spark.executor.cores 1
spark.executor.memory 1g
spark.submit.deployMode cluster
spark.cores.max 4

spark.shuffle.service.enabled true
spark.dynamicAllocation.enabled true
spark.dynamicAllocation.minExecutors 2
spark.dynamicAllocation.maxExecutors 5
spark.dynamicAllocation.cachedExecutorIdleTimeout 120s
spark.dynamicAllocation.schedulerBacklogTimeout 10s
spark.dynamicAllocation.sustainedSchedulerBacklogTimeout 20s
spark.mesos.executor.docker.volumes /tmp:/tmp:rw
spark.local.dir /tmp

来自 mesos 的日志:

16/12/08 19:20:42 INFO MemoryStore: Block broadcast_7_piece0 stored as bytes in memory (estimated size 18.0 KB, free 366.0 MB)
16/12/08 19:20:42 INFO TorrentBroadcast: Reading broadcast variable 7 took 21 ms
16/12/08 19:20:42 INFO MemoryStore: Block broadcast_7 stored as values in memory (estimated size 38.6 KB, free 366.0 MB)
16/12/08 19:20:42 INFO MapOutputTrackerWorker: Don't have map outputs for shuffle 0, fetching them
16/12/08 19:20:42 INFO MapOutputTrackerWorker: Doing the fetch; tracker endpoint = NettyRpcEndpointRef(spark://<a href="https://stackoverflow.com/cdn-cgi/l/email-protection" class="__cf_email__" data-cfemail="6d200c1d2218191d1819391f0c0e06081f2d5c5d435e5f435d4359" rel="noreferrer noopener nofollow">[email protected]</a>:45422)
16/12/08 19:20:42 INFO MapOutputTrackerWorker: Got the output locations
16/12/08 19:20:42 INFO ShuffleBlockFetcherIterator: Getting 4 non-empty blocks out of 58 blocks
16/12/08 19:20:42 INFO TransportClientFactory: Successfully created connection to /10.32.0.11:7337 after 2 ms (0 ms spent in bootstraps)
16/12/08 19:20:42 INFO ShuffleBlockFetcherIterator: Started 1 remote fetches in 13 ms
16/12/08 19:20:42 ERROR OneForOneBlockFetcher: Failed while starting block fetches java.lang.RuntimeException: java.lang.RuntimeException: Failed to open file: /tmp/blockmgr-d4df5df4-24c9-41a3-9f26-4c1aba096814/30/shuffle_0_0_0.index
at   org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.getSortBasedShuffleBlockData(ExternalShuffleBlockResolver.java:234)
...
at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:354)
...
 Caused by: java.io.FileNotFoundException: /tmp/blockmgr-d4df5df4-24c9-41a3-9f26-4c1aba096814/30/shuffle_0_0_0.index (No such file or directory)

来自马拉松 Spark-mesos-external-shuffle-service-tt 的日志:

...
16/12/08 19:20:29 INFO MesosExternalShuffleBlockHandler: Received registration request from app 704aec43-1aa3-4971-bb98-e892beeb2c45-0008-driver-20161208191958-0047 (remote address /10.32.0.4:49710, heartbeat timeout 120000 ms).
16/12/08 19:20:31 INFO ExternalShuffleBlockResolver: Registered executor AppExecId{appId=704aec43-1aa3-4971-bb98-e892beeb2c45-0008-driver-20161208191958-0047, execId=2} with ExecutorShuffleInfo{localDirs=[/tmp/blockmgr-14525ef0-22e9-49fb-8e81-dc84e5fba8b2], subDirsPerLocalDir=64, shuffleManager=org.apache.spark.shuffle.sort.SortShuffleManager}
16/12/08 19:20:38 ERROR TransportRequestHandler: Error while invoking RpcHandler#receive() on RPC id 8157825166903585542
java.lang.RuntimeException: Failed to open file: /tmp/blockmgr-14525ef0-22e9-49fb-8e81-dc84e5fba8b2/16/shuffle_0_55_0.index
at org.apache.spark.network.shuffle.ExternalShuffleBlockResolver.getSortBasedShuffleBlockData(ExternalShuffleBlockResolver.java:234)
...
at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:354)
Caused by: java.io.FileNotFoundException: /tmp/blockmgr-14525ef0-22e9-49fb-8e81-dc84e5fba8b2/16/shuffle_0_55_0.index (No such file or directory)
...

但文件存在于给定的从属盒上:

$ ls -l /tmp/blockmgr-14525ef0-22e9-49fb-8e81-dc84e5fba8b2/16/shuffle_0_55_0.index
-rw-r--r-- 1 root root 1608 Dec  8 19:20 /tmp/blockmgr-14525ef0-22e9-49fb-8e81-dc84e5fba8b2/16/shuffle_0_55_0.index


 stat shuffle_0_55_0.index 
  File: 'shuffle_0_55_0.index'
  Size: 1608        Blocks: 8          IO Block: 4096   regular file
  Device: 801h/2049d    Inode: 1805493     Links: 1
  Access: (0644/-rw-r--r--)  Uid: (    0/    root)   Gid: (    0/    root)
  Access: 2016-12-08 19:20:38.163188836 +0000
  Modify: 2016-12-08 19:20:38.163188836 +0000
  Change: 2016-12-08 19:20:38.163188836 +0000
  Birth: -

Answer 1

虽然我不熟悉 DCOS、Marathon 和 Azure，但我在 Mesos 和 Aurora 上使用 Docker 进行动态资源分配(Mesos external shuffle 服务)。

• 每个 Mesos Agent 节点都有自己的外部 Shuffle 服务(即，一个 Mesos Agent 对应一个外部 Shuffle 服务)？
• spark.local.dir 设置完全相同的字符串并指向相同的目录？不过，您用于随机播放服务的 spark.local.dir 是 /tmp，但我不知道 DCOS 设置。
• spark.local.dir 目录对于两者都是可读/可写的？如果 mesos Agent 和外部 shuffle 服务都由 docker 启动，则主机上的 spark.local.dir 必须安装到两个容器。

编辑

• 如果设置了SPARK_LOCAL_DIRS(mesos 或独立)环境变量，spark.local.dir 将被覆盖。