集群中的 Spark 错误：ModuleNotFoundError：没有名为“cst_utils”的模块

Question

我有一个 python 的 Spark 程序。程序的结构是这样的：

  cst_utils.py
  bn_utils.py
  ep_utils.py
  main.py

每个cst_utils.py、bn_utils.py、ep_utils.py都有一个名为Spark_Func(sc)的 function。 在 main 中，我创建了一个 Spark 上下文sc ，并将它发送到每个Spark_Func ，如下所示：

  import cst_utils as cu
  import bn_utils as bu
  import ep_utils as eu

  spark_conf = SparkConf().setAppName('app_name') \
    .setMaster("spark://x.x.x.x:7077") \
    .set('spark.executor.memory', "8g") \
    .set('spark.executor.cores', 4) \
    .set('spark.task.cpus', 2)

   sc = SparkContext(conf=spark_conf)

   cu.spark_func(sc)
   bu.spark_func(sc)
   eu.spark_func(sc) 

我用两个奴隶和一个主人配置 Spark 集群，他们都有 Ubuntu 20.04 操作系统。 我在 spark-env.sh 中设置了 Master IP并使 SSH 无密码，Master 节点无需身份验证即可访问每个 Slave 节点。 我在每个节点中运行这些命令：

主节点：

   ./start-master.sh

奴隶：

   ./start-worker.sh spark://x.x.x.x:7077

集群已创建，因为我可以在浏览器中使用此命令看到 SPARK UI：

  http://x.x.x.x:8080

但是当我想用这个命令运行程序时：

  /opt/spark/bin/spark-submit --master spark://x.x.x.x:7077 main.py

我收到此错误：

    22/02/16 16:39:20 INFO SparkContext: Starting job: count at /home/hs/Desktop/etl/cst_utils.py:442
    22/02/16 16:39:20 INFO DAGScheduler: Registering RDD 2 (reduceByKey at /home/hs/Desktop/etl/cst_utils.py:434) as input to shuffle 0
    22/02/16 16:39:20 INFO DAGScheduler: Got job 0 (count at /home/hs/Desktop/etl/cst_utils.py:442) with 1 output partitions
    22/02/16 16:39:20 INFO DAGScheduler: Final stage: ResultStage 1 (count at /home/hs/Desktop/etl/cst_utils.py:442)
    22/02/16 16:39:20 INFO DAGScheduler: Parents of final stage: List(ShuffleMapStage 0)
    22/02/16 16:39:20 INFO DAGScheduler: Missing parents: List(ShuffleMapStage 0)
    22/02/16 16:39:20 INFO DAGScheduler: Submitting ShuffleMapStage 0 (PairwiseRDD[2] at reduceByKey at /home/hs/Desktop/etl/cst_utils.py:434), which has no missing parents
    22/02/16 16:39:20 INFO MemoryStore: Block broadcast_0 stored as values in memory (estimated size 9.4 KiB, free 366.3 MiB)
    22/02/16 16:39:20 INFO MemoryStore: Block broadcast_0_piece0 stored as bytes in memory (estimated size 5.9 KiB, free 366.3 MiB)
    22/02/16 16:39:20 INFO BlockManagerInfo: Added broadcast_0_piece0 in memory on x.x.x.x:43875 (size: 5.9 KiB, free: 366.3 MiB)
    22/02/16 16:39:20 INFO SparkContext: Created broadcast 0 from broadcast at DAGScheduler.scala:1388
    22/02/16 16:39:20 INFO DAGScheduler: Submitting 1 missing tasks from ShuffleMapStage 0 (PairwiseRDD[2] at reduceByKey at /home/hs/Desktop/etl/cst_utils.py:434) (first 15 tasks are for partitions Vector(0))
    22/02/16 16:39:20 INFO TaskSchedulerImpl: Adding task set 0.0 with 1 tasks resource profile 0
    22/02/16 16:39:21 INFO CoarseGrainedSchedulerBackend$DriverEndpoint: Registered executor NettyRpcEndpointRef(spark-client://Executor) (z.z.z.z:39668) with ID 1,  ResourceProfileId 0
    22/02/16 16:39:21 INFO CoarseGrainedSchedulerBackend$DriverEndpoint: Registered executor NettyRpcEndpointRef(spark-client://Executor) (y.y.y.y:46330) with ID 0,  ResourceProfileId 0
    22/02/16 16:39:21 INFO BlockManagerMasterEndpoint: Registering block manager y.y.y.y:34159 with 4.1 GiB RAM, BlockManagerId(0, y.y.y.y, 34159, None)
    22/02/16 16:39:21 INFO BlockManagerMasterEndpoint: Registering block manager z.z.z.z:42231 with 4.1 GiB RAM, BlockManagerId(1, z.z.z.z, 42231, None)
    22/02/16 16:39:21 INFO TaskSetManager: Starting task 0.0 in stage 0.0 (TID 0) (y.y.y.y, executor 0, partition 0, PROCESS_LOCAL, 4481 bytes) taskResourceAssignments Map()
    22/02/16 16:39:21 INFO BlockManagerInfo: Added broadcast_0_piece0 in memory on y.y.y.y:34159 (size: 5.9 KiB, free: 4.1 GiB)
    22/02/16 16:39:22 WARN TaskSetManager: Lost task 0.0 in stage 0.0 (TID 0) (y.y.y.y executor 0): org.apache.spark.api.python.PythonException: Traceback (most recent call last):
    File "/opt/spark/python/lib/pyspark.zip/pyspark/worker.py", line 586, in main
func, profiler, deserializer, serializer = read_command(pickleSer, infile)
    File "/opt/spark/python/lib/pyspark.zip/pyspark/worker.py", line 69, in read_command
command = serializer._read_with_length(file)
    File "/opt/spark/python/lib/pyspark.zip/pyspark/serializers.py", line 160, in _read_with_length
    return self.loads(obj)
    File "/opt/spark/python/lib/pyspark.zip/pyspark/serializers.py", line 430, in loads
    return pickle.loads(obj, encoding=encoding)
    ModuleNotFoundError: No module named 'cst_utils'

程序路径与所有节点的路径相同，与SPARK路径相同。

事实上，当我以本地模式运行程序时，它运行没有任何问题。 但是，要在本地运行，我在 SPARK CONTEXT 中使用此配置：

 spark_conf = SparkConf().setAppName('app_name') \
    .setMaster("local[4]") \
    .set('spark.executor.memory', "8g") \
    .set('spark.executor.cores', 4) \
    .set('spark.task.cpus', 1)

    sc = SparkContext(conf=spark_conf)

更新 1：

我还使用虚拟环境并在其中安装所有包以在节点之间分发它们。 详情：

1. 要在 python 中创建虚拟环境，请运行以下命令：

    sudo apt install python3.8-venv

2. 创建虚拟环境：

    python3 -m venv my_venv

3. 进入环境：

    source my_vent/bin/activate

4. 我使用venv-pack打包您在项目中安装的所有包。

    pip install venv-pack

5. 打包包裹：

    venv-pack -o my_venv.tar.gz

此外，正如Spark网站所说，我将项目的所有.py文件放在一个文件夹中并将其压缩到.zip文件夹中。

最后在创建集群之后，我运行这个命令：

  /opt/spark/bin/spark-submit --master spark://x.x.x.x:7077 --archives my_venv.tar.gz#environment --py-files my_files.zip main.py

但是，它最终会出现此错误：

  Traceback (most recent call last):
  File "/home/spark/Desktop/etl/main.py", line 3, in <module>
  import cst_utils as cu
  File "/home/spark/Desktop/etl/cst_utils.py", line 5, in <module>
  import group_state as gs
  File "/home/spark/Desktop/etl/group_state.py", line 1, in <module>
  import numpy as np
  ModuleNotFoundError: No module named 'numpy'

你能指导我在集群中运行代码有什么问题吗？

任何帮助将非常感激。

Answer 1

问题解决了。

首先，我使用以下命令在每个节点中安装了所有包：

 python3 -m pip install PACKAGE

然后，当我运行程序时，我必须将程序中使用的所有 PY 文件写在--py-files前面，如下所示：

 /opt/spark/bin/spark-submit --master spark://x.x.x.x:7077 --files sparkConfig.json --py-files cst_utils.py,grouping.py,group_state.py,g_utils.py,csts.py,oracle_connection.py,config.py,brn_utils.py,emp_utils.py main.py    

然后我没有关于导入文件的任何错误。