如何从 Spark MLlib FP 增长模型中提取数据

Question

我在独立模式下运行 spark master 和 slave，没有 Hadoop 集群。 使用 spark-shell，我可以使用我的数据快速构建 FPGrowthModel。 构建模型后，我将尝试查看模型中捕获的模式和频率，但 spark 挂在具有更大数据集（200000 * 2000 矩阵等数据）的 collect() 方法（通过查看 Spark UI）上。 这是我在 spark-shell 中运行的代码：

import org.apache.spark.mllib.fpm.{FPGrowth, FPGrowthModel}
import org.apache.spark.rdd.RDD

val textFile = sc.textFile("/path/to/txt/file")
val data = textFile.map(_.split(" ")).cache()

val fpg = new FPGrowth().setMinSupport(0.9).setNumPartitions(8)
val model = fpg.run(data)

model.freqItemsets.collect().foreach { itemset =>
  println(itemset.items.mkString("[", ",", "]") + ", " + itemset.freq)
}

我尝试将 spark shell 内存从 512MB 增加到 2GB，但似乎并没有缓解挂起问题。我不确定是因为需要 Hadoop 才能执行此任务，还是我需要更多地增加 spark-shell 内存或其他原因。

15/08/10 22:19:40 ERROR TaskSchedulerImpl: Lost executor 0 on 142.103.22.23: remote Rpc client disassociated
15/08/10 22:19:40 WARN ReliableDeliverySupervisor: Association with remote system [akka.tcp://sparkExecutor@142.103.22.23:43440] has failed, address is now gated for [5000] ms. Reason is: [Disassociated].
15/08/10 22:19:40 INFO AppClient$ClientActor: Executor updated: app-20150810163957-0001/0 is now EXITED (Command exited with code 137)
15/08/10 22:19:40 INFO TaskSetManager: Re-queueing tasks for 0 from TaskSet 4.0
15/08/10 22:19:40 INFO SparkDeploySchedulerBackend: Executor app-20150810163957-0001/0 removed: Command exited with code 137
15/08/10 22:19:40 WARN TaskSetManager: Lost task 3.0 in stage 4.0 (TID 59, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 6.0 in stage 4.0 (TID 62, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 0.0 in stage 4.0 (TID 56, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 2.0 in stage 4.0 (TID 58, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 5.0 in stage 4.0 (TID 61, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 4.0 in stage 4.0 (TID 60, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 7.0 in stage 4.0 (TID 63, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 WARN TaskSetManager: Lost task 1.0 in stage 4.0 (TID 57, 142.103.22.23): ExecutorLostFailure (executor 0 lost)
15/08/10 22:19:40 ERROR SparkDeploySchedulerBackend: Asked to remove non-existent executor 0
15/08/10 22:19:40 INFO AppClient$ClientActor: Executor added: app-20150810163957-0001/1 on worker-20150810163259-142.103.22.23-48853 (142.103.22.23:48853) with 8 cores
15/08/10 22:19:40 INFO SparkDeploySchedulerBackend: Granted executor ID app-20150810163957-0001/1 on hostPort 142.103.22.23:48853 with 8 cores, 15.0 GB RAM
15/08/10 22:19:40 INFO AppClient$ClientActor: Executor updated: app-20150810163957-0001/1 is now LOADING
15/08/10 22:19:40 INFO DAGScheduler: Executor lost: 0 (epoch 2)
15/08/10 22:19:40 INFO AppClient$ClientActor: Executor updated: app-20150810163957-0001/1 is now RUNNING
15/08/10 22:19:40 INFO BlockManagerMasterEndpoint: Trying to remove executor 0 from BlockManagerMaster.
15/08/10 22:19:40 INFO BlockManagerMasterEndpoint: Removing block manager BlockManagerId(0, 142.103.22.23, 37411)
15/08/10 22:19:40 INFO BlockManagerMaster: Removed 0 successfully in removeExecutor
15/08/10 22:19:40 INFO ShuffleMapStage: ShuffleMapStage 3 is now unavailable on executor 0 (0/16, false)

Answer 1

如果数据集很大，你不应该运行 .collect()，比如它有几个 GB，你不应该使用它，它有助于加快执行多个评估的速度。 在不收集的情况下运行 foreach 循环。

Answer 2

Kryo 是比 org.apache.spark.serializer.JavaSerializer 更快的序列化程序。一种可能的解决方法是告诉 spark 不要使用 Kryo：

val conf = (new org.apache.spark.SparkConf()
.setAppName("APP_NAME")
.set("spark.serializer", "org.apache.spark.serializer.JavaSerializer")

然后尝试再次运行上面的代码。

请参阅此链接以供参考：

FPGrowth Algorithm in Spark

Answer 3

尝试将collect() 替换为local iterator。最终，您可能会遇到 FPGrowth 实现的限制。请参阅我的帖子 here 和 Spark JIRA issue。