Spark 之 partitons

Listing leaf files and directories

分析其并行化

org.apache.spark.util.HadoopFSUtils

      sc.parallelize(paths, numParallelism)
        .mapPartitions { pathsEachPartition =>
          val hadoopConf = serializableConfiguration.value
          pathsEachPartition.map { path =>
            val leafFiles = listLeafFiles(
              path = path,
              hadoopConf = hadoopConf,
              filter = filter,
              contextOpt = None, // Can't execute parallel scans on workers
              ignoreMissingFiles = ignoreMissingFiles,
              ignoreLocality = ignoreLocality,
              isRootPath = isRootLevel,
              parallelismThreshold = Int.MaxValue,
              parallelismMax = 0)
            (path, leafFiles)
          }
        }.collect()

    // Set the number of parallelism to prevent following file listing from generating many tasks
    // in case of large #defaultParallelism.
    val numParallelism = Math.min(paths.size, parallelismMax)

parallelismMax 最终由以下配置决定。

  val PARALLEL_PARTITION_DISCOVERY_PARALLELISM =
    buildConf("spark.sql.sources.parallelPartitionDiscovery.parallelism")
      .doc("The number of parallelism to list a collection of path recursively, Set the " +
        "number to prevent file listing from generating too many tasks.")
      .version("2.1.1")
      .internal()
      .intConf
      .createWithDefault(10000)

原文地址：https://blog.csdn.net/zhixingheyi_tian/article/details/144351939

免责声明：本站文章内容转载自网络资源，如本站内容侵犯了原著者的合法权益，可联系本站删除。更多内容请关注自学内容网（zxcms.com）！