从 spark-sql-streaming 写入 hdfs 和 s3 时如何避免小文件问题

Question

我在我的项目中使用 spark-sql-2.3.1v、kafka 和 java8。 与

--driver-memory 4g \
--driver-cores 2 \
--num-executors 120 \
--executor-cores 1 \
--executor-memory 768m \

在消费者方面，我试图在 hdfs 中写入文件 我在下面的代码中使用了类似的东西

                          dataSet.writeStream()
                                        .format("parquet")
                                        .option("path", parqetFileName)
                                        .option("mergeSchema", true)
                                        .outputMode("Append")
                                        .partitionBy("company_id","date")
                                        .option("checkpointLocation", checkPtLocation)
                                        .trigger(Trigger.ProcessingTime("25 seconds"))
                                        .start();

当我存储到 hdfs 文件夹时，它看起来如下，即每个文件大约 1.5k+，即几个 KB。

$ hdfs dfs -du -h /transactions/company_id=24779/date=2014-06-24/
1.5 K  /transactions/company_id=24779/date=2014-06-24/part-00026-1027fff9-5745-4250-961a-fd56508b7ea3.c000.snappy.parquet
1.5 K  /transactions/company_id=24779/date=2014-06-24/part-00057-6604f6cc-5b8d-41f4-8fc0-14f6e13b4a37.c000.snappy.parquet
1.5 K  /transactions/company_id=24779/date=2014-06-24/part-00098-754e6929-9c75-430f-b6bb-3457a216aae3.c000.snappy.parquet
1.5 K  /transactions/company_id=24779/date=2014-06-24/part-00099-1d62cbd5-7409-4259-b4f3-d0f0e5a93da3.c000.snappy.parquet
1.5 K  /transactions/company_id=24779/date=2014-06-24/part-00109-1965b052-c7a6-47a8-ae15-dea301010cf5.c000.snappy.parquet

由于文件很小，需要大量处理时间， 当我从 hdfs 读回更大的数据集并计算 行，然后它会导致以下堆空间错误。

2020-02-12 07:07:57,475 [Driver] ERROR org.apache.spark.deploy.yarn.ApplicationMaster - User class threw exception: java.lang.OutOfMemoryError: Java heap space
java.lang.OutOfMemoryError: Java heap space
        at java.util.Arrays.copyOfRange(Arrays.java:3664)
        at java.lang.String.<init>(String.java:207)
        at java.lang.String.substring(String.java:1969)
        at java.net.URI$Parser.substring(URI.java:2869)
        at java.net.URI$Parser.parse(URI.java:3049)
        at java.net.URI.<init>(URI.java:588)
        at org.apache.spark.sql.execution.streaming.SinkFileStatus.toFileStatus(FileStreamSinkLog.scala:52)
        at org.apache.spark.sql.execution.streaming.MetadataLogFileIndex$$anonfun$2.apply(MetadataLogFileIndex.scala:46)
        at org.apache.spark.sql.execution.streaming.MetadataLogFileIndex$$anonfun$2.apply(MetadataLogFileIndex.scala:46)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.TraversableLike$$anonfun$map$1.apply(TraversableLike.scala:234)
        at scala.collection.IndexedSeqOptimized$class.foreach(IndexedSeqOptimized.scala:33)
        at scala.collection.mutable.ArrayOps$ofRef.foreach(ArrayOps.scala:186)
        at scala.collection.TraversableLike$class.map(TraversableLike.scala:234)
        at scala.collection.mutable.ArrayOps$ofRef.map(ArrayOps.scala:186)
        at org.apache.spark.sql.execution.streaming.MetadataLogFileIndex.<init>(MetadataLogFileIndex.scala:46)
        at org.apache.spark.sql.execution.datasources.DataSource.resolveRelation(DataSource.scala:336)
        at org.apache.spark.sql.DataFrameReader.loadV1Source(DataFrameReader.scala:223)
        at org.apache.spark.sql.DataFrameReader.load(DataFrameReader.scala:211)
        at org.apache.spark.sql.DataFrameReader.load(DataFrameReader.scala:167)
        at com.spgmi.ca.prescore.utils.DbUtils.loadFromHdfs(DbUtils.java:129)
        at com.spgmi.ca.prescore.spark.CountRecords.main(CountRecords.java:84)
        at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
        at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
        at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
        at java.lang.reflect.Method.invoke(Method.java:498)
        at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$2.run(ApplicationMaster.scala:684)
2020-02-12 07:07:57,533 [Reporter] WARN  org.apache.spark.deploy.yarn.ApplicationMaster - Reporter thread fails 1 time(s) in a row.
java.io.IOException: Failed on local exception: java.nio.channels.ClosedByInterruptException; Host Details : local host is: dev1-dev.com":8030;
        at org.apache.hadoop.net.NetUtils.wrapException(NetUtils.java:805)
        at org.apache.hadoop.ipc.Client.getRpcResponse(Client.java:1497)
        at org.apache.hadoop.ipc.Client.call(Client.java:1439)
        at org.apache.hadoop.ipc.Client.call(Client.java:1349)
        at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.invoke(ProtobufRpcEngine.java:227)
        at org.apache.hadoop.ipc.ProtobufRpcEngine$Invoker.invoke(ProtobufRpcEngine.java:116)
        at com.sun.proxy.$Proxy22.allocate(Unknown Source)
        at org.apache.hadoop.yarn.api.impl.pb.client.ApplicationMasterProtocolPBClientImpl.allocate(ApplicationMasterProtocolPBClientImpl.java:77)
        at sun.reflect.GeneratedMethodAccessor32.invoke(Unknown Source)
        at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
        at java.lang.reflect.Method.invoke(Method.java:498)
        at org.apache.hadoop.io.retry.RetryInvocationHandler.invokeMethod(RetryInvocationHandler.java:422)
        at org.apache.hadoop.io.retry.RetryInvocationHandler$Call.invokeMethod(RetryInvocationHandler.java:165)
        at org.apache.hadoop.io.retry.RetryInvocationHandler$Call.invoke(RetryInvocationHandler.java:157)
        at org.apache.hadoop.io.retry.RetryInvocationHandler$Call.invokeOnce(RetryInvocationHandler.java:95)
        at org.apache.hadoop.io.retry.RetryInvocationHandler.invoke(RetryInvocationHandler.java:359)
        at com.sun.proxy.$Proxy23.allocate(Unknown Source)
        at org.apache.hadoop.yarn.client.api.impl.AMRMClientImpl.allocate(AMRMClientImpl.java:296)
        at org.apache.spark.deploy.yarn.YarnAllocator.allocateResources(YarnAllocator.scala:249)
        at org.apache.spark.deploy.yarn.ApplicationMaster.org$apache$spark$deploy$yarn$ApplicationMaster$$allocationThreadImpl(ApplicationMaster.scala:540)
        at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$1.run(ApplicationMaster.scala:606)
Caused by: java.nio.channels.ClosedByInterruptException
        at java.nio.channels.spi.AbstractInterruptibleChannel.end(AbstractInterruptibleChannel.java:202)
        at sun.nio.ch.SocketChannelImpl.finishConnect(SocketChannelImpl.java:753)
        at org.apache.hadoop.net.SocketIOWithTimeout.connect(SocketIOWithTimeout.java:206)
        at org.apache.hadoop.net.NetUtils.connect(NetUtils.java:531)
        at org.apache.hadoop.ipc.Client$Connection.setupConnection(Client.java:687)
        at org.apache.hadoop.ipc.Client$Connection.setupIOstreams(Client.java:790)
        at org.apache.hadoop.ipc.Client$Connection.access$3500(Client.java:411)
        at org.apache.hadoop.ipc.Client.getConnection(Client.java:1554)

问题：

1. 这些小文件会不会导致 Spark 处理中出现“小文件问题”？ 如果是这样如何处理这种情况。

2. 如果我想统计给定hdfs文件夹的记录总数，该怎么做？

3. 如何知道处理此类数据需要多少堆空间？

新变化后

--driver-memory 16g \
--driver-cores 1 \
--num-executors 120 \
--executor-cores 1 \
--executor-memory 768m \

运行成功的结果是：

2020-02-12 20:28:56,188 [Driver] WARN  com.spark.mypackage.CountRecords -  NUMBER OF PARTITIONS AFTER HDFS READ : 77926
+--------------------+-----+
|SPARK_PARTITION_ID()|count|
+--------------------+-----+
|               24354|   94|
|               26425|   96|
|               32414|   64|
|               76143|   32|
|               16861|   32|
|               30903|   64|
|               40335|   32|
|               64121|   64|
|               69042|   32|
|               32539|   64|
|               34759|   32|
|               41575|   32|
|                1591|   64|
|                3050|   98|
|               51772|   32|
+--------------------+-----+

2020-02-12 20:50:32,301 [Driver] WARN  com.spark.mypackage.CountRecords -  RECORD COUNT: 3999708

Answer 1

1. 是的。小文件不仅是 Spark 的问题。它会在您的 NameNode 上造成不必要的负载。在处理小文件时，您应该花更多时间压缩和上传较大的文件，而不是担心 OOM。您的文件小于 64MB / 128MB 的事实表明您对 Hadoop 的使用不佳。

2. spark.read("hdfs://path").count() 之类的东西会读取路径中的所有文件，然后计算 Dataframe 中的行数

3. 没有固定编号。您需要对您的作业启用 JMX 监控并查看堆大小达到了多少。否则，将您分配给工作的当前内存任意加倍，直到它开始没有 OOM。如果您开始接近 8GB 以上，那么您需要考虑通过添加更多并行化来减少每个作业中读取的数据。

---

FWIW，Kafka Connect 也可用于输出分区的 HDFS/S3 路径。