使用Spark缓慢将Parquet写入HDFS

Question

我正在使用Spark 1.6.1并写入HDFS。在某些情况下，似乎所有工作都由一个线程完成。这是为什么？

另外，我需要parquet.enable.summary-metadata将镶木地板文件注册到Impala。

Df.write().partitionBy("COLUMN").parquet(outputFileLocation);

同样，所有这些似乎都发生在执行器的一个cpu中。

16/11/03 14:59:20 INFO datasources.DynamicPartitionWriterContainer: Using user defined output committer class org.apache.parquet.hadoop.ParquetOutputCommitter
16/11/03 14:59:20 INFO mapred.SparkHadoopMapRedUtil: No need to commit output of task because needsTaskCommit=false: attempt_201611031459_0154_m_000029_0
16/11/03 15:17:56 INFO sort.UnsafeExternalSorter: Thread 545 spilling sort data of 41.9 GB to disk (3  times so far)
16/11/03 15:21:05 INFO storage.ShuffleBlockFetcherIterator: Getting 0 non-empty blocks out of 0 blocks
16/11/03 15:21:05 INFO storage.ShuffleBlockFetcherIterator: Started 0 remote fetches in 1 ms
16/11/03 15:21:05 INFO datasources.DynamicPartitionWriterContainer: Using user defined output committer class org.apache.parquet.hadoop.ParquetOutputCommitter
16/11/03 15:21:05 INFO codec.CodecConfig: Compression: GZIP
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Parquet block size to 134217728
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Parquet page size to 1048576
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Parquet dictionary page size to 1048576
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Dictionary is on
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Validation is off
16/11/03 15:21:05 INFO hadoop.ParquetOutputFormat: Writer version is: PARQUET_1_0
16/11/03 15:21:05 INFO parquet.CatalystWriteSupport: Initialized Parquet WriteSupport with Catalyst schema:

然后再：-

16/11/03 15:21:05 INFO compress.CodecPool: Got brand-new compressor [.gz]
16/11/03 15:21:05 INFO datasources.DynamicPartitionWriterContainer: Maximum partitions reached, falling back on sorting.
16/11/03 15:32:37 INFO sort.UnsafeExternalSorter: Thread 545 spilling sort data of 31.8 GB to disk (0  time so far)
16/11/03 15:45:47 INFO sort.UnsafeExternalSorter: Thread 545 spilling sort data of 31.8 GB to disk (1  time so far)
16/11/03 15:48:44 INFO datasources.DynamicPartitionWriterContainer: Sorting complete. Writing out partition files one at a time.
16/11/03 15:48:44 INFO codec.CodecConfig: Compression: GZIP
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Parquet block size to 134217728
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Parquet page size to 1048576
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Parquet dictionary page size to 1048576
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Dictionary is on
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Validation is off
16/11/03 15:48:44 INFO hadoop.ParquetOutputFormat: Writer version is: PARQUET_1_0
16/11/03 15:48:44 INFO parquet.CatalystWriteSupport: Initialized Parquet WriteSupport with Catalyst schema:

模式

以下行中的200条左右重复20次左右。

16/11/03 15:48:44 INFO compress.CodecPool: Got brand-new compressor [.gz]
16/11/03 15:49:50 INFO hadoop.InternalParquetRecordWriter: mem size 135,903,551 > 134,217,728: flushing 1,040,100 records to disk.
16/11/03 15:49:50 INFO hadoop.InternalParquetRecordWriter: Flushing mem columnStore to file. allocated memory: 89,688,651

大约200行以下

16/11/03 15:49:51 INFO hadoop.ColumnChunkPageWriteStore: written 413,231B for [a17bbfb1_2808_11e6_a4e6_77b5e8f92a4f] BINARY: 1,040,100 values, 1,138,534B raw, 412,919B comp, 8 pages, encodings: [RLE, BIT_PACKED, PLAIN_DICTIONARY], dic { 356 entries, 2,848B raw, 356B comp}

然后最后：-

16/11/03 16:15:41 INFO output.FileOutputCommitter: Saved output of task 'attempt_201611031521_0154_m_000040_0' to hdfs://PATH/_temporary/0/task_201611031521_0154_m_000040
16/11/03 16:15:41 INFO mapred.SparkHadoopMapRedUtil: attempt_201611031521_0154_m_000040_0: Committed
16/11/03 16:15:41 INFO executor.Executor: Finished task 40.0 in stage 154.0 (TID 8545). 3757 bytes result sent to driver

更新：parquet.enable.summary-metadata设置为false。
将分区减少到21。

Df.write().mode(SaveMode.Append).partitionBy("COL").parquet(outputFileLocation);

它确实提高了速度，但仍然需要一个小时才能完成。

更新：-大多数问题的原因是多个左外部联接，并且在写入之前就实现了非常小的数据。溢出是由于Append模式导致文件保持打开状态而发生的。在此模式下，默认限制为5个打开的文件。您可以使用属性“ spark.sql.sources.maxConcurrentWrites”来增加它

Answer 1

最后，在到达写入部分之前对代码进行一些优化后，我们获得了更好的写入时间。之前我们无法进行重新分区，因为 shuffle 超过 4-5 GB。在之前的更改之后，我将代码从合并更改为重新分区，通过为执行器中的每个 CPU 提供相同数量的要写入的数据，将数据分布到所有执行器上。因此，如果您发现作业创建的 parquet 文件大小不同，请尝试在写入之前重新分区 Dataframe。

此外，这也有助于提高写入性能：-

sc.hadoopConfiguration.set("parquet.enable.dictionary", "false")