从 S3 将嵌套文本文件读入 spark 时出现内存错误

Posted 2023-04-15

【中文标题】从 S3 将嵌套文本文件读入 spark 时出现内存错误

【英文标题】：Memory error when reading nested textfiles into spark from S3

【发布时间】：2016-07-25 20:24:42

【问题描述】：

我正在尝试从S3 将大约一百万个压缩文本文件读入spark。每个文件的压缩大小在 50 MB 到 80 MB 之间。总共大约有 6.5 TB 的数据。

不幸的是，我遇到了不知道如何解决的内存不足异常。一些简单的事情：

raw_file_list = subprocess.Popen("aws s3 ls --recursive s3://my-bucket/export/", shell=True, stdout=subprocess.PIPE).stdout.read().strip().split('\n')
cleaned_names = ["s3://my-bucket/" + f.split()[3] for f in raw_file_list if not f.endswith('_SUCCESS')]
dat = sc.textFile(','.join(cleaned_names))
dat.count()

产量：

 ---------------------------------------------------------------------------
Py4JJavaError                             Traceback (most recent call last)
<ipython-input-22-8ce3c7d1073e> in <module>() ----> 1 dat.count()

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/pyspark/rdd.pyc in count(self)
   1002         3
   1003         """
-> 1004         return self.mapPartitions(lambda i: [sum(1 for _ in i)]).sum()
   1005 
   1006     def stats(self):

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/pyspark/rdd.pyc in sum(self)
    993         6.0
    994         """
--> 995         return self.mapPartitions(lambda x: [sum(x)]).fold(0, operator.add)
    996 
    997     def count(self):

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/pyspark/rdd.pyc in fold(self, zeroValue, op)
    867         # zeroValue provided to each partition is unique from the one provided
    868         # to the final reduce call
--> 869         vals = self.mapPartitions(func).collect()
    870         return reduce(op, vals, zeroValue)
    871 

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/pyspark/rdd.pyc in collect(self)
    769         """
    770         with SCCallSiteSync(self.context) as css:
--> 771             port = self.ctx._jvm.PythonRDD.collectAndServe(self._jrdd.rdd())
    772         return list(_load_from_socket(port, self._jrdd_deserializer))
    773 

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/lib/py4j-0.9-src.zip/py4j/java_gateway.py in __call__(self, *args)
    811         answer = self.gateway_client.send_command(command)
    812         return_value = get_return_value(
--> 813             answer, self.gateway_client, self.target_id, self.name)
    814 
    815         for temp_arg in temp_args:

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/pyspark/sql/utils.pyc in deco(*a, **kw)
     43     def deco(*a, **kw):
     44         try:
---> 45             return f(*a, **kw)
     46         except py4j.protocol.Py4JJavaError as e:
     47             s = e.java_exception.toString()

/tmp/spark-tmp-lminer/spark-1.6.1-bin-hadoop2.6/python/lib/py4j-0.9-src.zip/py4j/protocol.py in get_return_value(answer, gateway_client, target_id, name)
    306                 raise Py4JJavaError(
    307                     "An error occurred while calling 012.\n".
--> 308                     format(target_id, ".", name), value)
    309             else:
    310                 raise Py4JError(

Py4JJavaError: An error occurred while calling z:org.apache.spark.api.python.PythonRDD.collectAndServe.
: java.lang.OutOfMemoryError: GC overhead limit exceeded

更新：

部分问题似乎已通过此post 解决。似乎火花很难从 S3 处理这么多文件。更新了错误，现在它只反映内存问题。

【问题讨论】：

您能否尝试为spark.driver.memory、spark.executor.memory 分配更多内存，并且还可能分配Java Heap 大小。

【参考方案1】：

问题是文件太多。解决方案似乎是通过读取文件子集并将它们合并为较小的数量来减少分区的数量。但是你不能让分区太大：500 - 1000 MB 的文件会导致它们自己的问题。