[HADOOP-16109] Parquet reading S3AFileSystem causes EOF - ASF JIRA

XML

Word

Printable

JSON

Details

Type: Sub-task
Status: Resolved
Priority: Blocker
Resolution: Fixed
Affects Version/s: 2.9.2, 2.8.5, 3.3.0, 3.1.2
Fix Version/s: 2.10.0, 3.0.4, 3.3.0, 2.8.6, 3.2.1, 2.9.3, 3.1.3, 3.1.4, 3.2.2
Component/s: fs/s3
Labels:
None

Description

When using S3AFileSystem to read Parquet files a specific set of circumstances causes an EOFException that is not thrown when reading the same file from local disk

Note this has only been observed under specific circumstances:
- when the reader is doing a projection (will cause it to do a seek backwards and put the filesystem into random mode)
- when the file is larger than the readahead buffer size
- when the seek behavior of the Parquet reader causes the reader to seek towards the end of the current input stream without reopening, such that the next read on the currently open stream will read past the end of the currently open stream.

Exception from Parquet reader is as follows:

Caused by: java.io.EOFException: Reached the end of stream with 51 bytes left to read
 at org.apache.parquet.io.DelegatingSeekableInputStream.readFully(DelegatingSeekableInputStream.java:104)
 at org.apache.parquet.io.DelegatingSeekableInputStream.readFullyHeapBuffer(DelegatingSeekableInputStream.java:127)
 at org.apache.parquet.io.DelegatingSeekableInputStream.readFully(DelegatingSeekableInputStream.java:91)
 at org.apache.parquet.hadoop.ParquetFileReader$ConsecutiveChunkList.readAll(ParquetFileReader.java:1174)
 at org.apache.parquet.hadoop.ParquetFileReader.readNextRowGroup(ParquetFileReader.java:805)
 at org.apache.parquet.hadoop.InternalParquetRecordReader.checkRead(InternalParquetRecordReader.java:127)
 at org.apache.parquet.hadoop.InternalParquetRecordReader.nextKeyValue(InternalParquetRecordReader.java:222)
 at org.apache.parquet.hadoop.ParquetRecordReader.nextKeyValue(ParquetRecordReader.java:207)
 at org.apache.flink.api.java.hadoop.mapreduce.HadoopInputFormatBase.fetchNext(HadoopInputFormatBase.java:206)
 at org.apache.flink.api.java.hadoop.mapreduce.HadoopInputFormatBase.reachedEnd(HadoopInputFormatBase.java:199)
 at org.apache.flink.runtime.operators.DataSourceTask.invoke(DataSourceTask.java:190)
 at org.apache.flink.runtime.taskmanager.Task.run(Task.java:711)
 at java.lang.Thread.run(Thread.java:748)

The following example program generate the same root behavior (sans finding a Parquet file that happens to trigger this condition) by purposely reading past the already active readahead range on any file >= 1029 bytes in size..

final Configuration conf = new Configuration();
conf.set("fs.s3a.readahead.range", "1K");
conf.set("fs.s3a.experimental.input.fadvise", "random");

final FileSystem fs = FileSystem.get(path.toUri(), conf);
// forward seek reading across readahead boundary
try (FSDataInputStream in = fs.open(path)) {
    final byte[] temp = new byte[5];
    in.readByte();
    in.readFully(1023, temp); // <-- works
}
// forward seek reading from end of readahead boundary
try (FSDataInputStream in = fs.open(path)) {
 final byte[] temp = new byte[5];
 in.readByte();
 in.readFully(1024, temp); // <-- throws EOFException
}