hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Arun C Murthy (JIRA)" <j...@apache.org>
Subject [jira] Updated: (HADOOP-3821) SequenceFile's Reader.decompressorPool or Writer.decompressorPool gets into an inconsistent state when calling close() more than once
Date Mon, 11 Aug 2008 18:28:44 GMT

     [ https://issues.apache.org/jira/browse/HADOOP-3821?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Arun C Murthy updated HADOOP-3821:
----------------------------------

    Status: Open  (was: Patch Available)

> SequenceFile's Reader.decompressorPool or Writer.decompressorPool gets into an inconsistent
state when calling close() more than once
> -------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: HADOOP-3821
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3821
>             Project: Hadoop Core
>          Issue Type: Bug
>          Components: io
>    Affects Versions: 0.17.1, 0.17.0, 0.18.0, 0.19.0
>            Reporter: Peter Voss
>            Assignee: Arun C Murthy
>             Fix For: 0.19.0
>
>         Attachments: HADOOP-3821_0_20080724.patch, HADOOP-3821_test1_20080805.patch
>
>
> SequenceFile.Reader uses a decompressorPool to reuse Decompressor instances. The Reader
obtains such an instance from the pool on object creation and returns it back to the pool
it when {{close()}} is called.
> SequenceFile.Reader implements the {{java.io.Closable}} interface and it's spec on the
{{close()}} method says:
> {quote}
> Closes this stream and releases any system resources associated 
> with it. If the stream is already closed then invoking this 
> method has no effect.
> {quote}
> This spec is violated by the Reader implementation, because calling {{close()}} multiple
times has really bad implications. 
> When you call {{close()}} twice, one and the same Decompressor instances will be returned
to the pool two times and the pool would now maintain duplicated references to the same Decompressor
instances. When other Readers now request instances from the pool it might happen that two
Readers get the same Decompressor instance.
> The correct behavior would be to just ignore a second call to {{close()}}.
> The exact same issue applies to the SequenceFile.Writer as well.
> We were having big trouble with this, because we were observing sporadic exceptions from
merge operations. The strange thing was that executing the same merge again usually succeeded.
But sometimes it took multiple attempts to complete a merge successfully. It was very hard
to debug that the root cause was some duplicated Decompressor references in the decompressorPool.
> Exceptions that we observed in production looked like this (we were using hadoop 0.17.0):
> {noformat}
> java.io.IOException: unknown compression method
> at org.apache.hadoop.io.compress.zlib.BuiltInZlibInflater.decompress(BuiltInZlibInflater.java:47)
> at org.apache.hadoop.io.compress.DecompressorStream.decompress(DecompressorStream.java:80)
> at org.apache.hadoop.io.compress.DecompressorStream.read(DecompressorStream.java:74)
> at java.io.DataInputStream.readFully(DataInputStream.java:178)
> at org.apache.hadoop.io.DataOutputBuffer$Buffer.write(DataOutputBuffer.java:56)
> at org.apache.hadoop.io.DataOutputBuffer.write(DataOutputBuffer.java:90)
> at org.apache.hadoop.io.SequenceFile$Reader.nextRawKey(SequenceFile.java:1995)
> at org.apache.hadoop.io.SequenceFile$Sorter$SegmentDescriptor.nextRawKey(SequenceFile.java:3002)
> at org.apache.hadoop.io.SequenceFile$Sorter$MergeQueue.next(SequenceFile.java:2760)
> at org.apache.hadoop.io.SequenceFile$Sorter.writeFile(SequenceFile.java:2625)
> at org.apache.hadoop.io.SequenceFile$Sorter.merge(SequenceFile.java:2644)
> {noformat}
> or 
> {noformat}
> java.io.IOException: zero length keys not allowed
> at org.apache.hadoop.io.SequenceFile$BlockCompressWriter.appendRaw(SequenceFile.java:1340)
> at org.apache.hadoop.io.SequenceFile$Sorter.writeFile(SequenceFile.java:2626)
> at org.apache.hadoop.io.SequenceFile$Sorter.merge(SequenceFile.java:2644)
> {noformat}
> The following snippet reproduces the problem:
> {code:java}
>     public void testCodecPool() throws IOException {
>         Configuration conf = new Configuration();
>         LocalFileSystem fs = new LocalFileSystem();
>         fs.setConf(conf);
>         fs.getRawFileSystem().setConf(conf);
>         // create a sequence file
>         Path path = new Path("target/seqFile");
>         SequenceFile.Writer writer = SequenceFile.createWriter(fs, conf, path, Text.class,
NullWritable.class, CompressionType.BLOCK);
>         writer.append(new Text("key1"), NullWritable.get());
>         writer.append(new Text("key2"), NullWritable.get());
>         writer.close();
>         // Create a reader which uses 4 BuiltInZLibInflater instances
>         SequenceFile.Reader reader = new SequenceFile.Reader(fs, path, conf);
>         // Returns the 4 BuiltInZLibInflater instances to the CodecPool
>         reader.close();
>         // The second close erroneously returns the same 4 BuiltInZLibInflater instances
to the CodecPool again
>         reader.close();
>         // The first reader gets 4 BuiltInZLibInflater instances from the CodecPool
>         SequenceFile.Reader reader1 = new SequenceFile.Reader(fs, path, conf);
>         // read first value from reader1
>         Text text = new Text();
>         reader1.next(text);
>         assertEquals("key1", text.toString());
>         // The second reader gets the same 4 BuiltInZLibInflater instances from the CodePool
as reader1
>         SequenceFile.Reader reader2 = new SequenceFile.Reader(fs, path, conf);
>         // read first value from reader2
>         reader2.next(text);
>         assertEquals("key1", text.toString());
>         // read second value from reader1
>         reader1.next(text);
>         assertEquals("key2", text.toString());
>         // read second value from reader2 (this throws an exception)
>         reader2.next(text);
>         assertEquals("key2", text.toString());
>         
>         assertFalse(reader1.next(text));
>         assertFalse(reader2.next(text));
>     }
> {code}
> It fails with the exception:
> {noformat}
> java.io.EOFException
> 	at java.io.DataInputStream.readByte(DataInputStream.java:243)
> 	at org.apache.hadoop.io.WritableUtils.readVLong(WritableUtils.java:324)
> 	at org.apache.hadoop.io.WritableUtils.readVInt(WritableUtils.java:345)
> 	at org.apache.hadoop.io.SequenceFile$Reader.next(SequenceFile.java:1835)
> 	at CodecPoolTest.testCodecPool(CodecPoolTest.java:56)
> {noformat}
> But this is just a very simple test that shows the problem. Much more weired things can
happen when running in a complex production environment. Esp. heavy concurrency makes the
behavior much more exciting. ;-)

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message