11

I'm storing files on HDFS in Snappy compression format. I'd like to be able to examine these files on my local Linux file system to make sure that the Hadoop process that created them has performed correctly.

When I copy them locally and attempt to de-compress them with the Google standard libarary, it tells me that the file is missing the Snappy identifier. When I try to go around this by inserting a Snappy identifier, it messes up the checksum.

What can I do to read these files without having to write a separate Hadoop program or pass it through something like Hive?

Robert Rapplean
  • 672
  • 1
  • 9
  • 30

4 Answers4

26

I finally found out that I can use the following command to read the contents of a Snappy compressed file on HDFS:

hadoop fs -text /path/filename

Using the latest commands on Cloudera or HDP:

hdfs dfs -text /path/filename

If the intent is to download the file in text format for additional examination and processing, the output of that command can be piped to a file on the local system. You can also use head to just view the first few lines of the file.

Robert Rapplean
  • 672
  • 1
  • 9
  • 30
  • How can I do this programatically in scala or java ? – Bunder Jan 25 '17 at 16:37
  • To write to a Snappy file programmatically, you need to import the Snappy codec class and get an instance of that class as part of the mapper or reducer setup. You need to pass your output stream through the codec's "createOutputStream" function to get the encoded output stream. Here's a snippet. Reading is the same in reverse codec = (CompressionCodec) ReflectionUtils.newInstance(codecClass, conf); fileOut = fs.create(targetPath, false); thiswriter = new LineRecordWriter(new DataOutputStream(codec.createOutputStream(fileOut))); – Robert Rapplean Mar 09 '17 at 21:23
3

Please take a look at this post on Cloudera blog. It explains how to use Snappy with Hadoop. Essentially, Snappy files on raw text are not splittable, so you cannot read a single file across multiple hosts.

The solution is to use Snappy in a container format, so essentially you're using Hadoop SequenceFile with compression set as Snappy. As described in this answer, you can set the property mapred.output.compression.codec to org.apache.hadoop.io.compress.SnappyCodec and setup your job output format as SequenceFileOutputFormat.

And then to read it, you should only need to use SequenceFile.Reader because the codec information is stored in the file header.

Community
  • 1
  • 1
Charles Menguy
  • 40,830
  • 17
  • 95
  • 117
  • 1
    Thanks, Charles, but I don't think that this addresses my question. Let me simplify it. I use hadoop fs -get filename to move a file from HDFS to my local directory on Linux. Now that I have it here, why can't I use the snappy java libraries to decompress it? – Robert Rapplean May 22 '13 at 16:51
0

Thats because snappy used by hadoop has some more meta data which is not undesrtood by libraries like https://code.google.com/p/snappy/, You need to use hadoop native snappy to unsnap the data file that you downloaded.

Jyotirmoy Sundi
  • 355
  • 1
  • 5
  • 12
0

If you land here and are trying to decompress a .snappy file via a local command line (like I was), try this tool:

https://github.com/kubo/snzip#hadoop-snappy-format

UnknownBeef
  • 418
  • 3
  • 17