Details
-
New Feature
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
-
None
Description
An InputFormat and OutputFormat for zip-format archives would be useful. It would provide a standard file format, friendly to foreign applications (e.g., streaming), that is also optimized for Hadoop.
The OutputFormat could generate archives containing a series of small (~1MB) files, each individually compressed. The InputFormat could then generate a split for each file in the archive. (Zip archives contain a directory at the end of the file that may be efficiently read when constructing splits.) Input splits could thus be sent to a node where they are stored locally for the map phase, providing good performance.
This would thus permit applications to (1) use a standard file format; (2) keep data compressed; and (3) efficiently split input into chunks substantially smaller than HDFS blocks. This is not available today.