[MAPREDUCE-573] reduce scans/copies while reading data in hadoop streaming

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: contrib/streaming
• Labels: newbie
• Assignee: Unassigned
• Reporter: Joydeep Sen Sarma
• Created: 15/Apr/08 00:00
• Updated: 17/Jul/14 21:57

Description

follow up from: http://issues.apache.org/jira/browse/HADOOP-2826

we copy over an entire line (from readLine) and then we break it into two strings by splitting on tab. So there is an extra scan of the input data and an extra copy based on splitting by tab.

instead if we generalized LineReader to instead read until it hits a delimiter - then we can do it with one less scan and copy. Something like:

byte [] tabDelimiter = new byte [1]; tabDelimiter[0] = '\t';
byte [] newlineDelimiter = new byte[2]; newlineDelimiter[0] = '\n'; newlineDelimiter[1] = '\r';

while()

{ lineReader.setDelimiter(tabDelimiter); lineReader.readLine(key); lineReader.setDelimiter(newlineDelimiter); lineReader.readLine(value); }

(take my proposed interfaces with a pinch of salt. just to convey the idea).