Affects Version/s: None
Fix Version/s: 0.12.0
Windows without Cygwin as a whole, but with some key binaries such as perl, diff, gawk, gzip, sed.
For pig to really work well on Windows, it needs hadoop core changes. Right now, those are in progress in branch-1-win. For this work, I am running Pig on Windows against branch-1-win and removing Cygwin dependencies as capabilities open up. Branch-1-win is fairly stable now, and has opened up enough functionality to see the few things needed in Pig to run E2E on top of a cross-platform Hadoop core without Cygwin. This uber-JIRA should track the whole of the work to get pig running well on Windows without Cygwin.
There are a few types of work that I think are needed right now (will break-out sub-jiras to track them):
1.) Tests that generate pig script strings with paths in them (e.g. dynamically build load/store commands) need to have Pig escape ("\") characters encoded – as they can now occur in both Hadoop and local paths.
2.) Tests that generate local temporary files with createTempFile, and then try to use those as HDFS paths need to remove ":" from the generated file name to create valid HDFS paths.
3.) Tests that hand-generate URIs via string concatenation (e.g. "file:" + strFileName) need to use Util.generateURI instead to get a valid URI for the target platform.
4.) Tests that assume the first line in a script (e.g. #!/bin/sh) auto-resolves interpreters need to explicitly call the interpreter (e.g. instead of calling "perlscript.pl" they should call "perl perlscript.pl".
5.) Changes in quotes or command syntax between shells (e.g. " or ', dir or ls) need to be tuned a little here and there.
1.) The streaming interface needs to be fixed to run without a Cygwin dependency.
2.) The pig.additional.jars separator is currently hardcoded to ":", and should be File.pathSeparator instead (":" on linux, ";" on Windows) to be able to accept Windows paths (C:\file.jar for instance).
3.) The Grunt "sh" command highly surfaces the behavior of the exec API. If you use a built-in, it fails with file not found. This surfaces a lot of differences in shell implementation differences (e.g. ls is an exe, but dir is builtin) – and many of the cases in TestGrunt end up running (sh bash -c "command"). For portability and ease of use, sh should actually exec "sh -c <command> on Linux and "cmd /C <command>" on Windows to improve usability and make it possible to use aliases and bat files on either platform to make the interface more platform independent to end-users.
4.) (eventual) Update Pig's dependencies to pick up a stable Hadoop core that runs on Windows from a release branch.