Doug Cutting wrote:
Konstantin Shvachko wrote:
HADOOP-5114. The problem seems to be in ConnectException.
The rest is just how it is revealed by different servers/clients.
If this is indeed just a single bug that's causing 30 tests to fail, I'm
not sure it's yet time to abandon Windows compatibility. It would
surely be nice to, but I suspect there are large numbers of folks who
use Hadoop on Windows but who are not well represented among Hadoop
developers (i.e., who regularly use trunk) and who would be dismayed if
we stopped supporting Windows.
If we really want to pursue this, we should perhaps first conduct a poll
on core-users.
Doug
1. I do think windows as client is important
2. I also think that windows standalone is useful, for people that
havent yet made the jump to unix.
3. I dont know of anyone who runs a production cluster on windows: and
if they do, they havent been reporting the problems which are bound to arise
4. It looks like nobody bothers to test the release regularly on
windows. I think at some point this is something I could set up with our
Hudson server, though I dont like VMWare-image tests as things are more
brittle than raw machine (timing issues, usually). First I need to move
my patches to Git so that hudson can see what I've been doing.
A very large percentage of people's client machines is windows; its only
recently I've moved my laptops over to Linux, and I do think that client
stuff is very important. For OSS applications, that means running the
tests matters, otherwise nobody who uses windows will be able to send in
patches that you will trust, because all the tests will fail.
The other thing is that failing tests are good, they show that the tests
are detecting trouble.
