program crashes with "bzr: ERROR: socket.error: (4, 'Interrupted system call')"

The cause of the EINTR is likely the same as in bug 572055, and the mitigation for that seems to be in 2.1.1 already so not much more we can do there.

The root problem is in socket._fileobject as tracked in <http://bugs.python.org/issue1628205> which is fixed in Python trunk, and the problem with our mitigation is tracked in <http://bugs.python.org/issue8354> which is fixed in Python trunk and the 2.6 branch.

When Debian finally upgrade to 2.6 this problem should vanish for them, but gettting the first Python bug fix backported as well would be good. Otherwise the only options are ripping out the SIGWINCH code, or changing bzrlib.transport.http._urllib2_wrappers not to use socket._fileobject but something more robust.

Martin [gz] wrote:
[...]
> Otherwise the only options are ripping out the SIGWINCH code, or

I'm pretty tempted to do that, and maybe replace it with a small C
module that can install a simple SIGWINCH handler that avoids Python's
bugs. The module would just set a “terminal sized changed” flag that
bzrlib.osutils.terminal_width() would check before returning a result.

I've gone ahead and implemented that C module.

Fix merged into 2.2b3.