socket leak in the test suite again
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
Bazaar |
Confirmed
|
High
|
Unassigned |
Bug Description
Or rather, an old bug is surfacing again leading to:
- occasional failures on Ubuntu (various releases)
http://
is such an occurrence (there are more)
- making the OSX builds fail with 'Too many open files' which
unfortunately doesn't get propagated to jenkins
We did some investigations with gz during the bzr sprint and the
leaked threads are related to the dummy connection used to
shutdown the test server.
There is a race condition there (currently not fully understood
but explaining the above cases).
Since this dummy connection was a hack to avoid involving some
form of thread kill, I think it's not worth diagosing and fixing
the race but instead, bite the bullet and go with the kill server
thread approach instead, especially since we dropped support for
python2.4/2.5.
Related branches
tags: |
added: selftest removed: test |
tags: | added: check-for-breezy |
tags: | removed: check-for-breezy |
Oh, I forgot to mention that the minimal investigation we did with lp:~vila/bzr/socket-leaks confirmed that the sockets were the ones reported by 'lsof -p nnn' on OSX 10.6 while running 'bzr selftest -s bt.per_branch'.