[openstack-dev] [all] testtools 1.2.0 release breaks the world
robertc at robertcollins.net
Sat Nov 15 20:06:18 UTC 2014
On 16 November 2014 03:25, Sean Dague <sean at dague.net> wrote:
> Testtools 1.2.0 release apparently broke subunit.run discover --list
> which breaks the way that tempest calls the test chain. No tempest runs
> have passed since it's release.
> https://review.openstack.org/#/c/134705/ is a requirements pin, though I
> think because of grenade this is actually going to have to be laddered
> up from icehouse => juno => master.
> It's probably half a day babysitting getting all the pins in place to
> make the world work again. I'm offline from here on out for the weekend,
> but I'll put a +2/+A on all of these so if someone wants to recheck them
> in the right order to land them, they can get things fixed.
> Also... lets try not to release libraries on Fridays before disappearing
> for the weekend... please. Pretty please.
So this has tweaked me - I'm going to rant a little here.
It wasn't Friday - it was saturday, when I had some time to do
personal stuff and instead chose to push forward on a long arc that
has been blocking oslo.db changes for a couple months. I released
knowing I'd need to be around to follow up, and the very first thing I
did when I got up Sunday morning after dealing with nappies etc was to
check for fallout.
The release was thoroughly tested upstream, and I explicitly tested it
with OpenStack trees as well.
I didn't disappear for the weekend, and I didn't even disappear
straight to bed... and given how well you know me, you know that I
rarely do disappear fully *anyway*, and finally - I'm reachable 24x7
if things really need that (but no-one rang me so clearly its not
panic button time) - nor did anyone ask any of the other other
testing-cabal committers to do an urgent action.
And, and this is perhaps the most annoying aspect, noone in OpenStack
tried to reproduce this upstream (which is a 4 command test - make a
virtualenv, pip install, cd to a tree, perform discovery) and if they
had, such a hypothetical person would have seen that the issue *isn't*
testtools, its the use of system-site-packages permitting the old
(0.5.1 - current is 0.8.0) unittest2 causing the issue. And thats
something thats entirely in the OpenStack space to fix - not upstream.
So by going 'oh its testtools problem', 8 or so hours when it could
have been fixed have passed with noone looking at it.
I totally get the omg its broken and that that makes everyone unhappy
when it happens. However I don't like the feeling of being accused of
irresponsibility when I was still around for some time (just not at
02:58 am when I was first pinged (< sdague> lifeless:
https://review.openstack.org/#/c/134705/ should you actually be
checking in this weekend).
I think everyone is aware that releases have some risk, and doing them
cavalierly would be bad. But running with stale dependencies isn't
going to be part of the test matrix upstream, since the entire project
goal is to fold all its learning into upstream: as things move
upstream we have to depend on newer releases of those components (via
things like unittest2, importlib2, traceback2 etc etc).
Robert Collins <rbtcollins at hp.com>
HP Converged Cloud
More information about the OpenStack-dev