Dear Greg, and anyone else that is involved in deployment
This is a follow-up from Dan Duvall's talk today during the metrics meeting about voting browser tests.
Background: The reading web team this quarter with the help of Dan Duvall has made huge strides in our QA infrastructure. The extensions Gather, MobileFrontend and now the new extension QuickSurveys are all running browser tests on a per commit basis. A selected set of MobileFrontend @smoke tests (a selected subset of all he tests) are running in 15minutes on every commit and the entire set of Gather browser tests is running in around 21minutes. It marginally slows down getting patches deployed... but I think this is a good thing. The results speak for themselves.
In the past month (August 4th-September 4th) only 3/33 builds failed for MobileFrontend's daily smoke test build [1] (all 3 due to issues with the Jenkins infrastructure). For the full set of tests only 10/33 failed in the Chrome daily build [3], 8 of which were due to tests being flakey and needing improvement or issues with the Jenkin infrastructure and the two others serious bugs [4,5] brought about by work the performance team had been doing that we were able to fix shortly after.
In Firefox [2] there were only 6 failures and only 2 of these were serious bugs, again caused by things outside MobileFrontend [4,6]. One of these was pretty serious - we had started loading JavaScript for users with legacy browsers such as IE6. These were caught prior to the daily builds when suddenly our MobileFrontend commits would not merge.
The future!: Given this success: 1) I would like to see us run @integration tests on core, but I understand given the number of bugs this might not be feasible so far. 2) We should run @integration tests prior to deployments to the cluster via the train and communicate out when we have failures (and make a decision to push broken code) 3) I'd like to see other extensions adopt browser test voting on their extensions. Please feel free to reach out to me if you need help with that. The more coverage across our extensions we have, the better.
We really have no excuse going forward to push broken code out to our users and at the very least we need to be visible to each other when we are deploying broken code. We have a responsibility to our users.
Thoughts? Reactions? Who's with me?!
[1] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [2] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [3] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [4] https://phabricator.wikimedia.org/T108045 [5] https://phabricator.wikimedia.org/T108191 [6] https://phabricator.wikimedia.org/T111233
Looping in QA list and dropping our team private list.
<quote name="Jon Robson" date="2015-09-03" time="11:45:47 -0700">
Dear Greg, and anyone else that is involved in deployment
This is a follow-up from Dan Duvall's talk today during the metrics meeting about voting browser tests.
Background: The reading web team this quarter with the help of Dan Duvall has made huge strides in our QA infrastructure. The extensions Gather, MobileFrontend and now the new extension QuickSurveys are all running browser tests on a per commit basis. A selected set of MobileFrontend @smoke tests (a selected subset of all he tests) are running in 15minutes on every commit and the entire set of Gather browser tests is running in around 21minutes. It marginally slows down getting patches deployed... but I think this is a good thing. The results speak for themselves.
In the past month (August 4th-September 4th) only 3/33 builds failed for MobileFrontend's daily smoke test build [1] (all 3 due to issues with the Jenkins infrastructure). For the full set of tests only 10/33 failed in the Chrome daily build [3], 8 of which were due to tests being flakey and needing improvement or issues with the Jenkin infrastructure and the two others serious bugs [4,5] brought about by work the performance team had been doing that we were able to fix shortly after.
In Firefox [2] there were only 6 failures and only 2 of these were serious bugs, again caused by things outside MobileFrontend [4,6]. One of these was pretty serious - we had started loading JavaScript for users with legacy browsers such as IE6. These were caught prior to the daily builds when suddenly our MobileFrontend commits would not merge.
The future!: Given this success:
- I would like to see us run @integration tests on core, but I
understand given the number of bugs this might not be feasible so far. 2) We should run @integration tests prior to deployments to the cluster via the train and communicate out when we have failures (and make a decision to push broken code) 3) I'd like to see other extensions adopt browser test voting on their extensions. Please feel free to reach out to me if you need help with that. The more coverage across our extensions we have, the better.
We really have no excuse going forward to push broken code out to our users and at the very least we need to be visible to each other when we are deploying broken code. We have a responsibility to our users.
Thoughts? Reactions? Who's with me?!
[1] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [2] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [3] https://integration.wikimedia.org/ci/view/Mobile/job/browsertests-MobileFron... [4] https://phabricator.wikimedia.org/T108045 [5] https://phabricator.wikimedia.org/T108191 [6] https://phabricator.wikimedia.org/T111233
Mobile-l mailing list Mobile-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/mobile-l
On 09/03/2015 02:45 PM, Jon Robson wrote:
The future!: Given this success:
- I would like to see us run @integration tests on core, but I
understand given the number of bugs this might not be feasible so far. 2) We should run @integration tests prior to deployments to the cluster via the train and communicate out when we have failures (and make a decision to push broken code) 3) I'd like to see other extensions adopt browser test voting on their extensions. Please feel free to reach out to me if you need help with that. The more coverage across our extensions we have, the better.
+100% I assume #2 should be "make a decision whether to push broken code".
Matt Flaschen
Just to hop on the bandwagon here: this seems like the only sane path going forward. One unmentioned benefit is that this is a step toward continuous deployment. Having integration tests run on every commit and then block when there are failures is pretty much a requirement if Wikimedia ever wants to get there.
On Thu, Sep 3, 2015 at 1:43 PM Pine W wiki.pine@gmail.com wrote:
I just want to say that I appreciate this overview.
Pine _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
In the services team, we found that prominent coverage metrics are a very powerful motivator for keeping tests in order. We have set up 'voting' coverage reports, which fail the overall tests if coverage falls, and make it easy to check which lines aren't covered yet (via coveralls). In all repositories we enabled this for, test coverage has since stabilized around 80-90%.
Gabriel
On Thu, Sep 3, 2015 at 4:31 PM, Steven Walling steven.walling@gmail.com wrote:
Just to hop on the bandwagon here: this seems like the only sane path going forward. One unmentioned benefit is that this is a step toward continuous deployment. Having integration tests run on every commit and then block when there are failures is pretty much a requirement if Wikimedia ever wants to get there.
On Thu, Sep 3, 2015 at 1:43 PM Pine W wiki.pine@gmail.com wrote:
I just want to say that I appreciate this overview.
Pine _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
<quote name="Gabriel Wicke" date="2015-09-03" time="17:03:03 -0700">
In the services team, we found that prominent coverage metrics are a very powerful motivator for keeping tests in order. We have set up 'voting' coverage reports, which fail the overall tests if coverage falls, and make it easy to check which lines aren't covered yet (via coveralls). In all repositories we enabled this for, test coverage has since stabilized around 80-90%.
We (RelEng), too, are interested in this. Given the nature of our projects we'll probably need to start this on a case-by-case basis, (un)fortunately. :)
There's two parts to this (as I see it): informational and enforcement.
Informational: * "Generate code coverage reports for extensions" ** https://phabricator.wikimedia.org/T71685 * Add ^^^ to "QA Health scoreboard" ** https://phabricator.wikimedia.org/T108768
Enforcement: * What Gabriel described above. ** There's no one ticket for tracking this cross repos right now, I'll create one... ** https://phabricator.wikimedia.org/T111546
Greg
PS: I didn't mean to, but I forked this thread across wikitech-l and qa lists (my bcc to wikitech-l didn't make it through mailman, I don't think). See the other sub-thread on adding @integration test runs on wmf deploy branch creation at: https://lists.wikimedia.org/pipermail/qa/2015-September/thread.html
(I've put the TO: field as the QA list only, and put everyone else on BCC now. If you're curious in this topic, please join the QA mailing list and follow along. It's not a very high traffic list.)
<quote name="Jon Robson" date="2015-09-03" time="11:45:47 -0700">
Dear Greg, and anyone else that is involved in deployment
Hi there :)
<successes over the past month>
Awesome :)
The future!: Given this success:
- I would like to see us run @integration tests on core, but I
understand given the number of bugs this might not be feasible so far.
https://integration.wikimedia.org/ci/view/BrowserTests/view/Core/
Those are pretty stable, but limited (rightfully) in number. Looks like the last run of the first job took 8 minutes. Not too bad.
- We should run @integration tests prior to deployments to the
cluster via the train and communicate out when we have failures (and make a decision to push broken code)
The way I hear this is: Run @integration tests on merge to wmfXX branches. Is that an accurate rephrasing?
If not, then it mean what we're planning on doing with respect to the "Staging" cluster work we started, paused (due to time constraints), and will restart in Q3. tl;dr: The staging cluster will be a cluster that runs nightly full blown tests against a git tag. In the morning we'll be able to make a go/no-go decision on deploying that tag.
That "nightly" part can, of course, be modified to whatever frequency we can support (which can probably be pretty fast).
- I'd like to see other extensions adopt browser test voting on their
extensions. Please feel free to reach out to me if you need help with that. The more coverage across our extensions we have, the better.
Thanks for the offer of help, Jon! That's awesome. I love the idea of teams/groups helping other teams/groups! You, personally, have been great this so far and I thank you for that.
Greg
On Thu, Sep 3, 2015 at 8:45 PM, Jon Robson jrobson@wikimedia.org wrote:
This is a follow-up from Dan Duvall's talk today during the metrics meeting about voting browser tests.
If you did not see it (34:30-44:30):
https://youtu.be/Hy307xn99-c?t=34m26s
Please notice the explanation of release engineering team, by Evil Greg: Delivering deliverables delivery since our delivery.
I am getting that tattooed. :)
Željko
wikitech-l@lists.wikimedia.org