Remi, thanks for the detailed post-mortem, it's a good read and great learning. I hope everyone reads it.
The one thing to emphasize is that we now have a very visible way to get code into master, we have folks investing time to provide review (great), we need the submitters to make due diligence and answer all comments in the reviews. In another project i work on, nothing can be added to the code without unit tests. I think we could go down the route of asking for new integration tests and unit tests for anything. If not, the PR does not get merged. But let's digest your post-mortem and we can discuss after 4.6.0. I see that you reverted one commit that was not made by you, that's great. Let's focus on the blockers now, everything else can wait. The big bonus of doing what we are doing is that once 4.6.0 is out, we can merge PRs again (assuming they are properly rebased and tested) and we can release 4.6.1 really quickly after. -sebastien On Sep 25, 2015, at 9:51 PM, Remi Bergsma <rberg...@schubergphilis.com> wrote: > Hi all, > > This mail is intended to be blameless. We need to learn something from it. > That's why I left out who exactly did what because it’s not relevant. There > are multiple examples but it's about the why. Let's learn from this without > blaming anyone. > > We know we need automated testing. We have integration tests, but we are > unable to run all of them on any Pull Request we receive. If we would have > that in place, it'd be much easier to spot errors, regression and so on. It'd > also be more rewarding to write more tests. > > Unfortunately we're not there yet. So, we need to do something else instead > until we get there. If we do nothing, we know we have many issues because a > master that breaks on a regular basis is the most frustrating things. We said > we'd use Pull Requests with at least two humans to review and give their OK > for a Pull Request. In the form of LGTM: Looks Good To Me. Ok, so the LGTMs > are there because we have no automated testing. Keep that in mind. You are > supposed to replace automated testing until it's there. > > Since we do this, master got a lot more stable. But every now and then we > still have issues. Let's look at how we do manual reviews. Again, this is not > to blame anyone. It's to open our eyes and make us realise what we're doing > and what results we get out of that. > > > Example Pull Request #784: > Title: CLOUDSTACK-8799 fixed the default routes > > That's nice, it has a Jira id and a short description (as it should be). > > The first person comes along and makes a comment: > "There was also an issue with VPC VRs" ... "Have you seen this issue? Does > your change affects the VPC VR (single/redundant)?" > > Actually a good question. Unfortunaly there comes no answer. After a > reminder, it was promised to do tests against VPC networks. Great! > > The Jenkins builds both succeed and also Travis is green. But how much value > does this have? They have the impression to do automated testing, and > although you could argue they do, it's far from complete. If it breaks, you > know you have an issue. But it doesn’t work the other way around. > > Back to our example PR. In the mean time, another commit gets pushed to it: > "CLOUDSTACK-8799 fixed for vpc networks." But if you look at the Jira issue, > you see it is about redundant virtual routers. The non-VPC ones. So this is > vague at best. But a reviewer gives a LGTM because the person could create a > VPC. That doesn't have anything to do with the problem being fixed in this PR > nor with the comments made earlier. But, at least the person said what he did > and we should all do that. What nobody knew back then, was that this broke > the default route on VPCs. > > Then something strange happens: the two commits from the PR end up on master > as direct commits. With just one LGTM and no verification from the person > commenting about the linked issue. This happened on Friday September 11th. > > That day 21 commits came in, from 7 Pull Request and unfortunately also from > some direct commits. We noticed the direct commits and notified the list > (http://cloudstack.markmail.org/message/srmszloyipkxml36). As a lot came in > at the same time, it was decided not to revert them. Looking back, we should > have done it. > > From this point on, VPCs were broken as they wouldn't get a default route. > So, no public internet access from VMs in VPC tiers, no VPNs working, etc. > This was mentioned to the list on Thursday September 15th, after some chats > and debugging going on over the weekend > (http://cloudstack.markmail.org/message/73ulpu4p75ex24tc) > > Here we are, master is broken functionality wise and new Pull Requests come > in to fix blockers. But we cannot ever test their proper working, because > VPCs are broken in master and so also in the PRs branched off of it. With or > without change in the PR. > > It starts to escalate as the days go by. > > I’ll leave out the bit on how this frustrated people. Although it’s good to > know we do not want to be in this situation. > > Eventually Wilder and I spent an evening and a day working on a branch where > we loaded 7 PRs on top of each other (all VR related) only to find the VPC is > still broken. It allowed us to zoom in and find the default route was missing > again. We said it worked 3 weeks before, because the same tests that > succeeded then, now were broken. We had already fixed this in PR #738 on > August 25 so were sure about it. > > After some digging we could trace it back to Pull Request #784. Imagine the > feeling seeing your own comment there mentioning the previous issue on the > default gateways. Fair to say our human review process clearly failed here. > Many many hours were spent on this problem over the past two weeks. Could we > have prevented this from happening? I think so, yes. > > > This example clearly shows why: > > - we should use Pull Requests > It made the change visible: Great! > > - we do reviews and ask for feedback > We got feedback and questions: Also great! > > - we should always respond to feedback and verify it is resolved, before > merging > We need to improve here. Even with two reviewers that say LGTM, we should > still address any feedback before merging. > > - we should have two humans doing a review > We need to improve here as well. Not one reviewer, we need two. Really. > > - we need to document why we say LGTM. > Another improvement. It’s nice to say LGTM, but a review of only 4 > characters and nothing more is useless. We need to know what was tested and > how. Test results, screen shots or anything that shows what's been verified. > If you only reviewed the code, also fine but at least say that. Then the next > reviewer should do another type of review to get the comlete picture. > Remember you're replacing automated testing! > > - we should always merge Pull Requests > We made it easy, merging is the de facto standard, and it has even more > benefits. You can trace commits back to their Pull Request (and find all > comments and discussion there: saves time, trust me). It also allows for > easier reverting of a Pull Request. We’ll see even more benefits once 4.7 is > there. Although the intentions to merge the Pull Request were there, it still > didn't happen. We should always check before we push. As a committer we just > need to be sure. > > - we need automated testing! > The sooner the better. It’s all about the missing automated testing. After > 4.6, we all need to focus on this. Saves a lot of time. And frustrations. > > > > We're doing final testing on PR #887 and will merge it soon. From that point > on we can look into new issues. Be aware that any PR out there that was > created after September 10 needs to be rebased with current master (when #887 > is merged). Without that, no serious testing can be done. > > Let's be careful what to land on master. I'll only be merging Pull Requests > that have had proper reviews with information on what was tested. At least > one reviewer needs to actually verify it works (and show the rest of us). We > simply cannot assume it will work. > > If we do this, I think we can start resolving the remaining blockers > one-by-one and go into the first RC round. Please help out where you can so > we can make this a success together. Thanks! > > Looking forward to the day we have our automated testing in place ;-) > > Regards, > Remi >
