Post-merge test automation failures

Recently we implemented selenium grid for one of our automated suites. I've written about our reasons for this change, but in short we wanted to improve the speed and stability of our automation. Happily we've seen both those benefits.

We've also seen a noticeable jump in the number of pull requests that are successfully merged back to our master branch each day. This gives some weight to the idea that our rate of application code change was previously impeded by our test infrastructure.

The increase in volume occasionally causes a problem when two feature branches are merged back to master in quick succession. Our tests fail on the second build of the master branch post-merge.

To illustrate, imagine that there are two open pull requests for two feature branches: orange and purple. We can trigger multiple pull request (PR) builds in parallel, so the two delivery teams who are behind these feature branches can receive feedback about their code simultaneously.

When a PR build passes successfully and the code has been through peer review, it can be merged back to the master branch. Each time the master branch changes it triggers the same test suite that executes for a pull request.

We do not trigger multiple builds against master in parallel. If two pull requests are merged in quick succession the first will build immediately and the second will trigger a build that waits for the first to complete before executing. Sometimes the second build will fail.

1. Failing tests after multiple PR merges to master

As the person who had driven sweeping test infrastructure changes, when this happened the first time I assumed that the test automation was somehow faulty. The real issue was that the code changes in orange and purple, while not in conflict with each other at a source code level, caused unexpected problems when put together. The failing tests reflected this.

We hadn't seen this problem previously because our pull requests were rarely merged in such quick succession. They were widely spaced, which meant that when the developer pulled from master to their branch at the beginning of the merge process these type of failures were discovered and resolved.

I raised this as a topic of conversation during Lean Coffee at CAST2016 to find out how other teams move quickly with continuous integration. Those present offered up some possible options to resolve the problem as I described it.

Trunk based development

Google and Facebook move a lot faster than my organisation. Someone suggested that I research these companies to learn about their branching and merging strategy.

I duly found Google's vs Facebook's Trunk Based Development by Paul Hammant and was slightly surprised to see a relevant visualisation at the very top of the article:

2. Google's vs Facebook's Trunk Based Development by Paul Hammant

It seems that, to move very quickly with a large number of people contributing to a code base, trunk-based development is preferred. As the previous diagram illustrates, we currently use a mainline approach with feature branches. This creates larger opportunities for conflicts due to merging.

I had assumed that all possible solutions to these tests failing on master would be a testing-focused. However, a switch to trunk-based development would be a significant change to our practices for every person writing code. I think this solution is too big for the problem.

Sequential build

Someone else suggested that perhaps we were just going faster than we should be. If we weren't running any build requests in parallel and instead triggered everything sequentially, would there still be a problem?

I don't think that switching to sequential builds would fix our issue as the step to trigger the merge is a manual one. A pull request might have successfully passed tests but be waiting on peer review from other developers. In the event that no changes are required by reviewers, the pull request could be merged to master at a time that still creates conflict:

3. Sequential PR build with rapid merge timing

The pull request build being sequential would slow our feedback loop to the delivery teams with no certain benefit.

Staged Build

Another suggestion was to look at introducing an interim step to our branching strategy. Instead of feature branches to master, we'd have a staging zone that might work something like this:

4. Introducing a staging area

The staging branch would use sequential builds. If a test passes there, then it can go to master. If a test fails there, then it doesn't go to master. The theory is that master is always passing.

Where this solution gets a little vague is how the staging branch might automatically rollback a merge. I'm not sure whether it's possible to automatically back changes off a branch based on a test result from continuous integration. If this were possible, why wouldn't we just do this with master instead of introducing an interim step?

I'm relatively sure that the person who suggested this hadn't seen such an approach work in practice.

Do Nothing

After querying the cost of the problem that we're experiencing, the last suggestion that I received was to do nothing. This is the easiest suggestion to implement but one that I find challenging. It feels like I'm leaving a problem unresolved.

However, I know that the build can't always pass successfully. Test automation that is meaningful should fail sometimes and provide information about potential problems in the software. I'm coming to terms with the idea that perhaps the failures we see post-merge are valuable, even though they have become more prevalent since we picked up our pace.

While frustrating, the failures are revealing dependencies between teams that might have been hidden. They also encourage collaboration as people from across the product work together on rapid solutions once the master branch is broken.

While I still feel like there must be a better way, for now it's likely that we will do nothing.



Other posts from CAST2016:

• Fostering professional development
• Human centered test automation

Human centered test automation

The opening keynote at CAST2016 was Nicholas Carr. Though his talk was live streamed, unfortunately a recording is not going to be published. If you missed it, much of the content is available in the material of a talk he delivered last June titled "Media takes command".

Nicholas spoke about typology of automation, the substitution myth, automation complacency, automation bias and the automation paradox. His material focused on the application of technology in traditionally non-technical industries e.g. farming, architecture, personal training.

As he spoke, I started to wonder about the use of automation within software development itself. Specifically, as a tester, I thought about the execution of test automation to determine whether a product is ready to release.

Automation providing answers

Nicholas shared an academic study of a group of young students who were learning about words that are opposite in meaning e.g. hot and cold. The group of students were divided in two. Half of the students received flashcards to study that stated a word with the first letter of it's opposite e.g. hot and c. The other half of the students received flashcards that stated both words in their entirety e.g. hot and cold.

The students who were in the first group performed better in their exam than those in the second group. Academics concluded that this was because when we need to generate an answer rather than simply study an answer, then we are more likely to learn it. This phenomenon is labelled the generation effect.

On the flip side, the degeneration effect is where the answers are simply provided, as in many automated solutions. Nicholas stated that this approach is "a great way to stop humans from developing rich talents".

It's interesting to consider which of these effects are most prevalent in processing the results provided by our continuous integration builds. I believe that the intent of build results is to provide an answer: the build will pass or fail. However, I think the reality of the result is that it can rarely be taken at face value.

I like to confirm that a successful build has truly succeeded by checking the execution time and number of tests that were run. When a build fails, there is a lot of investigative work to determine the real root cause. I dig through test results, log files and screenshots.

I have previously thought that this work was annoying, but in the context of the degeneration effect perhaps the need to manually determine an answer is how we continue to learn about our system. If continuous integration were truly hands-off, what would we lose?

Developing human expertise

Nicholas also introduced the idea of human centered automation. This is a set of five principles by which we see the benefits of automation but continue to develop rich human expertise.

1. Automate after mastery
2. Transfer control between computer and operator
3. Allow a professional to assess the situation before providing algorithmic assistance
4. Don't hide feedback
5. Allow friction for learning

This list is interesting to consider in the context of test automation. The purpose of having test automation is to get fast feedback, so I think it meets the fourth point listed above. But against every other criteria, I start to question our approach.

When I think about the test automation for our products, there is one suite in particular that has been developed over the past five years. This suite has existed longer than most of our testers have been working within my organisation. There is logic coded into the suite for which we no longer have a depth of human expertise. So we do not automate after mastery.

The suite executes without human interaction, there is no transfer of control between computer and operator. Having no tester involvement is a goal for us. Part of providing rapid feedback is reliably provide results within a set period of time regardless of whether there is a tester available.

The suite provides a result. There is human work to assess the result, as I describe above, but the suite has already provided algorithmic assistance that will bias our investigation. It has decided whether the build has passed or a failed.

Finally, the suite is relatively reliable. When the tests usually pass, there is no friction for learning. When the tests are failing and flaky, that is when testers have the opportunity to deeply understand a particular area of the application and associated test code. This happens, but ideally not very much.

So, what are the long term impacts of test automation on our testing skills? Are we forfeiting opportunities to develop rich human expertise in testing by prioritising fast, reliable, automated test execution?

I plan to think more about this topic and perhaps experiment with ways to make our automation more human centered. I'd be curious to hear if other organisations are already exploring in this area.

Other posts from CAST2016:

• Fostering professional development
• Post-merge test automation failures

Fostering professional development

One of the sessions that I attended at CAST2016 was titled "How do I reach the congregation when I'm preaching to the choir?" presented by Rob Bowyer and Erik Davis. One of the main themes of discussion focused on whether people should "sell" professional development to their colleagues or team.

In the introduction to the session, Rob and Erik spoke a little bit about their own contexts. They shared some of the challenges that they've encountered in trying to foster a culture of professional development in both their organisations and their local testing community.

Two particular challenges stuck out for me and I noted them down. Firstly, that "most of the people didn't care" about professional development. Secondly, that "I've been struggling to get people to see the value" in professional development. It struck me that these two challenges in creating a culture of learning could be related.

Do I see value?

I have an ISTQB Foundation certificate. I did this early in my career because I believed that getting this qualification was necessary to find employment in the software testing field. I could see the certificate being mentioned in a lot of job advertisements for testers. 

I saw a clear benefit to me in downloading the syllabus, doing some independent study and taking an exam. This activity was going to open up opportunities in a field of work that I might otherwise be unable to enter. I wanted to be a tester, so I wanted to get the certification.

At that time, I saw the value in this professional development for my career.

On the other hand, I have never completed the BBST Foundation course. I have heard a lot about this qualification and investigated the material that is available online. I have advocated for people in my team to attend this course and published the business benefits I used to argue for this opportunity. But I have not completed the course personally.

I did not learn about BBST Foundation until I had reached a point where I had learned many, but not all, of the concepts in the course via other means. I had heard a lot about the time investment required to complete the course successfully. When given the opportunity to take the class, I decided not to.

At that time, I did not see the value in this professional development for my career.

Do I care?

In the case of ISTQB, a manager might have assumed by my actions that I cared about my professional development. In the case of BBST, a manager might have assumed by my actions that I did not care about my professional development. Both conclusions are reached by assumptions, which are present in any communication.

The Satir Interaction Model describes what happens inside of us as we communicate - the process we go through as we take in information, interpret it, and decide how to respond [Ref].

Ref: "I think we have an issue" -- Delivering unwelcome messages
Fiona Charles

The steps in the Satir Interaction Model between intake and response are hidden. This means that the end result of the process that assigns meaning and significance can be quite surprising to the recipient, which can be a catalyst for conflict.

For example, imagine that I give a manager an input of "I do not want to take the BBST Foundation course". I would be surprised by a response from that manager expressing disappointment that I don't care about my professional development.

We can also climb a Ladder of Inference in our interactions, which refers the idea that there's "a common mental pathway of increasing abstraction, often leading to misguided beliefs". In essence, this is about leaping to conclusions.

For example, imagine the same manager who received my negative response to the BBST Foundation course receiving a promotional email for the RST course. They might extrapolate from my previous negative response that I will not want to attend RST, that I don't want to take any training courses, and that it would be a waste of time to forward me an email that describes this opportunity. I haven't had any input into this flow of reasoning. The manager has independently climbed a ladder of inference.

I think we need to be aware of both of these communication models when assessing an apparent lack of interest in professional development - particularly when we're labeling what we see as "most of the people didn't care".

Empathy & Understanding

Let's return to the question of whether there is a need to sell professional development. I don't think so. However, I agree with an alternative phrasing suggested in the session: that we should foster professional development.

When I sell, I am trying to be persuasive and articulate the merits of an activity. My communication is broadcast oriented. I want to share my reasoning and rationale. I try to explain why people should participate. My intent is to advertise.

That "I've been struggling to get people to see the value" is a failure to sell.

When I foster, I am seeking to encourage the development of an activity through understanding the obstacles that prevent it from happening. I want to be mindful of the ladder of inference and the judgments that I am applying to the responses of my colleagues and team. I want to be aware of where I've assigned significance and meaning might have distorted the message that I have been given, particularly when people are saying "no" to an opportunity.

That "most of the people didn't care" is a failure to foster.

I believe that there are relatively few people who truly don't care about their professional development. If there are people around you who you would label in this manner, I'd challenge you to think about how you have communicated and the responses that you've received.

What have they actually said? What meaning have you prescribed? Have you really understood?

I believe that in reflection and inquiry there is opportunity to successfully foster professional development.



Other posts from CAST2016:

• Human centered test automation
• Post-merge test automation failures