At Calavista we’ve been helping a couple of our clients use Selenium for automated testing of web applications. For one client we’re slowly introducing a slightly different, but still .Net-focused technical stack that allows for much more effective test automation without having to resort to quite so many Selenium tests. For another client we’re trying to help them optimize the execution time of their large Selenium test suite.
At this point, they’re only running the Selenium test suite in a scheduled run overnight, with their testers and developers needing to deal with any test failures the next day. Ideally, they want to get to the point where developers could optionally execute either the whole suite or a targeted subset of the Selenium tests on their own development branches whenever they want.
I think it’s unlikely that we’ll get the full Selenium test suite to where it executes fast enough that a developer would be willing to run those tests as part of their normal “check in dance” routine. To thread the needle a bit between letting a developer get quick feedback from their own local builds or the main continuous integration builds and the desire to run the Selenium suite much more often for faster feedback, we’re suggesting they split the build activity up with what I’ve frequently seen called the “fast build, slow build” pattern (I couldn’t find anybody to attribute this to tonight as I wrote this, but I can’t take credit for it).
First off, let’s assume your project is following the idea of the “testing pyramid” one way or another such that your automated tests probably fall into one of three broad categories:
- Unit tests that don’t touch the database or other external services so they generally run pretty quickly. This would probably include things like business logic rules or validation rules.
- Integration tests that test a subset of the system and frequently use databases or other external services. HTTP contract tests are another example.
- End to end tests that almost inevitably run slowly compared to other types of tests. Selenium tests are notoriously slow and are the obvious example here.
The general idea is to segment the automated build something like this:
- Local developer’s build — You might only choose to compile the code and run fast unit tests as a check before you try to push commits to a GitHub/BitBucket/Azure DevOps/whatever you happen to be using branch. If the integration tests in item #2 are fast enough, you might include them in this step. At times, I’ve divided a local build script into “full” and “fast” modes so I can easily choose how much to run at one time for local commits versus any kind of push (I’m obviously assuming that everybody uses Git by this point, so I apologize if the Git-centric terminology isn’t helpful here).
- The CI “fast build” — You’d run a superset of the local developer’s build, but add the integration tests that run reasonably quickly and maybe a small smattering of the end to end tests. This is the “fast build” to give the developer reasonable assurance that their push built successfully and didn’t break anything
- The CI “slow build” of the rest of the end to end tests. This build would be triggered as a cascading build by the success of the “fast build” on the build server. The “slow build” wouldn’t necessarily be executed for every single push to source control, but there would at least be much more granularity in the tracking from build results to the commits picked up by the “slow build” execution. The feedback from these tests would also be much more timely than running overnight. The segregation into the “fast build / slow build” split allows developers not to be stuck waiting for long test runs before they can check in or continue working, but still get some reasonable feedback cycle from those bigger, slower, end to end tests.