Regression testing with Selenium GRID
- by Ben Adderson
A lot of software teams out there are tasked with supporting and maintaining systems that have grown organically over time, and the web team here at Red Gate is no exception. We're about to embark on our first significant refactoring endeavour for some time, and as such its clearly paramount that the code be tested thoroughly for regressions. Unfortunately we currently find ourselves with a codebase that isn't very testable - the three layers (database, business logic and UI) are currently tightly coupled. This leaves us with the unfortunate problem that, in order to confidently refactor the code, we need unit tests. But in order to write unit tests, we need to refactor the code :S To try and ease the initial pain of decoupling these layers, I've been looking into the idea of using UI automation to provide a sort of system-level regression test suite. The idea being that these tests can help us identify regressions whilst we work towards a more testable codebase, at which point the more traditional combination of unit and integration tests can take over. Ending up with a strong battery of UI tests is also a nice bonus :) Following on from my previous posts (here, here and here) I knew I wanted to use Selenium. I also figured that this would be a good excuse to put my xUnit [Browser] attribute to good use. Pretty quickly, I had a raft of tests that looked like the following (this particular example uses Reflector Pro). In a nut shell the test traverses our shopping cart and, for a particular combination of number of users and months of support, checks that the price calculations all come up with the correct values. [BrowserTheory] [Browser(Browsers.Firefox3_6, "http://www.red-gate.com")] public void Purchase1UserLicenceNoSupport(SeleniumProvider seleniumProvider) { //Arrange _browser = seleniumProvider.GetBrowser(); _browser.Open("http://www.red-gate.com/dynamic/shoppingCart/ProductOption.aspx?Product=ReflectorPro"); //Act _browser = ShoppingCartHelpers.TraverseShoppingCart(_browser, 1, 0, ".NET Reflector Pro"); //Assert var priceResult = PriceHelpers.GetNewPurchasePrice(db, "ReflectorPro", 1, 0, Currencies.Euros); Assert.Equal(priceResult.Price, _browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl01_Price")); Assert.Equal(priceResult.Tax, _browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl02_Tax")); Assert.Equal(priceResult.Total, _browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl02_Total")); } These tests are pretty concise, with much of the common code in the TraverseShoppingCart() and GetNewPurchasePrice() methods. The (inevitable) problem arose when it came to execute these tests en masse. Selenium is a very slick tool, but it can't mask the fact that UI automation is very slow. To give you an idea, the set of cases that covers all of our products, for all combinations of users and support, came to 372 tests (for now only considering purchases in dollars). In the world of automated integration tests, that's a very manageable number. For unit tests, it's a trifle. However for UI automation, those 372 tests were taking just over two hours to run. Two hours may not sound like a lot, but those cases only cover one of the three currencies we deal with, and only one of the many different ways our systems can be asked to calculate a price. It was already pretty clear at this point that in order for this approach to be viable, I was going to have to find a way to speed things up. Up to this point I had been using Selenium Remote Control to automate Firefox, as this was the approach I had used previously and it had worked well. Fortunately, the guys at SeleniumHQ also maintain a tool for executing multiple Selenium RC tests in parallel: Selenium Grid. Selenium Grid uses a central 'hub' to handle allocation of Selenium tests to individual RCs. The Remote Controls simply register themselves with the hub when they start, and then wait to be assigned work. The (for me) really clever part is that, as far as the client driver library is concerned, the grid hub looks exactly the same as a vanilla remote control. To create a new browser session against Selenium RC, the following C# code suffices: new DefaultSelenium("localhost", 4444, "*firefox", "http://www.red-gate.com"); This assumes that the RC is running on the local machine, and is listening on port 4444 (the default). Assuming the hub is running on your local machine, then to create a browser session in Selenium Grid, via the hub rather than directly against the control, the code is exactly the same! Behind the scenes, the hub will take this request and hand it off to one of the registered RCs that provides the "*firefox" execution environment. It will then pass all communications back and forth between the test runner and the remote control transparently. This makes running existing RC tests on a Selenium Grid a piece of cake, as the developers intended. For a more detailed description of exactly how Selenium Grid works, see this page. Once I had a test environment capable of running multiple tests in parallel, I needed a test runner capable of doing the same. Unfortunately, this does not currently exist for xUnit (boo!). MbUnit on the other hand, has the concept of concurrent execution baked right into the framework. So after swapping out my assembly references, and fixing up the resulting mismatches in assertions, my example test now looks like this: [Test] public void Purchase1UserLicenceNoSupport() { //Arrange ISelenium browser = BrowserHelpers.GetBrowser(); var db = DbHelpers.GetWebsiteDBDataContext(); browser.Start(); browser.Open("http://www.red-gate.com/dynamic/shoppingCart/ProductOption.aspx?Product=ReflectorPro"); //Act browser = ShoppingCartHelpers.TraverseShoppingCart(browser, 1, 0, ".NET Reflector Pro"); var priceResult = PriceHelpers.GetNewPurchasePrice(db, "ReflectorPro", 1, 0, Currencies.Euros); //Assert Assert.AreEqual(priceResult.Price, browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl01_Price")); Assert.AreEqual(priceResult.Tax, browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl02_Tax")); Assert.AreEqual(priceResult.Total, browser.GetText("ctl00_content_InvoiceShoppingItemRepeater_ctl02_Total")); } This is pretty much the same as the xUnit version. The exceptions are that the attributes have changed, the //Arrange phase now has to handle setting up the ISelenium object, as the attribute that previously did this has gone away, and the test now sets up its own database connection. Previously I was using a shared database connection, but this approach becomes more complicated when tests are being executed concurrently. To avoid complexity each test has its own connection, which it is responsible for closing. For the sake of readability, I snipped out the code that closes the browser session and the db connection at the end of the test. With all that done, there was only one more step required before the tests would execute concurrently. It is necessary to tell the test runner which tests are eligible to run in parallel, via the [Parallelizable] attribute. This can be done at the test, fixture or assembly level. Since I wanted to run all tests concurrently, I marked mine at the assembly level in the AssemblyInfo.cs using the following: [assembly: DegreeOfParallelism(3)] [assembly: Parallelizable(TestScope.All)] The second attribute marks all tests in the assembly as [Parallelizable], whilst the first tells the test runner how many concurrent threads to use when executing the tests. I set mine to three since I was using 3 RCs in separate VMs. With everything now in place, I fired up the Icarus* test runner that comes with MbUnit. Executing my 372 tests three at a time instead of one at a time reduced the running time from 2 hours 10 minutes, to 55 minutes, that's an improvement of about 58%! I'd like to have seen an improvement of 66%, but I can understand that either inefficiencies in the hub code, my test environment or the test runner code (or some combination of all three most likely) contributes to a slightly diminished improvement. That said, I'd love to hear about any experience you have in upping this efficiency. Ultimately though, it was a saving that was most definitely worth having. It makes regression testing via UI automation a far more plausible prospect. The other obvious point to make is that this approach scales far better than executing tests serially. So if ever we need to improve performance, we just register additional RC's with the hub, and up the DegreeOfParallelism. *This was just my personal preference for a GUI runner. The MbUnit/Gallio installer also provides a command line runner, a TestDriven.net runner, and a Resharper 4.5 runner. For now at least, Resharper 5 isn't supported.
