Mocking WebService consumed by a Biztalk Request-Response port

Question

I'm using BizUnit to unit-tests my Biztalk orchestrations, but some orchestrations consume a WebService,and testing these seems more like integration testing than unit testing.

I'm familiar with using a mocking framework to mock the generated proxy objects, in order to test a web service from a Windows Forms application, but I would like to be able to do it in a more integrated way in a request-response port?

How would you approach this problem?

Answer 1

This goes to the heart of one of my main irritations as a BizTalk developer - BizTalk does not lend it self to unit testing. From the fact the 99% of your interfaces into BizTalk applications are message based and have a huge number of possible inputs, through to the opaque nature of orchestrations, BizTalk offers no real way of testing units of functionality as... well... units.

For BizTalk, integration tests are sadly often the only game in town.

That results in, due to no fault on the part of Kevin Smith, BizUnit being (IMO) a misnomer. A better name would perhaps be BizIntegrationIt. BizUnit offers a range of tools that assist in integration testing, the majority of its tests, like checking if a file has been written to a given directory or sending an HTTPRequest to a BizTalk HTTPReceive location are all strictly speaking, testing integration.

Now that I've gotten that rant out, what you are asking for is something I've been thinking about for a long time, the ability to create automated unit tests that give some real confidence that my making a small change to a map won't suddenly break something else downstream, as well as a way to remove dependance on external services.

I've never thought of any nice way of doing this but below is a solution that should work, I've done variations of each part of this in isolation but never tried to but them all together in this specific form.

So given the desire to mock a call to some external service (that may not even exist yet) without needing to actually make any external call and wanting to have the ability to set expectations for that service call and to specify the nature of the response, the only method I can think of is to develop a custom adapter.

Mock webservice using custom adapter

If you build a custom request-response adapter you can plug it into your send port in place of the SOAP adapter. You can then specify properties for the adapter that allow it to behave as a mock of your webservice. The adapter would be similar in concept to a loopback adapter but would allow internal mocking logic.

Things that you might want to include as adapter properties:

• Expected document (perhaps a disk location that specifies an example of what you expect your BizTalk applicaiton to send to the webservice).
• Response document - the document that the adapter will send back to the messaging engine.
• Specific expectations for the test such as lookup values in document elements.

You could also have the custom adapter write to disk and setup a BizUnit step to validate the file that was written out.

Building a custom adapter is non-trivial, but possible, you can get a good start from the BizTalk Adapter Wizard and there is an article on deploying custom adapters here.

There is a bug in the code generated by the wizard, you will need to change new Guid(""), to new Guid().

There are also some examples of building custom adapters in the BizTalk SDK.

Another option is to use a plain http page and the HTTP solicit response as discussed here, all your logic goes in the http page. This is probably simpler if you are happy having an http call, and setting up an IIS port to listen for your test.

Initialising unit tests

You can import binding files into a BizTalk application using a .bat file.

If you make a new binding file for each test you run, as well as for your standard applicaiton set up, you can then run the appropriate batch file to apply the right binding.

Each binding file would change your webservice sendport to use the mock custom adapter and set the specific properties for that test.

You could then even make a custom BizUnit step that (perhaps) generated binding settings based on settings in the test step and then ran the shell commands to update the bindings.

Testing Message Contents

A final thing that you might want to consider, to really tie all this together, is some way of testing the contents of messages. You could do this in your mock adapter, but that would get tedious very quickly for large messages, or for a large range of possible input messages.

One option is to make a custom pipeline that calls Schematron to validate files that it receives. Schematron is a schema language that allows a much richer level of file inspection that xsd, so you can check things like "If element x contains this content, I expect element y to be present".

If you built a custom pipeline that took a schematron schema as a parameter, you could then swap in a testing file for a specific unit test, validating that for this test, when you call the webservice you get a file that actually matches what you want (and doesn't just match the xsd)

Answer 2

As a co-author of BizUnitExtensions (www.codeplex.com/bizunitextensions) i agree that the name "unit" in BizUnit can be confusing but for Biztalk, the 'integration test' is the unit test. Some Biztalk folk have successfully used mocks to test pipeline components and other test harnesses (+ BizUnit/Extensions) to test schemas and maps.

Orchestrations unfortunately are opaque. But theres are good reasons for that.

(a) Because of the huge subscription system in the message box - that orchestrations use when being activated etc, it is not possible to fire up some "virtual" process to host the orchestration (which can be done for pipelines. Tomas Restrepo has done something along these lines).

(b) Also, how would this virtual process handle persistence and dehydration?. I'd wager that people using WF would have the same problem in trying to test the workflow fully.

(c) we dont work with the C# directly, so there is no way we can "inject" a mock interface into the orchestration code.

(d) An orchestration is not really a "unit". its a composite element. The units are the messages going to and from the message box and the external components called through expression shapes.So even if you could inject a mock webservice interface you cannot inject mock message boxes and correlation sets and other things.

One thing that can be done for orchestrations (and i've been considering an addition to the BizUnitExtensions library to do this) is to link in with the OrchestrationProfiler tool as that tool gives a pretty detailed report of all the shapes and somehow check that individual steps were executed (and perhaps the time it took for execution). This could go quite far in making the orchestration a bit more of a white box.Also considering that the orchestration debugger shows a lot of the variable values, surely it must be possible to get that info via an API to show what the values of variables were at a given point for a given instance.

Back to Richard's question though, my previous dev team had a solution. Basically what we did was to write a generic configurable HttpHandler that parsed incoming service requests and returned pre-set responses. The response sent back was configurable based on conditions such as XPath. In the BUILD and DEV binding files, the webservice end point was the mock. This worked brilliantly in isolating the BUILD and DEV environments from the actual third party webservices. This also helped in a "contract first" approach where we built the mock and the orch developer used it while the webservice author went ahead and built the actual service.

[Update:17-FEB-09: this tool is now on codeplex : http://www.codeplex.com/mockingbird. If this approach sounds interesting check it out and let me know what you think of the tool ]

Now, before someone throws the old "WHAT ABOUT MOCK OBJECT FRAMEWORKS" chestnut in, let me say that the utility above was used for both Biztalk 'consumers' as well as non Biztalk consumers, BUT i have also worked with NMock2 and found that to be an excellent way to mock interfaces and set expectations when writing CLR consumers. (I'm going to be looking into MoQ and TypeMock etc soon). However, it wont work with orchestrations for the reasons described above.

Hope this helps.

Regards,

Benjy

Answer 3

Don't.

Don't test against arbitrary interfaces, and don't create mocks for them.

Most people seem to see developer (unit) testing as intended for testing nontrivial, individual units of functionality such as a single class. On the other hand, it is also important to perform customer (acceptance/integration) testing of major subsystems or the entire system.

For a web service, the nontrivial unit of functionality is hidden in the classes that actually perform the meaningful service, behind the communication wiring. Those classes should have individual developer test classes that verify their functionality, but completely without any of the web-service-oriented communication wiring. Naturally, but maybe not obviously, that means that your implementation of the functionality must be separate from your implementation of the wiring. So, your developer (unit) tests should never ever see any of that special communication wiring; that is part of integration and it can be viewed (appropriately) as a "presentation" issue rather than "business logic".

The customer (acceptance/integration) tests should address a much bigger scale of functionality, but still not focused on "presentation" issues. This is where the use of the Facade pattern is common--exposing a subsystem with a unified, coarse-grained, testable interface. Again, the web service communication integration is irrelevant and is implemented separately.

However, it is very useful to implement a separate set of tests that actually do include the web service integration. But I strongly recommend against testing only one side of that integration: test it end-to-end. That means building tests that are web service clients just like the real production code; they should consume the web services exactly the way that the real application(s) do(es), which means that those tests then serve as examples to anyone who must implement such applications (like your customers if you are selling a library).

So, why go to all that trouble?

1. Your developer tests verify that your functionality works in-the-small, regardless of how it is accessed (independent of presentation tier since it is all inside the business logic tier).

2. Your customer tests verify that your functionality works in-the-large, again regardless of how it is accessed, at the interface boundary of your business logic tier.

3. Your integration tests verify that your presentation tier works with your business logic tier, which is now managable since you can now ignore the underlying functionality (because you separately tested it above). In other words, these tests are focused on a thin layer of a pretty face (GUI?) and a communication interface (web services?).

4. When you add another method of accessing your functionality, you only have to add integration tests for that new form of access (presentation tier). Your developer and customer tests ensure that your core functionality is unchanged and unbroken.

5. You do not need any special tools, such as a test tool specifically for web services. You use the tools/components/libraries/techniques that you would use in production code, exactly as you would use them in such production code. This makes your tests more meaningful, since you are not testing someone else's tools. It saves you lots of time and money, since you are not buying, deploying, developing for, and maintaining for a special tool. However, if you are testing through a GUI (don't do that!), you might need one special tool for that part (e.g., HttpUnit?).

So, let's get concrete. Assume that we want to provide some functionality for keeping track of the cafeteria's daily menu ('cause we work in a mega-corp with its own cafe in the building, like mine). Let's say that we are targeting C#.

We build some C# classes for menus, menu items, and other fine-grained pieces of functionality and its related data. We establish an automated build (you do that, right?) using nAnt that executes developer tests using nUnit, and we confirm that we can build a daily menu and look at it via all these little pieces.

We have some idea of where we are going, so we apply the Facade pattern by creating a single class that exposes a handful of methods while hiding most of the fine-grained pieces. We add a separate set of customer tests that operate only through that new facade, just as a client would.

Now we decide that we want to provide a web page for our mega-corp knowledge workers to check today's cafeteria menu. We write an ASP.NET page, have it invoke our facade class (which becomes our model if we are doing MVC), and deploy it. Since we have already thoroughly tested the facade class via our customer tests, and since our single web page is so simple, we forego writing automated tests against the web page--a manual test using a few fellow knowledge workers will do the trick.

Later, we start adding some major new functionality, like being able to preorder our lunch for the day. We extend our fine-grained classes and the corresponding developer tests, knowing that our pre-existing tests guard us against breaking existing functionality. Likewise, we extend our facade class, perhaps even splitting off a new class (e.g., MenuFacade and OrderFacade) as the interface grows, with similar additions to our customer tests.

Now, perhaps, the changes to the website (two pages is a website, right?) make manual testing unsatisfactory. So, we bring in a simple tool comparable to HttpUnit that allows nUnit to test web pages. We implement a battery of integration/presentation tests, but against a mock version of our facade classes, because the point here is simply that the web pages work--we already know that the facade classes work. The tests push and pull data through the mock facades, only to test that the data successfully made it to the other side. Nothing more.

Of course, our grand success prompts the CEO to request (demand) that we expose the web application to mega-corp's BlackBerrys. So we implement some new pages and a new battery of integration tests. We don't have to touch the developer or customer tests, because we have added no new core functionality.

Finally, the CTO requests (demands) that we extend our cafeteria application to all of mega-corp's robotic workers--you did notice them over the last few days? So, now we add a web services layer that communicates through our facade. Again, no changes to our core functionality, our developer tests, or our customer tests. We apply the Adapter/Wrapper pattern by creating classes that expose the facade with an equivalent web service API, and we create client-side classes to consume that API. We add a new battery of integration tests, but they use plain nUnit to create client-side API classes, which communicate over the web service wiring to the service-side API classes, which invoke mock facade classes, which confirm that our wiring works.

Note that throughout this whole process, we did not need anything significant beyond our production platform and code, our chosen development platform, a few open-source components for automated building and testing, and a few well-defined batteries of tests. Also note that we didn't test anything that we don't use in production, and we didn't test anything twice.

We ended up with a solid core of functionality (business logic tier) that has proven itself mature (hypothetically). We have three separate presentation tier implementations: a website targeted to desktops, a website targeted to BlackBerrys, and a web service API.

Now, please forgive me for the long answer--I tire of inadequate answers and I did not want to provide one. And please note that I have actually done this (though not for a cafeteria menu).