What every developer should know about testing - part 2

The short version of this post:

pyVmomi contributions from this point forward will be required to follow this pattern. They include a fixture based unit test build around vcrpy.

Note: While these posts are python specific, the techniques I advocate are not. If you happen to be working in a JVM language, then I recommend Betamax. And, if you happen to be on Ruby, then use vcr. The point of any unit test is that it should be stand alone, automated, and deterministic. These posts discuss how to create deterministic and automated tests in situations where people often argue it's impossible to create such tests.

Overview

Last week I covered some basics about unit testing. What is a unit? What do you need to test? What don't you? This week we'll dive a bit deeper and cover why a Stub, a Mock, or any other trick may not solve your problems and how we create a unit test where there's arguably no 'unit' only a client.

In this post we'll briefly recap what a unit test is. Then we'll cover stubs and mocks and why they are useful but aren't sufficient. Finally I'll cover fixtures.

If I'm in the mood, I may cover property based testing at some future date ... but that's an intermediate to advanced topic in my book.

Recap: What a Unit Test is, what it isn't

The dividing line between units is 'human hard' because it's not just 'a method is a unit' ... a method very frequently is a unit but then so is a class and there can be hidden classes and methods. I'm advocating a somewhat controversial position based on the back-lash to TDD you'll find around the blogosphere recently. My position in a nutshell is that tests are best defined at 'unit boundaries' which are necessarily abstract borders between software units (and units themselves are a term of art).

What is a good Unit Test?

There is some debate over what makes a good unit test. In general, your tests should be stand alone, automated, and deterministic (and those three words in this context have very precise meanings). This is simple enough when a unit is simple. For example, the canonical roman numeral example is easy enough to unit test.

In the roman numeral example the unit is easily isolated since the concept of a roman numeral does not require the introduction of additional units. This is the most basic tier of unit testing that every developer  must learn.

The next set of techniques are to use Fakes and Stubs, and then later maybe Mocks, and these work as long as system interactions remain relatively simple. But, these are not sufficient tools for your tool box. You have to learn one more trick before you have all the basic tools you'll need for modern development. You need to know how to build fixtures and the resources that go with them.

When to Stub

If your programming language supports interfaces and dependency injection then creating stubs will feel natural to you. If your unit uses several other objects by composition to accomplish it's work, one of the simplest things to do is write a fake version of the objects your unit under test uses.

That means you have to create an object that implements all the methods of the dependent class that your particular test unit will use. The stub will have to cover as many calls as your unit of code uses. If you call on method 'foo' you need to write something that reasonably reproduces the expected output of 'foo'. This is fine when the call number is small, but it can quickly get out of hand. Take for example that someone felt compelled to write an entire fake server in Java.

The fake HTTP server author calls their product a mock server but this is in fact a stub. You have no facility to assert call order, parameters, or the absence of calls.

When to Mock

A mock is not a stub. With a stub you have to provide a stub implementation of all possible call paths and have no facility to later go back and make assertions on call order. If you do, you probably wrote it yourself... and you've probably neglected what your day job really is.

A mock is about coding for expected calls and call order. The problem with doing this with your stub is that you will have to either be organically grow your stub into a mock to get this information (and that's a whole new project worth of complexity) or you'll have to invent a whole system of semaphores and messages to watch for these details.

With mocks, you can assert how a method was called; you can make assessments about call order; you can also assert that there is an absence of calls. This is all very important for you to be able to do in your unit tests. With these tools you can continue to assert that no matter how you evolve your intervening code the units code continues to use it's underlying API within acceptable parameters.

A concrete example is if you are consuming pyVmomi or rbVmomi as your client library bindings, you can mock the calls to the client binding library. If you observe that vim.VirtualMachine's PowerOff method is still called properly even after you refactor your own library then there's no need for a VCSIM to run your tests at the unit testing phase.

Too much of this, however, and you can't make assertions about the code's behavior outside of test. Not to mention that creating mocks and stubs are programming efforts of their own. This can lead to wonderful code coverage numbers, tightly coupled designs, and a frustrating body of work.

Enter the Network

What if you're writing a client library? On the pyVmomi project, that's what we're doing. A highly efficient and specifically tailored API binding in Python for vSphere. How can we unit test something that's intended for use with a networked appliance?

Fixtures

In the case of building tests for networked software, the simulator problem is a common problem. In order to deal with this it's vital to have some tool that can record your interactions then play them back. The ability to deliberately tamper with the interaction recording is vital. There are bound to be states that the complex server or simulator on the other end can't reach.

If you only use a simulator for your tests, then they are by definition not stand-alone. You have to exit the testing context and enter an administration context to build an orchestrate the situation you want to test. This can be costly and it can be a problem very similar to the simulator problem. You will inevitably have to build a complex environment that will harm the determinism of your tests.

Next week, I'll get specific and cover how to build a test with a recording, we'll have to get into details and I'll cover how I use VCSIM to create the basis for complex situations I can't even create in my environment.


More on that next time...

What every developer should know about testing - part 1

The short version of this blog post is that this week and next week on pyVmomi will be spent on radically improved testing code and processes. This testing process will become part of the commit cycle. And, it's about damn time.

Once these measures are in place over the next few days the project will be better able to absorb new commits from interested parties.

Overview 

Over the last three weeks I've been working on pyVmomi's next release. If you've not been following along, it's a python client for a SOAP server and it's a code base dating back to at least Python 2.3 but it has only been in the wild as OpenSource now since December 2013. The pyVmomi library has a long internal history at VMware and has served many teams well over the years. But, it does have problems.

The OpenSource version of the library has *no* unit testing shipping with it this makes it hard for interested third parties to contribue. It's time to change that. But, it's a client library for a network service. How do you test such a beast?

In this post I will cover what unit testing is and what integration testing is and how this impacts the design choices made on libraries. This discussion is directly applicable to the long-term evolution of a library like pyVmomi and is generally applicable to Software Design. I'm bothering to write all this down because I want everyone to be on the same page and I don't particularly want to repeat myself too much. Over the years I expect to be involved with this project I expect to point at this post frequently.

The Problem

Work on pyVmomi has been rather painful. For much of it, I have had to spend vast amounts of time deep in the debugger. Testing on this library involves building a VCSIM and simulating a vCenter environment. This in turn means the creation of a suitable inventory and potentially setting up a suitable Fault to work with. This is a lot of yak shaving to get to the point you can even start considering doing development work.

The root of the specific problem

The problem in specific is that pyVmomi as a library speaks to a server. No other thing can completely simulate all the inputs, outputs, and exposed states that such a thing achieves except the big complex thing itself. This problem is routinely solved by developers in this entire cloud infrastructure space by spinning up virtual environments to create the scenarios they need.

This is a natural inclination since you have a beautiful hammer, why not nail all the bugs with this beautiful hammer? Virtualization is powerful and has transformed our industry. One day I will be an old man and tell stories of how infrastructure and development worked in the bad old days of Dot-Com Boom but this inclination is an example of the hype-cycle in full detrimental effect.

The problem in general

Because client library code development starts at the integration phase the units that end up defined by the client library programmer are inherently integrations. How do you test integrations? With integration tests. But, how do you do integration testing when the thing you are testing isn't even on your build machine? If it's a server (such as our case) you have to fall back to either a simulator or you have to stand up a whole new of your environment just for testing.

Unsurprisingly, this is fairly standard practice for every step of IaaS and PaaS development. You stand up the universe to author a new function then you retest the whole thing on a fresh copy of the universe. Then, you wash-rinse-repeat for the whole integrated system. It's so easy. It's also so very wrong. Because code that is hard to test (or completely untestable) in isolation is poorly designed. If you're defending the fact that it's tested, you're missing the point.

This isn't just a problem with the one library I'm working on now. I've seen this repeatedly in development environments of every kind at huge shops and tiny shops. You build up a pile-o-software that glues systems together and to test it you build a pile-o-infrastructure that you bring to a pile-o-state so you can validate the right calls and responses.

When you test this way (bringing a whole simulated universe into existence to test your new 'if' statement), invariably something's state gets out of sync and what do you do? You have to test the test environment to validate that you don't have false positives for your failure report, then you have to retest and you re-start the whole process which typically grows into hours. This is, frankly, an extremely expensive way to develop software.

And, for the record I've seen this in JEE, Spring, Grails, Python, Bash, Perl, C, C++, projects on Solaris, Linux, Irix, BSD, and now ESX based environments. This is not a problem unique to those crappy developers on that crappy environment. This is an intrinsic integration development problem that crops up when you routinely write code that takes big complex systems and makes them work together. It's a far too easy trap to fall into and a far too difficult of a pit to climb out of.

Unit Testing?

So the story so-far is that we have a library, maybe that library talks to things "not of this machine". Maybe, it speaks over the wire, talks to other things we can't see or directly control. These are things well outside of anything we could define as our unit of code. So if that's our fundamental unit (because what *else* is something like pyVmomi?) How the heck do we test it?

The term unit is deliberately ambiguous in this context. Did we mean class? Did we mean method? The answer is it depends. Getting the logical border of the unit is hard. It's actually human intelligence hard. It's "why AI do not yet write code" hard. Why is it hard? It's hard in the same what the making beautiful art is hard it's a fuzzy problem that requires aesthetics.

The definition of where a unit is, is hard and simultaneously critical to get right. Define the wrong unit and pay the price. This doesn't mean testing is wrong, it just means testing is a programming-hard problem. Looking for easy answers here just means you don't know what you're doing.

Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.
        — Brian W. Kernighan and P. J. Plauger in The Elements of Programming Style.

I mock your Mocks!

A simple answer to the problem is to Mock and Stub everything. (Mock objects are not Stubs, and you should know the difference. Edit: this is a whole topic unto itself and I will cover it separately at a later date.) The problem is when you work with a sufficiently complex interactions you will be forced to write sufficiently complex mocks and stubs. You will back your way into the simulator problem I mentioned before. In our specific case this means essentially re-inventing something like VCSIM except all in Python mock objects and that's absurd.

What are you forging, o' library author?

Consider also, where is the unit boundary when it comes to a client library? The library absolutely has a boundary at its public interfaces. If the bit-o-code you just wrote is private why did you write it? The private method is not part of the interface and so therefore it's not part of the library's unit definition. The unit in this context only makes sense as a test-first tested component if it's going to be exposed. By definition a private method isn't exposed so it's an implementation detail and we don't test our programs to make sure implementation details work. We don't test if the 'if' statement works. So where is the detail and where is the interface?

This means your test-first tests should be your surface. To develop a library that you intend on providing to people to interact with you should model sets of expected interactions. Each interaction should then be codified as a set of calls to the library.

Code as little as possible, test as little as possible

Tests are code. The mark of a good programmer is not how much code they write, but how much they can accomplish with how little. If you are following the aesthetic of minimalist code, then this attitude should also follow into your tests. Your tests should cover as many lines as is needed to validate the code and should do this as effectively as possible. Ideally, you should have an efficiency to your tests. No two tests should cover exactly the same unit. Covering a unit multiple times is effectively wasted effort.

This is a much harder philosophy and practice to follow than the lazy 'cover all the lines' strategy. It requires you to understand the functional cases your code can cover. In a rarified ideal world, this might mean you get 100% coverage anyway but the percentage isn't what we care about. You can have 100% code coverage and still have horrible comprehension of what your project even is.

Breaking down your tests to cover all the methods, even the private ones is a horrible idea. If you cover your private methods you will be tying your tests that matter to what (by defining them as private methods) you have decided are implementation details. That equals tight coupling; tight coupling is bad.

How do you test something where it provides very little function other than basic interactions with a service? How do you exercise a library that is arguably mostly private and hidden code?

Introducing Fixtures

The concept of a testing fixture is a very old testing concept. It even predates software as a thing and yet I rarely if ever see a shop using fixtures. The truly sad thing is that for most development languages fixtures are old as the hills. So, WHY are so few projects using them?

A Specific Solution: vcrpy

I reviewed several Python fixture libraries this week and was fairly well impressed with vcrpy for our purposes. The description may mention 'mocking' but function of the library is to provide you with testing fixtures at the socket level. In libraries like pyVmomi we are effectively a skin over a very complex back end web service. This 'skin' nature of ours means that a simple set of library interactions may hide dozens of network conversations.

Manually creating dozens of HTTP interaction mocks to explore a single high-level test can be so painful that you are likely to just not do it. Fortunately tools like vcrpy exist and can record your HTTP traffic. Now you can do the lazy thing and toy with your client-server a bit, record the on the wire interactions, and then later (and more importantly) edit the conversations to represent the larger API cases you want to cover.

With the recorded HTTP fixtures at our disposal we can now work with the binding in much more predictable and controlled ways.


More on that next week... (or skip to the end)

pyVmomi and the road to Python3 - part 3

The Good News

It works! And without major surgery!

I currently have a working branch of the pyVmomi code base that runs on python 3.x interpreters. We are working to get this version usable under Python 2.6, 2.7, 3.3, and 3.4 as well and hope to post on that next week. I'd also like to thank Rackspace for offering to host our CI servers. Kudos to Michael Rice for getting that to happen.

The bad bits...

One of the most frustrating aspects of porting pyVmomi on to Python 3.x has been the fact that many of the language changes caused pyVmomi to fail to load or fail silently. This is in part due to design considerations made in the library early on, and in part due to how the language has changed.

In particular the way imports are done were instrumental in basically masking the failures. The try-except-swallow pattern is frequently used and it tended to hide why a vim.* or vmodl.* data type wasn't loading into memory.

---

---

Simple porting techniques and tools also helped mask the problems, however, now that they are known I may back off some of the changes in favor of cleaner more generic solutions.

Simple 2 to 3

A number of changes are rather trivial, take the case of exceptions for example...

... there were a number of these trivial cases where tools like the 2 to 3 tool could help. But, in general libraries like six and tools like 2to3 tended to mask underlying problems with more complex issues such as unicode and type inheritance in the library.

There were a number of these kinds of changes

When is a Type not a Type?

Inheriting off of base classes is considered very bad practice. It tends to gain you very little and tends to complicate code unnecessarily. Consider the Link class in pyVmomi's VmomiSupport.py which very naughtily extends the core unicode class in Python.

---

---

The Link class here serves the sole purpose of allowing a string to be recognized as a type temporarily so that when we do type mapping during SOAP deserialization we have a type to latch on to in key branching logic.

This is clearly not within the use cases most people are thinking about when they are thinking about Python 2to3 and while the practice is naughty, it does provide an interesting orthogonality to the mapping code.

So, because of that, while it's bad practice to extend a built in type I can't think of something cleaner that wouldn't immediately ugly up the code-base. However, I'm also not sure how that bit of logic currently performs since any use of Link is going to return ostensibly a string (albeit a unicode one) and said string is never going to to be of type Link. So the first point still stands... extending a base class is more confusing than it is useful.

That's just out of scope for strictly Python portability work, so for the time being the class stays in. It also seems mostly benign at the moment. There's a significant lack of testing around the lib and that needs to be addressed before I will be confident doing any major refactoring work. IMHO: The python 3 support work is pushing our luck as it is.

Unicode uncoded

Of particular interest is in testing our Unicode related code no longer make sense in Python 3.

Hopefully, we can do away with the uglier bits I've hacked on here to make things work. We will also need to back-track the assumptions made when the 'unicode' and 'basestring' symbols were used and make sure they still hold.

Special thanks to the friendly folks on #python for their help with understanding the issues around Python and unicode.

Next steps: testing, testing, testing

My next set of tasks will be to clean up and refine the python 3 related changes to try and make them as targeted an small as possible. I can't realistically do this with much confidence if I can't make assertions about the library's operation. While VMware has internal build systems and tests that validate and verify the library against a variety of product builds I can't realistically expose these since they are intimately related to shipping product itself.

The pyVit (pronounced pi-vət) project is a first step toward creating a test suite for pyVmomi that a 3rd party could easily consume to validate the library and their environments. This will help us move the library from its current status of beta to a 'stable' category. I also hope this will enable broader stewardship of the pyVmomi library allowing it to take on a life of its own. Ideally, this will also enable anyone to fearlessly refactor the library.

More on that next week...

pyVmomi and the road to Python3 - part 2

Short week this week, so a short update ...

I've spent most of my time with my head in the debugger flipping between Python 2.7 and Python 3.4 interpreters. I've been observing the differences in behavior between the two versions and finding ways to make pyVmomi behave with both. Some of the more interesting work has had to do with how Python 3 handles class loading and namespaces. I could probably write a whole post on that alone.

In my current attempt, even with a good debugger I've been forced to resort to bear-skins-and-bone-knives development. The problem being that the critical fault occurs during interpreter load. That has to do with how pyVmomi builds its binding.

At runtime pyVmomi uses a LazyModule class to hold instances of class definitions it needs. This leads to some efficiencies in cases where very little of the actual API is being used. The API is enormous as you can see from the generated documents I posted so this efficiency makes sense. It does mean that the pyVmomi classes that you use in your own code are not python types as you normally experience them. The binding does produce a convincing illusion, however.

I've bumped back the next release milestone from July 15th to July 31st to give us time to validate the changes for Python 3 support since they will be slightly more invasive than I had hoped for.  I want the next post to pypi to include that Python 3 support patch if possible.

Lessons learned from porting pyVmomi to Python 3 will be generally valuable beyond just pyVmomi but, I'm still in the thick of it. So far progress is encouraging but there is still a way to go before declaring the new version stable.

More on that next week...