I'm trying to apply the TDD approach to an I/O library that I'm writing to manage custom archive files that contain my game's resources and assets. This is a learning process so I'm trying to follow the rules of TDD rigorously, but I'm falling on a few issues for which I don't see an easy way out without breaking TDD. I'd like to have some advice on this.

TDD says "test behavior, not implementation", but when the whole point of a class is its implementation, how does TDD go about it? I'm asking this because in my game's I/O library, I have the OutputStream class which is the base class for all stream objects that write output somewhere, and then I have the BufferedOutputStream that implements buffering and flushes the buffer when it's full. (Into another OutputStream that may be a FileOutputStream, a MemoryOutputStream, a mock object for unit tests, etc.)

Basically, I'd like to test that the buffering works, but this is about implementation, not about behavior. It's easy to test that the Flush method works by using a mock object for the buffered stream's output and checking that its Write method is called, but how do I test that the actual buffering works correctly? The simplest method to make every tests pass is to just not bother with buffering and write the input directly in the output stream. Or more realistically, I could have a bugged buffering algorithm that flushes the buffer when it is not full, and I don't know how I could write a unit test that fails without breaking encapsulation and knowing volatile and private implementation details such as the size of the buffer. This is obviously private stuff, and if I add public method to change and access the buffer, I'm breaking encapsulation, bloating the API and potentially adding more ways to break the object's internal state.

Also, I'd like to know if I'm right to assume that classes and methods that interface directly with the operating system cannot be tested until integration tests, and thus should not be written until then? Again, in my I/O library, I have the FileOutputStream which is a specializtion of the OutputStream that writes to files. Sounds like I have to break TDD principles, otherwise I'll never write that class without unit tests.

I imagine a good answer might involve breaking the buffering out into its own class (one responsibility for one class, etc.).

Also, I've occasionally made a unit test a "friend" of the class it tests so it can screw with internals without exposing them in general.

After reading that a few times, it looks like you have a unit test for the buffers, and an acceptance test / integration test for the archiving system as a whole.

For the buffering system, I would think this kind of test would work for the problem you described:

[TestFixture]
public class BufferedOutputStreamShould {
  [Test]
  public void StoreMaxBytesWithoutFlushing() {...}
  [Test]
  public void StoreMaxPlusOneBytesWithFlushing_AndLeaveOneByteBuffered() {...}
  [Test]
  public void BeEmptyAfterAddingOneByteAndFlushing() {...}
}

These things could be testable by exposing the maximum number of bytes that can be buffered and the bytes still in the buffer. Neither requires great insights into the details, and you can modify the actual number of bytes later without changing the interface. Simple functions like GetBufferTotalSize() and GetBufferRemaining() should suffice, no need for friends.

This should address your concerns about it dumping the data before it is completely full, and verify that buffering is dumping the data when it becomes exactly full.


The archiving system as a whole can operate without regard to buffering. Here's the things I would expect in the integration tests:

[TestFixture]
public class ArchivingSystemShould  {
  [Test]
  public void GiveIdenticalFooAfterSaveLoad() {...}
  [Test]
  public void GiveIdenticalBarAfterSaveLoad() {...}
  [Test]
  public void HandleBatchOfFiftyLoadSave() {...}
}

Is that about it?

I hope you dont mind, but I think your design is a bit overcomplicated and makes testing more difficult than it needs to be.

In your design there are many layers and every layer needs to know and call the next layer which needs to know and call an even lower layer and so on. Also it looks like you maybe intermingled data formatting/encoding with writing.

You could just have a few completely separate systems where each just gets a memory buffer containing a whole asset and then gives you a memory buffer with the whole encoded/decoded asset. The archive system would just get asset ids and gives you a simple object which just contains start and end index and which file to write to and possibly another+a little data for updating the archive index. The file layer takes the info object with filename, start, end and data buffer and writes or reads it.

That makes testing easier as you dont need mockups and you spare yourself from implementing buffering logic and knowledge about other systems into all layers.

@frob

Sorry if my original post was unclear, but you nailed my problem rather well. Thanks a lot for your input. Yeah, that sounds like the best way to do it. I designed my BufferedOutputStream class to be usable as a base class because some kinds of streams are begging for buffering (compression and decompression streams come to mind) so the functions you suggested are already there for the child classes, they're just protected. I don't know if it would be considered good practice, but I could just create a new class that inherits BufferedOutputStream to expose those protected methods.

I still don't like the idea of testing an implementation, because I could change it for optimization reasons, but to test a class that does buffering, I guess I need a spec to test.

@wintertime

Oh I don't mind your comment at all. In fact, I wouldn't make an entire set of stream classes normally. Besides, my old archiving library worked similar to the way you suggest. The reason I have all these streams is because I want something more modular this time around. This is a library I want to use in many projects, and I also want it to be more of a I/O library than something limited to archives. The archive manager is mostly workable and actually works pretty well.