It's interesting. After Charlie made an immense effort and rewrote our IO system, basically from scratch, I have started to find bugs. But these are generally not bugs in the IO code, but bugs in Ruby libraries that depend on the way MRI usually works. One of the more annoying ones are IO#read(n), where n is the length you want to read.
This method is not guaranteed to return a string of length n, even if we haven't hit EOF yet. You can NEVER be sure that what you get back is the length you requested. Ever. If you have code that doesn't check the length of the returned string from read, you are almost guaranteed to have a bug just waiting to happen.
Of course, it might work perfectly on your machine and every other machine you test it on. The reason for this is that read(n) will usually return n bytes, but that depends on the socket implementation or file reading implementation of the operating system, it depends on the size of the cache in the network interface, it depends on network latency, and many other things. Please, just make sure to check the return values length before going ahead and using it.
Case in point: net/ldap has this exact problem. Look in net/ber.rb and you will see. There are also two - possibly three - bugs (couple of years old too) that reports different failures because of this.
One thing that makes this problem hard to find is the fact that if you insert debug statement, you will affect the timing in such a way that the code might actually work with debug statement but not without them.
Oh, did I mention that StringIO#read works the same way? It has exactly the same guarantees as IO#read.