To cut the long story short, I’ll send you directly to the very interesting link:
You can find there an amusing story about Facebook engineers making the decision of giving up on Apache Hadoop (Map-Reduce implementation) and going for their own build-from-scratch solution named Corona. You won’t find many technical details here, but you can read a bit about the scale of their business and why they think Map-Reduce isn’t enough for them. This may be the scalability level you won’t even think about on any of your project (but I wish you you will :D), but the educational aspect is important as well.