Errata

The image for single hop is incorrect. The image should have one less hop. Thanks to Aleks Shulman for pointing this out.

The section on Binary data says "If the splittable unit of binary data is larger than 64 MB, you may consider putting the data in its own file, without using a container format."

I have a different opinion here, I don't think that is always the case. Here is my argument, if the file size is large say 10 GB, we can store it is as is i.e. without any container format like SequenceFile.
But suppose we have many (around 10,000) binary files each has 300 MB (larger than 64 MB), we are still benefited by storing it in container like SequenceFile because that way we have one large SequenceFile containing small files (where Key is file name and Value is the content of the file). Now if we have stored this file as is we are increasing the memory footprint of Namenode (1 big file Vs 10,000 small file). Although we don't have appreciable gains in terms of space but we have substantial improvement in terms of memory of the Namenode which I think is a major architectural decision.