In contrast, I found it rather lacking. No discussion of the most common way to sort floats as bytes (shift the sign bit down and XOR the other 31 bits with the resulting masks), nor NaNs and +/-0 for that matter. Varint sorting introduces its own homegrown serialization but doesn't discuss the issue of overlong encodings. Nothing about string collation or Unicode issues in general. Composite data suggests adding NULs, but what if there are NULs in the actual data? (It is briefly mentioned, but only as in “you can't”.)