TITLE: Why does the algebraic condition of flatness on the structure sheaves  give a good definition of family?
QUESTION [18 upvotes]: Hartshorne remarks that is is something of a mystery as to why the algebraic condition of flatness on the structure sheaves gives a good definition of a family (see below). Are there any known enlightening explanations that help serve to unravel this mystery?  Below is Hartshorne's introductory motivation to flat families containing said remark:

For many reasons it is important to have a good notion of an algebraic 
  family of varieties or schemes. The most naive definition would be just to 
  take the fibres of a morphism. To get a good notion, however, we should 
  require that certain numerical invariants remain constant in a family, such 
  as the dimension of the fibres. It turns out that if we are dealing with non- 
  singular (or even normal) varieties over a field, then the naive definition is 
  already a good one. Evidence for this is the theorem (9.13) that in such a 
  family, the arithmetic genus is constant. 
On the other hand, if we deal with nonnormal varieties, or more general 
  schemes, the naive definition will not do. So we consider a flat family of 
  schemes, which means the fibres of a flat morphism, and this is a very good 
  notion. Why the algebraic condition of flatness on the structure sheaves 
  should give a good definition of a family is something of a mystery. But 
  at least we will justify this choice by showing that flat families have many 
  good properties, and by giving necessary and sufficient conditions for 
  flatness in some special cases. In particular, we will show that a family 
  of closed subschemes of projective space (over an integral scheme) is flat if 
  and only if the Hilbert polynomials of the fibres are the same. 
  -- Hartshorne, Algebraic Geometry, 1977, III.9.5, p. 256

REPLY [5 votes]: There is also the following (probably unhistorical) point of view (it is a version of Hailong Dao's answer). Namely, you don't have to work with flat families at all, so if you want, you can just declare all morphisms to be 'families'. The problem with this approach is that this is a family of `derived' objects. Here's an example:
Let $S$ be a scheme, and let $F$ be a coherent sheaf on S. When is it a 'family' of its fibers? If it is flat, it definitely deserves to be called a family of vector spaces (a vector bundle). But even if it is not flat, you can still view it as a family, but the family of what? The (derived) fibers of $F$ are no longer vector spaces, they are complexes of vector spaces (precisely because $F$ fails to be flat), so we can view $F$ as a nice family of complexes of vector spaces, even though $F$ itself is a sheaf, not a complex.
To summarize: by all means, let's forget about flatness and declare any morphism to be a family. . . of some kind of derived objects. If we now want members of the family to be actual objects (schemes, vector spaces, sheaves, or whatever it is we are trying to include in a family), flatness is forced on you more or less by definition.