# Are all databases reducible to this ultimate abstract database design?

I've designed a few databases in my time, and on more than one occasion the drive to abstract common elements from specific tables has led me to create generic top-level tables which contain those common elements. For example:

Table      Column         Column

Hamburgers Item           Topping
Cheeseburger   Tomatoes
Mushroomburger Swiss


Could be "simplified" ("normalized") as:

Table      Column         Column

FoodTypes  ID             Name
1              Hamburger
2              Topping

Food       Item           TypeID
Cheeseburger   1
Mushroomburger 1
Tomatoes       2
Swiss          2


Recently I've gone over the deep end with this approach, abstracting and re-abstracting a fairly complex database design until I was left with something both very simple and yet completely un-resembling of the actual data being stored.

This has led me to the conclusion that all databases could be "summarized" in a single monstrous table called "Entries" with columns:

ID       Type     Value1     Value2


For example:

ID       Type     Value1     Value2
4321     Item
8746     Descrip  4321       Food
5673     Item
9876     Descrip  5673       Hamburger
0341     Item
1234     Descrip  0341       Lettuce
5478     Relation 5673       0341
2381     Descrip  5478       Topping
2244     Relation 5673       4321
2160     Descrip  2244       Class
4436     Relation 0341       4321
7547     Descrip  4436       Class


Here, using these 4 columns in 1 table, I have created two objects sharing a common superclass, given them an attribute, and defined not only a relationship between them but the class of that relationship as well. We could now say "Lettuce is a Topping of Hamburger, both of which are Foods".

There would of course be a set of rules for this system, but that is beyond the scope of this question.

My question is, is this not logically the case? If so (or if there is a different, "correct" answer), what is this in relation to real databases? Does such a system exist, or should it not?

I'm not sure if I've gone far enough in my analysis, and I feel like I'm on the verge of some insight which is profoundly obvious to mathematicians and computer scientists (like "yes, all relational data can be described in terms of binary operands like F[a, b]).

• I think that works (assuming your columns have no type, or you store everything as string); what you are doing is model the database as graph -- that always works! -- and write it down as nodes and edges in one list. But why would you? It's horrible in terms of readability, maintainability, and probably also efficiency. – Raphael May 16 '13 at 6:06
• Why not just use a JSON object at this point? – Wildcard Jul 12 '16 at 20:51

This insight was made at least as far back as 1886, by Alfred Kempe. In fact, he observed that any set of relationships (not just those that are already relations) can be captured as a binary relation. At the time this seems to have come as a shock to Charles Sanders Peirce, who had built his relation algebra on the idea that any relation can be captured by means of ternary relations, and who up to this time apparently believed this to be the best possible. Kempe used graphs to illustrate his notion, while Peirce used edge-labelled graphs. In a long and admiring review Peirce revised his theories to accommodate binary relations as a possible foundation, and distinguished the two approaches:

Mr. Kempe's conception depends upon considering the diagram purely in its self-contained relations, the idea of its representing anything being altogether left out of view; while my doctrine depends upon considering how the diagram is to be connected with nature.

In terms of modern database representations, RDF corresponds to Peirce's ternary relations, while key-value stores correspond to Kempe's graphs. As Raphael pointed out in a comment, these are essentially the ideas behind graph databases.

As an aside, the arguments about binary/ternary/quaternary/quinternary/... tuples being the "best" foundation for data representation still rage on in the W3C semantic web working groups. I believe that the original insight of the relational model was that for efficiency, one should use relations of the right arity; relationships that are naturally represented as tuples should not be transmogrified into more complicated data structures just for reasons of purity. However, this argument also goes the other way: relationships where different instances naturally have different arities should not be shoehorned into a rigid relational schema just to preserve the purity of the data model.

• A. B. Kempe, A Memoir on the Theory of Mathematical Form, Philosophical Transactions of the Royal Society of London 177, 1886, 1–70.
• C. S. Peirce, The Critic of Arguments (1892 essay), in Collected Papers of Charles Sanders Peirce, Vol. III, Harvard University Press, 1933, 250–265.
• Outrageously good answer, thank you. This is exactly what I was wondering. I was thinking that a key-value store was similar to my conclusion, and so you say as well. – SG1 May 16 '13 at 17:22
• It should perhaps be added that if we assume some kind of logic for working with the data, allowing more than 2-place relations may still increase expressive power due to limitations within the logic - see e.g. en.wikipedia.org/wiki/Relation_algebra#Expressive_power – reinierpost May 16 '13 at 22:46
• Reading Peirce, I think what he calls a ternary relation is actually quaternary, subject–predicate–indirect object–object, and what he calls "dual relations", Kempe's graphs, are actually ternary, subject-predicate-object, cf. RDF triples. E.g. “In a certain act, D, something is given by A”: subject is act D, predicate is 'giver', object is A. – David Moles Jul 19 '16 at 18:21