You are on page 1of 11

Information Retrieval P.

BAXENDALE, Editor

A Relational Model of Data for The relational view (or model) of data described in
Section 1 appears to be superior in several respects to the
Large Shared Data Banks graph or network model [3,4] presently in vogue for non-
inferential systems. It provides a means of describing data
with its natural structure only-that is, without superim-
E. F. CODD posing any additional structure for machine representation
IBM Research Laboratory, San Jose, California purposes. Accordingly, it provides a basis for a high level
data language which will yield maximal independence be-
tween programs on the one hand and machine representa-
Future users of large data banks must be protected from tion and organization of data on the other.
having to know how the data is organized in the machine (the A further advantage of the relational view is that it
internal representation). A prompting service which supplies forms a sound basis for treating derivability, redundancy,
such information is not a satisfactory solution. Activities of users and consistency of relations-these are discussedin Section
at terminals and most application programs should remain 2. The network model, on the other hand, has spawned a
unaffected when the internal representation of data is changed number of confusions, not the least of which is mistaking
and even when some aspects of the external representation the derivation of connections for the derivation of rela-
are changed. Changes in data representation will often be tions (seeremarks in Section 2 on the “connection trap”).
needed as a result of changes in query, update, and report Finally, the relational view permits a clearer evaluation
traffic and natural growth in the types of stored information. of the scope and logical limitations of present formatted
Existing noninferential, formatted data systems provide users data systems, and also the relative merits (from a logical
with tree-structured files or slightly more general network standpoint) of competing representations of data within a
models of the data. In Section 1, inadequacies of these models single system. Examples of this clearer perspective are
are discussed. A model based on n-ary relations, a normal cited in various parts of this paper. Implementations of
form for data base relations, and the concept of a universal systems to support the relational model are not discussed.
data sublanguage are introduced. In Section 2, certain opera-
1.2. DATA DEPENDENCIESIN PRESENTSYSTEMS
tions on relations (other than logical inference) are discussed
The provision of data description tables in recently de-
and applied to the problems of redundancy and consistency
veloped information systems represents a major advance
in the user’s model.
toward the goal of data independence [5,6,7]. Such tables
KEY WORDS AND PHRASES: data bank, data base, data structure, data facilitate changing certain characteristics of the data repre-
organization, hierarchies of data, networks of data, relations, derivability, sentation stored in a data bank. However, the variety of
redundancy, consistency, composition, join, retrieval language, predicate
calculus, security, data integrity
data representation characteristics which can be changed
CR CATEGORIES: 3.70, 3.73, 3.75, 4.20, 4.22, 4.29 without logically impairing some application programs is
still quite limited. Further, the model of data with which
users interact is still cluttered with representational prop-
erties, particularly in regard to the representation of col-
lections of data (as opposed to individual items). Three of
the principal kinds of data dependencies which still need
1. Relational Model and Normal Form to be removed are: ordering dependence, indexing depend-
ence, and accesspath dependence. In some systems these
1.I. INTR~xJ~TI~N dependencies are not clearly separable from one another.
This paper is concerned with the application of ele- 1.2.1. Ordering Dependence. Elements of data in a
mentary relation theory to systems which provide shared data bank may be stored in a variety of ways, someinvolv-
accessto large banks of formatted data. Except for a paper ing no concern for ordering, some permitting each element
by Childs [l], the principal application of relations to data to participate in one ordering only, others permitting each
systems has been to deductive question-answering systems. element to participate in several orderings. Let us consider
Levein and Maron [2] provide numerous referencesto work those existing systems which either require or permit data
in this area. elements to be stored in at least one total ordering which is
In contrast, the problems treated here are those of data closely associated with the hardware-determined ordering
independence-the independence of application programs of addresses.For example, the records of a file concerning
and terminal activities from growth in data types and parts might be stored in ascending order by part serial
changesin data representation-and certain kinds of data number. Such systems normally permit application pro-
inconsistency which are expected to become troublesome grams to assumethat the order of presentation of records
even in nondeductive systems. from such a file is identical to (or is a subordering of) the

Volume 13 / Number 6 / June, 1970 Communications of the ACM 377


stored ordering. Those application programs which take Structure 1. Projects Subordinate to Parts
advantage of the stored ordering of a file are likely to fail File Segment Fields
to operate correctly if for some reason it becomes necessary F PART part #
to replace that ordering by a different one. Similar remarks part name
part description
hold for a stored ordering implemented by means of quantity-on-hand
pointers. quantity-on-order
It is unnecessary to single out any system as an example, PROJECT project #
because all the well-known information systems that are project name
marketed today fail to make a clear distinction between project description
quantity committed
order of presentation on the one hand and stored ordering
on the other. Significant implementation problems must be
solved to provide this kind of independence. Structure 2. Parts Subordinate to Projects
1.2.2. Indexing Dependence. In the context of for- File Sqmeut Fields
matted data, an index is usually thought of as a purely F PROJECT project #
performance-oriented component of the data representa- project name
tion. It tends to improve response to queries and updates project description
PART part #
and, at the same time, slow down response to insertions part name
and deletions. From an informational standpoint, an index part description
is a redundant component of the data representation. If a quantity-on-hand
system uses indices at all and if it is to perform well in an quantity-on-order
environment with changing patterns of activity on the data quantity committed
bank, an ability to create and destroy indices from time to
time will probably be necessary. The question then arises: Structure 3. Parts and Projects as Peers
Can application programs and terminal activities remain Commitment Relationship Subordinate to Projects
invariant as indices come and go? File Segment Fields
Present formatted data systems take widely different F PART part #
approaches to indexing. TDMS [7] unconditionally pro- part name
part description
vides indexing on all attributes. The presently released quantity-on-hand
version of IMS [5] provides the user with a choice for each quantity-on-order
file: a choice between no indexing at all (the hierarchic se- G PROJECT project #
quential organization) or indexing on the primary key project name
only (the hierarchic indexed sequent,ial organization). In project description
PART part #
neither case is the user’s application logic dependent on the quantity committed
existence of the unconditionally provided indices. IDS
[8], however, permits the fle designers to select attributes Structure 4. Parts and Projects as Peers
to be indexed and to incorporate indices into the file struc- Commitment Relationship Subordinate to Parts
ture by means of additional chains. Application programs File Segnren1 Fields
taking advantage of the performance benefit of these in- F PART part #
dexing chains must refer to those chains by name. Such pro- part description
grams do not operate correctly if these chains are later quantity-on-hand
removed. quantity-on-order
PROJECT project #
1.2.3. Access Path Dependence. Many of the existing quantity committed
formatted data systems provide users with tree-structured G PROJECT project #
files or slightly more general network models of the data. project name
Application programs developed to work with these sys- project description
tems tend to be logically impaired if the trees or networks
are changed in structure. A simple example follows. Structure 5. Parts, Projects, and
Suppose the data bank contains information about parts Commitment Relationship as Peers
FCZC .%&-,,ZC,,t Ficlds
and projects. For each part, the part number, part name,
part description, quantity-on-hand, and quantity-on-order F PART part #
part name
are recorded. For each project, the project number, project part description
name, project description are recorded. Whenever a project quantity-on-hand
makes use of a certain part, the quantity of that part com- quantity-on-order
mitted to the given project is also recorded. Suppose that G PROJECT project #
the system requires the user or file designer to declare or project name
project description
define the data in terms of tree structures. Then, any one H COMMIT part #
of the hierarchical structures may be adopted for the infor- project #
mation mentioned above (see Structures l-5). quantity committed

378 Communications of the ACM Volume 13 / Number 6 / June, 1970


Now, consider the problem of printing out the part ray which represents an n-ary relation R has the following
number, part name, and quantity committed for every part properties :
used in the project whose project name is “alpha.” The (1) Each row represents an n-tuple of R.
following observations may be made regardless of which (2) The ordering of rows is immaterial.
available tree-oriented information system is selected to (3) All rows are distinct.
tackle this problem. If a program P is developed for this (4) The ordering of columns is significant-it corre-
problem assuming one of the five structures above-that sponds to the ordering S1, Sz , . . . , S, of the do-
is, P makes no test to determine which structure is in ef- mains on which R is defined (see, however, remarks
fect-then P will fail on at least three of the remaining below on domain-ordered and domain-unordered
structures. More specifically, if P succeeds with structure 5, relations ) .
it will fail with all the others; if P succeeds with structure 3 (5) The significance of each column is partially con-
or 4, it will fail with at least 1,2, and 5; if P succeeds with veyed by labeling it with the name of the corre-
1 or 2, it will fail with at least 3, 4, and 5. The reason is sponding domain.
simple in each case. In the absence of a test to determine The example in Figure 1 illustrates a relation of degree
which structure is in effect, P fails because an attempt is 4, called supply, which reflects the shipments-in-progress
made to exceute a reference to a nonexistent file (available of parts from specified suppliers to specified projects in
systems treat this as an error) or no attempt is made to specified quantities.
execute a reference to a file containing needed information.
supply (supplier part project quantity)
The reader who is not convinced should develop sample
programs for this simple problem. 1 2 5 17
1 3 5 23
Since, in general, it is not practical to develop applica-
2 3 7 9
tion programs which test for all tree structurings permitted 2 7 5 4
by the system, these programs fail when a change in 4 1 1 12
&ructure becomes necessary. FIG. 1. A relation of degree 4
Systems which provide users with a network model of
the data run into similar difficulties. In both the tree and One might ask: If the columns are labeled by the name
network cases, the user (or his program) is required to of corresponding domains, why should the ordering of col-
exploit a collection of user access paths to the data. It does umns matter? As the example in Figure 2 shows, two col-
not matter whether these paths are in close correspondence umns may have identical headings (indicating identical
with pointer-defined paths in the stored representation-in domains) but possessdistinct meanings with respect to the
IDS the correspondence is extremely simple, in TDMS it is relation. The relation depicted is called component. It is a
just the opposite. The consequence, regardless of the stored ternary relation, whose first two domains are called part
representation, is that terminal activities and programs be- and third domain is called quantity. The meaning of com-
come dependent on the continued existence of the user ponent (2, y, z) is that part x is an immediate component
access paths. (or subassembly) of part y, and z units of part 5 are needed
One solution to this is to adopt the policy that once a to assembleone unit of part y. It is a relation which plays
user access path is defined it will not be made obsolete un- a critical role in the parts explosion problem.
til all application programs using that path have become
obsolete. Such a policy is not practical, because the number component (part part quantity)

of access paths in the total model for the community of 1 5 9


users of a data bank would eventually become excessively 2 5 7
3 5 2
large.
2 6 12
1.3. A RELATIONAL VIEW OF DATA 3 6 3
The term relation is used here in its accepted mathe- 4 7 1
6 7 1
matical sense.Given sets X1, S, , . . . , S, (not necessarily
distinct), R is a relation on these n sets if it is a set of n- FIG. 2. A relation with-two identical domains
tuples each of which has its first element from S1, its
second element from Sz , and so on.’ We shall refer to Si as It is a remarkable fact that several existing information
the jth domain of R. As defined above, R is said to have systems (chiefly those based on tree-structured files) fail
degreen. Relations of degree 1 are often called unary, de- to provide data representations for relations which have
gree 2 binary, degree 3 ternary, and degree n n-ary. two or more identical domains. The present version of
For expository reasons, we shall frequently make use of IMS/360 [5] is an example of such a system.
an array representation of relations, but it must be re- The totality of data in a data bank may be viewed as a
membered that this particular representation is not an es- collection of time-varying relations. These relations are of
sential part of the relational view being expounded. An ar- assorted degrees. As time progresses,each n-ary relation
may be subject to insertion of additional n-tuples, deletion
1 More concisely, R is a subset of the Cartesian product 81 X of existing ones, and alteration of components of any of its
sz x *.* x 87%. existing n-tuples.

Volume 13 / Number 6 / June, 1970 Communications of the ACM 379


In many commercial, governmental, and scientific data names, and part numbers are. We shall call the set of
banks, however, some of the relations are of quite high de- values represented at someinstant the active domain at that
gree (a degree of 30 is not at all uncommon). Users should instant.
not normally be burdened with remembering the domain Normally, one domain (or combination of domains) of a
ordering of any relation (for example, the ordering supplier, given relation has values which uniquely identify each ele-
then part, then project, then quantity in the relation supply). ment (n-tuple) of that relation. Such a domain (or com-
Accordingly, we propose that users deal, not with relations bination) is called a primary key. In the example above,
which are domain-ordered, but with relationships which are part number would be a primary key, while part color
their domain-unordered counterparts.2 To accomplish this, would not be. A primary key is nonredundant if it is either
domains must be uniquely identifiable at least within any a simple domain (not a combination) or a combination
given relation, without using position. Thus, where there such that none of the participating simple domains is
are two or more identical domains, we require in each case superfluous in uniquely identifying each element. A rela-
that the domain name be qualified by a distinctive role tion may possessmore than one nonredundant primary
name, which serves to identify the role played by that key. This would be the casein the example if different parts
domain in the given relation. For example, in the relation were always given distinct names. Whenever a relation
component of Figure 2, the first domain part might be has two or more nonredundant primary keys, one of them
qualified by the role name sub, and the second by super, so is arbitrarily selected and called the primary key of that re-
that users could deal with the relationship component and lation.
its domains-sub.part super.part, quantity-without regard A common requirement is for elements of a relation to
to any ordering between these domains. cross-reference other elements of the same relation or ele-
To sum up, it is proposed that most usersshould interact ments of a different relation. Keys provide a user-oriented
with a relational model of the data consisting of a collection means (but not the only means) of expressing such cross-
of time-varying relationships (rather than relations). Each references. We shall call a domain (or domain combma-
user need not know more about any relationship than its tion) of relation R a foreign key if it is not the primary key
name together with the names of its domains (role quali- of R but its elements are values of the primary key of some
fied whenever necessary): Even this information might be relation S (the possibility that S and R are identical is not
offered in menu style by the system (subject to security excluded). In the relation supply of Figure 1, the combina-
and privacy constraints) upon request by the user. tion of supplier, part, project is the primary key, while each
There are usually many alternative ways in which a re- of these three domains taken separately is a foreign key.
lational model may be established for a data bank. In In previous work there has been a strong tendency to
order to discuss a preferred way (or normal form), we treat the data in a data bank as consisting of two parts, one
must first introduce a few additional concepts (active part consisting of entity descriptions (for example, descrip-
domain, primary key, foreign key, nonsimple domain) tions of suppliers) and the other part consisting of rela-
and establish some links with terminology currently in use tions between the various entities or types of entities (for
in information systems programming. In the remainder of example, the supply relation). This distinction is difficult
this paper, we shall not bother to distinguish between re- to maintain when one may have foreign keys in any rela-
lations and relationships except where it appears advan- tion whatsoever. In the user’s relational model there ap-
tageous to be explicit. pears to be no advantage to making such a distinction
Consider an example of a data bank which includes rela- (there may be some advantage, however, when one applies
tions concerning parts, projects, and suppliers. One rela- relational concepts to machine representations of the user’s
tion called part is defined on the following domains: set of relationships).
(1) part number So far, we have discussedexamples of relations which are
(2) part name defined on simple domains-domains whose elements are
(3) part color atomic (nondecomposable) values. Nonatomic values can
(4) part weight be discussedwithin the relational framework. Thus, some
(5) quantity on hand domains may have relations as elements. These relations
(6) quantity on order may, in turn, be defined on nonsimple domains, and so on.
and possibly other domains as well. Each of these domains For example, one of the domains on which the relation em-
is, in effect, a pool of values, some or all of which may be ployee is defined might be salary history. An element of the
represented in the data bank at any instant. While it is salary history domain is a binary relation defined on the do-
conceivable that, at some instant, all part colors are pres- main date and the domain salary. The salary history domain
ent, it is unlikely that all possible part weights, part is the set of all such binary relations. At any instant of time
2 In mathematical terms, a relationship is an equivalence class of there are as many instances of the salary history relation
those relations that are equivalent under permutation of domains in the data bank as there are employees. In contrast, there
(see Section 2.1.1). is only one instance of the employeerelation.
* Naturally, as with any data put into and retrieved from a com-
puter system, the user will normally make far more effective use The terms attribute and repeating group in present data
of the data if he is aware of its meaning. base terminology are roughly analogous to simple domain

380 Communications of the ACM Volume 13 / Number 6 / June, 1970


and nonsimple domain, respectively. Much of the confusion If normalization as described above is to be applicable,
in present terminology is due to failure to distinguish be- the unnormalized collection of relations must satisfy the
tween type and instance (as in “record”) and between following conditions :
components of a user model of the data on the one hand (1) The graph of interrelationships of the nonsimple
and their machine representation counterparts on the domains is a collection of trees.
other hand (again, we cite “record” as an example). (2) No primary key has a component domain which is
1.4. NORMAL FORM nonsimple.
A relation whose domains are all simple can be repre- The writer knows of no application which would require
sented in storage by a two-dimensional column-homo- any relaxation of these conditions. Further operations of a
geneous array of the kind discussed above. Some more normalizing kind are possible. These are not discussedin
complicated data structure is necessary for a relation with this paper.
one or more nonsimple domains. For this reason (and others The simplicity of the array representation which becomes
to be cited below) the possibility of eliminating nonsimple feasible when all relations are cast in normal form is not
domains appears worth investigating! There is, in fact, a only an advantage for storage purposes but also for com-
very simple elimination procedure, which we shall call munication of bulk data between systems which usewidely
normalization. different representations of the data. The communication
Consider, for example, the collection of relations ex- form would be a suitably compressedversion of the array
hibited in Figure 3 (a). Job history and children are non- representation and would have the following advantages:
simple domains of the relation employee.Salary history is a (1) It would be devoid of pointers (address-valued or
nonsimple domain of the relation job history. The tree in displacement-valued ) .
Figure 3 (a) shows just these interrelationships of the non- (2) It would avoid all dependence on hash addressing
simple domains. schemes.
(3) It would contain no indices or ordering lists.
employee If the user’s relational model is set up in normal form,
I names of items of data in the data bank can take a simpler
form than would otherwise be the case. A general name
jobhistory children would take a form such as
I
salaryhistory R (g).r.d

employee (man#, name, birthdate, jobhistory, children) where R is a relational name; g is a generation identifier
jobhistory (jobdate, title, salaryhistory) (optional); r is a role name (optional); d is a domain name.
salaryhistory (salarydate, salary) Since g is needed only when several generations of a given
children (childname, birthyear)
relation exist, or are anticipated to exist, and r is needed
FIG. 3(a). Unnormalized set only when the relation R has two or more domains named
d, the simple form R.d will often be adequate.
employee’ (man#, name, birthdate)
jobhistory’ (man#, jobdate, title) 1.5. SOME LINGUISTIC ASPECTS
salaryhistory’ (man#, jobdate, salarydate, salary) The adoption of a relational model of data, as described
children’ (man#, childname, birthyear) above, permits the development of a universal data sub-
FIG. 3(b). Normalized set language based on an applied predicate calculus. A first-
order predicate calculus s&ices if the collection of relations
Normalization proceeds as follows. Starting with the re- is in normal form. Such a language would provide a yard-
lation at the top of the tree, take its primary key and ex- stick of linguistic power for all other proposed data Ian-
pand each of the immediately subordinate relations by guages, and would itself be a strong candidate for embed-
inserting this primary key domain or domain combination. ding (with appropriate syntactic modification) in a variety
The primary key of each expanded relation consists of the of host Ianguages (programming, command- or problem-
primary key before expansion augmented by the primary oriented). While it is not the purpose of this paper to
key copied down from the parent relation. Now, strike out describe such a language in detail, its salient features
from the parent relation all nonsimple domains, remove the would be as follows.
top node of the tree, and repeat the same sequence of Let us denote the data sublanguage by R and the host
operations on each remaining subtree. language by H. R permits the declaration of relations and
The result of normalizing the collection of relations in their domains. Each declaration of a relation identifies the
Figure 3 (a) is the collection in Figure 3 (b). The primary primary key for that relation. Declared relations are added
key of each relation is italicized to show how such keys to the system catalog for use by any members of the user
are expanded by the normalization. community who have appropriate authorization. H per-
4 M. E. Sanko of IBM, San Jose, independently recognized the mits supporting declarations which indicate, perhaps less
desirability of eliminating nonsimple domains. permanently, how these relations are represented in stor-

Volume 13 / Number 6 / June, 1970 Communications of the ACM 381


age. R permits the specification for retrieval of any subset 4-ary relation supply of Figure 1, which entails 5 names in
of data from the data bank. Action on such a retrieval re- n-ary notation, would be represented in the form
quest is subject to security constraints.
P (supplier, & (part, R (project, quantity)))
The universality of the data sublanguage lies in its
descriptive ability (not its computing ability). In a large in nested binary notation and, thus, employ 7 names.
data bank each subset of the data has a very large number -4 further disadvantage of this kind of expression is its
of possible (and sensible) descriptions, even when we as- asymmetry. Although this asymmetry does not prohibit
sume (as we do) that there is only a finite set of function symmetric exploitation, it certainly makes some basesof
subroutines to which the system has access for use in interrogation very awkward for the user to express (con-
qualifying data for retrieval. Thus, the class of qualification sider, for example, a query for those parts and quantities
expressions which can be used in a set specification must related to certain given projects via & and R).
have the descriptive power of the class of well-formed 1.6. EXPRESSIBLE, NAMED, AND STORED RELATIONS
formulas of an applied predicate calculus. It is well known Associated with a data bank are two collections of rela-
that to preserve this descriptive power it is unnecessary to tions: the named set and the expressibleset. The named set
express (in whatever syntax is chosen) every formula of is the collection of all those relations that the community of
the selected predicate calculus. For example, just those in userscan identify by means of a simple name (or identifier).
prenex normal form are adequate [9]. A relation R acquires membership in the named set when a
Arithmetic functions may be needed in the qualification suitably authorized user declares R; it loses membership
or other parts of retrieval statements. Such functions can when a suitably authorized user cancels the declaration of
be defined in H and invoked in R. R.
A set so specified may be fetched for query purposes The expressible set is the total collection of relations that
only, or it may be held for possible changes. Insertions t,ake can be designated by expressionsin the data language. Such
the form of adding new elements to declared relations with- expressionsare constructed from simple names of relations
out regard to any ordering that may be present in their in the named set; names of generations, roles and domains;
machine representation. Deletions which are effective for logical connectives; the quantifiers of the predicate calcu-
the community (as opposed to the individual user or sub- 1~s;~and certain constant relation symbols such as = , > .
communities) take the form of removing elements from de- The named set is a subset of the expressible set-usually a
clared relations. Some deletions and updates may be trig- very small subset.
gered by others, if deletion and update dependencies be- Since some relations in the named set may be time-inde-
tween specified relations are declared in R. pendent combinations of others in that set, it is useful to
One important effect that the view adopted toward data consider associating with the named set a collection of
has on the language used to retrieve it is in the naming of statements that define these time-independent constraints.
data elements and sets. Some aspects of this have been dis- We shall postpone further discussion of this until we have
cussed in the previous section. With the usual network introduced several operations on relations (seeSection 2).
view, users will often be burdened with coining and using One of the major problems confronting the designer of a
more relat,ion names than are absolutely necessary, since data system which is to support a relational model for its
names are associated with paths (or path types) rather users is that of determining the class of stored representa-
than with relations. tions to be supported. Ideally, the variety of permitted
Once a user is aware that a certain relation is stored, he data representations should be just adequate to cover the
will expect to be able to exploit5 it using any combination spectrum of performance requirements of the total col-
of its arguments as “knowns” and the remaining argu- lection of installations. Too great a variety leads to un-
ments as “unknowns,” because the information (like necessary overhead in storage and continual reinterpreta-
Everest) is there. This is a system feature (missing from tion of descriptions for the structures currently in effect.
many current informat.ion systems) which we shall call For any selected class of stored representations the data
(logically) symmetric expZoitation of relations. Naturally, system must provide a means of translating user requests
symmetry in performance is not to be expected. expressedin the data language of the relational model into
To support symmetric exploitation of a single binary re- corresponding-and elhcient-actions on the current
lation, two directed paths are needed. For a relation of de- stored representation. For a high level data language this
gree n, the number of paths to be named and controlled is presents a challenging design problem. Nevertheless, it is a
n factorial. problem which must be solved-as more users obtain con-
Again, if a relational view is adopted in which every n- current accessto a large data bank, responsibility for pro-
ary relation (n > 2) has to be expressed by the user as a viding efficient response and throughput shifts from the
nested expression involving only binary relations (see individual user to the data system.
Feldman’s LEAP System [lo], for example) then 2n - 1
names have to be coined instead of only n + 1 with direct 6 Because each relation in a practical data bank is a finite set at
every instant of time, the existential and universal quantifiers
n-ary notation as described in Section 1.2. For example, the can be expressed in terms of a function that counts the number of
6 Exploiting a relation includes query, update, and delete. elements in any finite set.

382 Communications of the ACM Volume 13 / Number 6 / June, 1970


2. Redundancy and Consistency ternary relation which preserves all of the information in
the given relations?
2.1. OPERATIONS ON RELATIONS
The example in Figure 5 shows two relations R, S, which
Since relations are sets, all of the usual set operations are
are joinable without loss of information, while Figure 6
applicable to them. Nevertheless, the result may not be a
shows a join of R with S. A binary relation R is joinable
relation; for example, the union of a binary relation and a
with a binary relation S if there exists a ternary relation U
ternary relation is not a relation.
such that 7r12 (U) = R and ‘1~23 (U) = S. Any such ternary
The operations discussedbelow are specifically for rela-
relation is called a join of R with S. If R, S are binary rela-
tions. These operations are introduced becauseof their key
tions such that ~2(R) = ~1(S), then R is joinable with S.
role in deriving relations from other relations. Their
One join that always exists in such a case is the natural
principal application is in noninferential information sys-
join of R with S defined by
tems-systems which do not provide logical inference
R*S = {(a, b, c):R(a, b) A S(b, c))
services-although their applicability is not necessarily
where R (a, b) has the value true if (a, b) is a member of R
destroyed when such services are added.
and similarly for S(b, c). It is immediate that
Most users would not be directly concerned with these
operations. Information systems designersand people con- TB(R*S) = R
cerned with data bank control should, however, be thor-
and
oughly familiar with them.
2.1.1. Permutation. A binary relation has an array T33(R*S) = S.
representation with two columns. Interchanging these col-
Note that the join shown in Figure 6 is the natural join
umns yields the converse relation. More generally, if a
of R with S from Figure 5. Another join is shown in Figure
permutation is applied to the columns of an n-ary relation,
7.
the resulting relation is said to be a permutation of the
given relation. There are, for example, 4! = 24 permuta-
tions of the relation supply in Figure 1, if we include the II31 hPPb/) (project supplier)
identity permutation which leaves the ordering of columns 5 1
unchanged. 5 2
Since the user’s relational model consists of a collection 1 4
7 2
of relationships (domain-unordered relations), permuta-
tion is not relevant to such a model considered in isolation. FIG. 4. A permuted projection of the relation in Figure 1
It is, however, relevant to the consideration of stored
representations of the model. In a system which provides
symmetric exploitation of relations, the set of queries R (supplier Part) S (part project)
answerable by a stored relation is identical to the set 1 1 1 1
2 1 1 2
answerable by any permutation of that relation. Although
2 2 2 1
it is logically unnecessary to store both a relation and some
permutation of it, performance considerations could make FIG. 5. Two joinable relations
it advisable.
2.1.2. Projection. Suppose now we select certain col-
R*S (supplier part project)
umns of a relation (striking out the others) and then re-
move from the resulting array any duplication in the rows. 1 1 1
1 1 2
The final array represents a relation which is said to be a
2 1 1
projection of the given relation. 2 1 2
A selection operator ?r is used to obtain any desired 2 2 1
permutation, projection, or combination of the two opera-
FIG. 6. The natural join of R with S (from Figure 5)
tions. Thus, if L is a list of lc indices7 L = i1, ii, - . - , ik
and R is an n-ary relation (n 2 k ), then rrL(R ) is the k-ary
relation whose jth column is column ii of R (j = 1,2, * * . ,k) U (supplier part project)
except that duplication in resulting rows is removed. Con-
1 1 2
sider the relation supply of Figure 1. A permuted projection 2 1 1
of this relation is exhibited in Figure 4. Note that, in this 2 2 1
particular case, the projection has fewer n-tuples than the
FIG. 7. Another join of R with S (from Figure 5)
relation from which it is derived.
2.1.3. Join. Suppose we are given two binary rela-
tions, which have some domain in common. Under what Inspection of these relations reveals an element (ele-
circumstances can we combine these relations to form a ment 1) of the domain part (the domain on which the join
7 When dealing with relationships, we use domain names (role- is to be made) with the property that it possessesmore
qualified whenever necessary) instead of domain positions. than one relative under R and also under S. It is this ele-

Volume 13 / Number 6 / June, 1970 Communications of the ACM 383


ment which gives rise to the plurality of joins. Such an ele- y), and T with R (say a), and, furthermore, y must be a
ment in the joining domain is called a point of ambiguity relative of x under S, z a relative of y under T, and x a
with respect to the joining of R with S. relative of z under R. Note that in Figure 8 the points
If either ~1 (R) or S is a function: no point of ambiguity 2 = a; y = d; x = 2 have this property.
can occur in joining R with S. In such a case, the natural The natural linear 3-join of three binary relations R, S,
join of R with S is the only join of R with S. Note that the T is given by
reiterated qualification “of R with S” is necessary, because
S might be joinable with R (as well as R with S), and this R*S*T = { (a, b, c, d):R (a, b) A S (b, c) A T (c, d)}
join would be an entirely separate consideration. In Figure
5, none of the relations R, 7r21(R), S, ?rzl(S) is a function. where parentheses are not needed on the left-hand side be-
Ambiguity in the joining of R with S can sometimes be cause the natural 2-join (*) is associative. To obtain the
resolved by means of other relations. Suppose we are given, cyclic counterpart, we introduce the operator y which pro-
or can derive from sources independent of R and S, a rela- duces a relation of degree n - 1 from a relation of degree n
tion T on the domains project and supplier with the follow- by tying its ends together. Thus, if R is an n-ary relation
ing properties : (n 2 2), the tie of R is defined by the equation

(1) m(T) = m(S), r(R) = {(a~, a2, .-. , a,-l):R(ul, az, ... , a,-~, a,)
A al = a,).
(2) m(T) = al(R),
We may now represent the natural cyclic S-join of R, S, T
(3) T(i s> +~P(R(& P> A S(P,~))~
by the expression
(4) R(s, P> -+ 3j(Soj\,.i) * T(.A s)),
y (R*S*T).
(5) S@,.i) + 3s(T(.?, s> A R(s, P>>,
then we may form a three-way join of R, S, T; that is, a Extension of the notions of linear and cyclic S-join and
ternary relation such that their natural counterparts to the joining of n binary rela-
mz(U) = R, 7r23(U) = s, ml(U) = T. tions (where n 2 3) is obvious. A few words may be ap-
propriate, however, regarding the joining of relations which
Such a join will be called a cyclic 3-join to distinguish it are not necessarily binary. Consider the case of two rela-
from a linear S-join which would be a quaternary relation tions R (degree r ), S (degree s) which are to be joined on
V such that p of their domains (p < T, p < s). For simplicity, sup-
pose these p domains are the last p of the r domains of R,
m(V) = R, lr23W) = s, lr34(V) = T.
and the first p of the s domains of S. If this were not so, we
While it is possiblefor more than one cyclic 3-join to exist could always apply appropriate permutations to make it
(seeFigures 8,9, for an example), the circumstances under so. Now, take the Cartesian product of the first r-p do-
which this can occur entail much more severe constraints mains of R, and call this new domain A. Take the Car-
tesian product of the last p domains of R, and call this B.
R (s P) s (P 23 T 0’ s) Take the Cartesian product of the last s-p domains of S
1 a a d d 1 and call this C.
2 a d 2 We can treat R as if it were a binary relation on the
2 b b&Ii e 2 domains A, B. Similarly, we can treat S as if it were a bi-
b e e 2 nary relation on the domains B, C. The notions of linear
FIG. 8. Binary relations with a plurality of cyclic 3-joins and cyclic S-join are now directly applicable. A similar ap-
proach can be taken with the linear and cyclic n-joins of n
u b P 8 TJ’ (s P i) relations of assorted degrees.
2.1.4. Composition. The reader is probably familiar
1 a d 1 a d
2 a e 2 a d with the notion of composition applied to functions. We
2 b d 2 a e shall discussa generalization of that concept and apply it
2 b e 2 b d first to binary relations. Our definitions of composition
2 b e and composability are based very directly on the definitions
FIG. 9. Two cyclic 3-joins of the relations in Figure 8 of join and joinability given above.
Suppose we are given two relations R, S. T is a cam-
than those for a plurality of 2-joins. To be specific, the re- position of R with S if there exists a.join U of R with S such
lations R, S, T must possesspoints of ambiguity with that T = aI3(U) . Thus, two relations are composable if
respect to joining R with S (say point x), S with T (say and only if they are joinable. However, the existence of
more than one join of R with S doesnot imply the existence
8 A function is a binary relation, which is one-one or many-one, of more than one composition of R with S.
but not one-many. Corresponding to the natural join of R with S is the

384 Communications of the ACM Volume 13 / Number 6 / June, 1970


ndural composition9 of R with S defined by 2.1.5. Restriction. A subset of a relation is a relation.
One way in which a relation S may act on a relation R to
R.S = TH(R*S).
generate a subset of R is through the operation restriction
Taking the relations R, S from Figure 5, their natural com- of R by S. This operation is a generalization of the restric-
position is exhibited in Figure 10 and another composition tion of a function to a subset of its domain, and is defined
is exhibited in Figure 11 (derived from the join exhibited as follows.
in Figure 7). Let L, M be equal-length lists of indices such that
R. S (project supplier) L = iI,&., *** ,ik,M = jI,j2, e-s , jk where k 5 degree
of R and k 6 degree of S. Then the L, M restriction of R by
1 1
1 2 S denoted RLIMS is the maximal subset R’ of R such that
2 1
?rL(R’) = TM(S).

FIG. 10. The natural composition of R with S (from Figure 5) The operation is defined only if equality is applicable be-
tween elements of ?T+, (R) on the one hand and rjh (S) on
T (project supplier) the other for all h = 1, 2, . . . , k.
1 2 The three relations R, S, R’ of Figure 13 satisfy the equa-
2 1 tion R' = Rw~wsS.
FIQ. 11. Another composition of R with S (from Figure 5) R (8 P ~3 s (P 8 R' (8 P 3
When two or more joins exist, the number of distinct 1aA a A 1aA
compositions may be as few as one or as many as the num- 2 a A c B 2 a A
2 a B b B 2 b B
ber of distinct joins. Figure 12 shows an example of two
2bA
relations which have several joins but only one composition. 2 b B
Note that the ambiguity of point c is lost in composing R FIG. 13. Example of restriction
with S, because of unambiguous associationsmade via the
points a, b, d, e. We are now in a position to consider various applications
R (supplier part) S (part project)
of these operations on relations.
1 2.2. REDUNDANCY
1 z ; ; Redundancy in the named set of relations must be dis-
1 c c f tinguished from redundancy in the stored set of representa-
2 is tions. We are primarily concerned here with the former.
2 8 :
To begin with, we need a precise notion of derivability for
2 e e f
relations.
FICA 12. Many joins, only one composition
Suppose 0 is a collection of operations on relations and
Extension of composition to pairs of relations which are each operation has the property that from its operands it
not necessarily binary (and which may be of different de- yields a unique relation (thus natural join is eligible, but
grees) follows the same pattern as extension of pairwise join is not ). A relation R is O-derivablefrom a set S of rela-
joining to such relations. tions if there exists a sequence of operations from the col-
A lack of understanding of relational composition has led lection 0 which, for all time, yields R from members of S.
several systems designers into what may be called the The phrase “for all time” is present, becausewe are dealing
connection trap. This trap may be described in terms of the with time-varying relations, and our interest is in derivabil-
following example. Suppose each supplier description is ity which holds over a significant period of time. For the
linked by pointers to the descriptions of each part supplied named set of relationships in noninferential systems, it ap-
by that supplier, and each part description is similarly pears that an adequate collection & contains the following
linked to the descriptions of each project which uses that operations: projection, natural join, tie, and restriction.
part. A conclusion is now drawn which is, in general, er- Permutation is irrelevant and natural composition need
roneous: namely that, if all possiblepaths are followed from not be included, becauseit is obtainable by taking a natural
a given supplier via the parts he supplies to the projects join and then a projection. For the stored set of representa-
using those parts, one will obtain a valid set of all projects tions, an adequate collection e2of operations would include
supplied by that supplier. Such a conclusion is correct permutation and additional operations concerned with sub-
only in the very special case that the target relation be- setting and merging relations, and ordering and connecting
tween projects and suppliers is, in fact, the natural com- their elements.
position of the other two relations-and we must normally 2.2.1. Strong Redundancy. A set of relations is strongly
add the phrase “for all time,” because this is usually im- redundant if it contains at least one relation that possesses
plied in claims concerning path-following techniques. a projection which is derivable from other projections of
0 Other writers tend to ignore compositions other than the na- relations in the set. The following two examples are in-
tural one, and accordingly refer to this particular composition as tended to explain why strong redundancy is defined this
the composition-see, for example, Kelley’s “General Topology.” way, and to demonstrate its practical use. In the first ex-

Volume 13 / Number 6 / June, 1970 Communications of the ACM 385


ample the collection of relations consists of just the follow- The relations al2(P), 7r12(Q), 7r12
(R ) are complexlo relations
ing relation : with the possibility of points of ambiguity occurring from
time to time in the potential joining of any two. Under
employee (serial #, name, manager#, managername)
these circumstances, none of them is derivable from the
with serial# as the primary key and manager# as a foreign other two. However, constraints do exist between them,
key. Let us denote the active domain by A,, and suppose since each is a projection of some cyclic join of the three of
that them. One of the weak redundancies can be characterized
by the statement: for all time, 1r12(P) is somecomposition
A, (munuger#) c A, (serial#)
of ~12(Q) with ‘~~1(R). The composition in question might
and be the natural one at some instant and a nonnatural one at
another instant.
At (managername) C At (name)
Generally speaking, weak redundancies are inherent in
for all time t. In this case the redundancy is obvious: the the logical needs of the community of users. They are not
domain managernameis unnecessary. To see that it is a removable by the system or data base administrator. If
strong redundancy as defined above, we observe that they appear at all, they appear in both the named set and
m4(employee) = ~12(empZoyee)&r3(employee). the stored set of representations.
In the secondexample the collection of relations includes a 2.3. CONSISTENCY
relation S describing suppliers with primary key s#, a re- Whenever the named set of relations is redundant in
lation D describing departments with primary key d#, a either sense,we shall associatewith that set a collection of
relation J describing projects with primary key j#, and the statements which define all of the redundancies which hold
following relations : independent of time between the member relations. If the
lJ (s#, d#, - * * >t Q(s#,j#, .-->, R(d#,j#, **->, information system lacks-and it most probably will-de-
tailed semantic information about each named relation, it
where in each case - - - denotes domains other than s#, d#, cannot deduce the redundancies applicable to the named
j#. Let us supposethe following condition C is known to set. It might, over a period of time, make attempts to
hold independent of time: supplier s supplies department induce the redundancies, but such attempts would be fal-
d (relation P ) if and only if supplier s suppliessomeproject lible.
j (relation Q) to which d is assigned (relation R). Then, we Given a collection C of time-varying relations, an as-
can write the equation sociated set Z of constraint statements and an instantaneous
m(P) = m(Q)-w(R) value V for C, we shall call the state (C, 2, V) consistent
or inconsistent according as V does or does not satisfy 2.
and thereby exhibit a strong redundancy. For example, given stored relations R, S, T together with
An important reason for the existence of strong re- the constraint statement “nlz(T) is a composition of
dundancies in the named set of relationships is user con- 5~12(R) with ~12(X)“, we may check from time to time that
venience. A particular caseof this is the retention of semi- the values stored for R, S, T satisfy this constraint. An al-
obsolete relationships in the named set so that old pro- gorithm for making this check would examine the first two
grams that refer to them by name can continue to run cor- columns of each of R, S, T (in whatever way they are repre-
rectly. Knowledge of the existence of strong redundancies sented in the system) and determine whether
in the named set enables a system or data base adminis-
trator greater freedom in the selection of stored representa- (1) ?rl(T) = rl(R),
tions to cope more efficiently with current traffic. If the (2) ?rz(T) = 7r2@),

strong redundancies in the named set are directly reflected (3) for every element pair (a, c) in the relation al2(T)
in strong redundancies in the stored set (or if other strong there is an element b such that (a, b) is in a12(R)
redundancies are introduced into the stored set), then, gen- and (b, c) is in 7r12(S).
erally speaking, extra storage space and update time are
consumed with a potential drop in query time for some There are practica1 problems (which we shall not discuss
queries and in load on the central processing units. here) in taking an instantaneous snapshot of a collection
2.2.2. Weak Redundancy. A second type of redun- of relations, some of which may be very large and highly
dancy may exist. In contrast to strong redundancy it is not variable.
characterized by an equation. A colIection of relations is It is important to note that consistency as defined above
weakly redundant if it contains a relation that has a projec- is a property of the instantaneous state of a data bank, and
tion which is not derivable from other members but is at is independent of how that state came about. Thus, in
all times a projection of somejoin of other projections of particular, there is no distinction made on the basis of
relations in the collection. whether a user generated an inconsistency due to an act of
We can exhibit a weak redundancy by taking the second omission or an act of commission. Examination of a simple
example (cited above) for a strong redundancy, and as- IOA binary relation is complex if neither it nor its converse is a
suming now that condition C does not hold at all times. function.

386 Communications of the AMC Volume 13 / Number 6 / June, 1970


example will show the reasonableness of this (possibly un- In Section 2 operations on relations and two types of
conventional) approach to consistency. redundancy are defined and applied to the problem of
Suppose the named set C includes the relations S, J, D, maintaining the data in a consistent state. This is bound to
P, Q, R of the example in Section 2.2 and that P, Q, R become a serious practical problem as more and more dif-
possess either the strong or weak redundancies described ferent types of data are integrated together into common
therein (in the particular case now under consideration, it data banks.
does not matter which kind of redundancy occurs). Further, Many questions are raised and left unanswered. For
suppose that at some time t the data bank state is consistent example, only a few of the more important properties of
and contains no project j such that supplier 2 supplies the data sublanguagein Section 1.4 are mentioned. Neither
project j and j is assigned to department 5. Accordingly, the purely linguistic details of such a language nor the
there is no element (2,5) in ~12(P). Now, a user introduces implementation problems are discussed.Nevertheless, the
the element (2, 5) into 7~12(P) by inserting some appropri- material presented should be adequate for experienced
ate element into P. The data bank state is now inconsistent. systems programmers to visualize several approaches. It
The inconsistency could have arisen from an act of omis- is also hoped that this paper can contribute to greater pre-
sion, if the input (2, 5) is correct, and there does exist a cision in work on formatted data systems.
project j such that supplier 2 supplies j and j is assigned to Acknowledgment. It was C. T. Davies of IBM Pough-
department 5. In this case, it is very likely that the user keepsie who convinced the author of the need for data
intends in the near future to insert elements into Q and R independence in future information systems. The author
which will have the effect of introducing (2, j) into al2 (Q) wishes to thank him and also F. P. Palermo, C. P. Wang,
and (5, j) in W(R). On the other hand, the input (2, 5) E. B. Altman, and M. E. Senko of the IBM San JoseRe-
might have been faulty. It could be the case that the user search Laboratory for helpful discussions.
intended to insert some other element into P-an element
whose insertion would transform a consistent state into
a consistent state. The point is that the system will RECEIVED SEPTEMBER, 1969; REVISED FEBRUARY, 1970
normally have no way of resolving this question without
interrogating its environment (perhaps the user who cre- REFERENCES
ated the inconsistency ). 1. CHILDS, D. L. Feasibility of a set-theoretical data structure
There are, of course, several possible ways in which a -a general structure based on a reconstituted definition of
system can detect inconsistencies and respond to them. relation. Proc. IFIP Cong., 1968, North Holland Pub. Co.,
Amsterdam, p. 162-172.
In one approach the system checks for possible inconsist- 2. LEVEIN, R. E., AND MARON, M. E. A computer system for
ency whenever an insertion, deletion, or key update occurs. inference execution and data retrieval. Comm. ACM 10,
Naturally, such checking will slow these operations down. 11 (Nov. 1967), 715-721.
If an inconsistency has been generated, details are logged 3. BACHMAN, C. W. Software for random access processing.
internally, and if it is not remedied within some reasonable Datumation (Apr. 1965), 3641.
4. MCGEE, W. C. Generalized file processing. In Annual Re-
time interval, either the user or someone responsible for view in Automatic Programming 6, 13, Pergamon Press,
the security and integrity of the data is notified. Another New York, 1969, pp. 77-149.
approach is to conduct consistency checking as a batch 5. Information Management System/360, Application Descrip-
operation once a day or less frequently. Inputs causing the tion Manual H20-0524-1. IBM Corp., White Plains, N. Y.,
inconsistencies which remain in the data bank state at July 1968.
6. GIS (Generalized Information System), Application Descrip-
checking time can be tracked down if the system main- tion Manual H20-0574. IBM Corp., White Plains, N. Y.,
tains a journal of all state-changing transactions. This 1965.
latter approach would certainly be superior if few non- 7. BLEIER, R. E. Treating hierarchical data structures in the
transitory inconsistencies occurred. SDC time-shared data management system (TDMS).
Proc. ACM 22nd Nat. Conf., 1967, MD1 Publications,
2.4. SUMMARY Wayne, Pa., pp. 41-49.
In Section 1 a relational model of data is proposed as a 8. IDS Reference Manual GE 625/635, GE Inform. Sys. Div.,
basis for protecting users of formatted data systems from Pheonix, Ariz., CPB 1093B, Feb. 1968.
the potentially disruptive changes in data representation 9. CHURCH, A. An Introduction to Mathematical Logic I. Prince-
ton U. Press, Princeton, N.J., 1956.
causedby growth in the data bank and changes in traffic. 10. FELDMAN, J. A., AND ROVNER, P. D. An Algol-based associ-
A normal form for the time-varying collection of relation- ative language. Stanford Artificial Intelligence Rep. AI-66,
ships is introduced. Aug. 1, 1968.

Volume 13 / Number 6 / June, 1970 Communications of the ACM 38’7