Network Working Group Richard Winter, Jeffrey Hill, Warren Greiff RFC # 610 CCA NIC # 21352 December 15, 1973 Further Datalanguage Design Concepts Richard Winter Jeffrey Hill Warren Greiff Computer Corporation of America December 15, 1973
Acknowledgment During the course of the Datacomputer Project, many people have contributed to the development of datalanguage. The suggestions and criticisms of Dr. Gordon Everest (University of Minnesota), Dr. Robert Taylor (University of Massachusetts), Professor Thomas Cheatham (Harvard University) and Professor George Mealy (Harvard University) have been particularly useful. Within CCA, several people in addition to the authors have participated in the language design at various stages of the project. Hal Murray, Bill Bush, David Shipman and Dale Stern have been especially helpful.
1. Introduction 1.1 The Datacomputer System The datacomputer is a large-scale data utility system, offering data storage and data management services to other computers. The datacomputer differs from traditional data management systems in several ways. First, it is implemented on dedicated hardware, and comprises a separate computing system specialized for data management. Second, the system is implemented on a large scale. Data is intended to be stored on mass storage devices, with capacities in the range of a trillion bits. Files on the order of one hundred billion bits are to be kept online. Third, it is intended to support sharing of data among processes operating in diverse environments. That is, the programs which share a given data base may be written in different languages, execute on different hardware under different operating systems, and support end users with radically different requirements. To enable such shared use of a data base, transformations between various hardware representations and data structuring concepts must be achieved. Finally, the datacomputer is designed to function smoothly as a component of a much larger system: a computer network. In a computer network, the datacomputer is a node specialized for data management, and acting as a data utility for the other nodes. The Arpanet, for which the datacomputer is being developed, is an international network which has over 60 nodes. Of these, some are presently specialized for terminal handling, others are specialized for computation (e.g., the ILLIAC IV), some are general purpose service nodes (e.g., MULTICS) and one (CCA) is specialized for data management. 1.2 Datalanguage Datalanguage is the language in which all requests to the datacomputer are stated. It includes facilities for data description and creation, for retrieval of or changes to stored data, and for access to a variety of auxiliary facilities and services. In datalanguage it is possible to specify any operation the datacomputer is capable of performing. Datalanguage is the only language accepted by the datacomputer and is the exclusive means of access to data and services.
1.3 Present Design Effort We are now engaged in developing complete specifications for datalanguage; this is the second iteration in the language design process. A smaller, initial design effort developed some concepts and principles which are described in the third working paper in this series. These have been used as the basis of software implementations resulting in an initial network service capability. A user manual for this system was published as working paper number 7. As a result of experience gained in implementation and service, through further study of user requirements and work with potential users, and through investigation of other work in the data management field, quite a few ideas have been developed for the improvement of datalanguage. These are being assimilated into the language design in the iteration now in progress. When the language design is complete, it will be incorporated into the existing software (requiring changes to the language compiler, but having little impact on the rest of the system). Datacomputer users will first have access to the new language during 1975. 1.4 Purpose of this Paper This paper presents concepts and preliminary results, rather than a completed design. There are two reasons for publishing now. The first is to provide information to those planning to use the datacomputer. They may benefit from knowledge of our intentions for development. The second is to enable system and language designers to comment on our work before the design is frozen. 1.5 Organization of the Paper The remainder of the paper is divided into four sections. Section 2 discusses the most global considerations for language design. This comprises our view of the problem; it has influenced our work to date and will determine most of our actions in completion of the design. This section provides background for section 3, and reviews some
material that will be familiar to those who have been following our work closely. Section 3 discusses some of the specific issues we have worked on. The emphasis is on solutions and options for solution. In sections 2 and 3 we are presenting our "top-down" work: this is the thinking we have done based on known requirements and our conception of the desirable properties of datalanguage. We have also been working from the opposite end, developing the primitives from which to construct the language. Section 4 presents our work in this area: a model datacomputer which will ultimately provide a precise semantic definition of datalanguage. Section 4 explains that part of the model which is complete, and relates this to our other work. Section 5 discusses work that remains, both on the model and in our top-down analysis.
2. Considerations for Language Design 2.1 Introduction Data management is the task of managing data as a resource, independent of hardware and applications programs. It can be divided it into five major sub-tasks: (1) _creating_ databases in storage, (2) making the data _available_ (e.g., satisfying queries), (3) _maintaining_ the data as information is added, deleted and modified, (4) assuring the _integrity_ of the data (e.g., through backup and recovery systems, through internal consistency checks), (5) _regulating_access_, to protect the databases, the system, and the privacy of users. These are the major data-related functions of the datacomputer; while the system will ultimately provide other services (such as accounting for use, monitoring performance) these are really auxiliary and common to all service facilities. This section presents global considerations for the design of datalanguage, based on our observations about the problem and the environment in which it is to be solved. The central problem is data management, and the datacomputer shares the same goals as many currently available data management systems. Several aspects of the datacomputer create a unique set of problems to be solved. 2.2 Hardware Considerations 2.2.1 Separate Box The datacomputer is a complete data management utility in a separate, closed box. That is, the hardware, the data and the data management software are segregated from any general-purpose processing facilities. There is a separate installation dedicated to data management. Datalanguage is the only means users have for communicating with the datacomputer and the sole activity of the datacomputer is to process datalanguage requests. Dedicating hardware provides an obvious advantage: one can specialize it for data management. The processor(s) can be modified to have data management "instructions"; common low-level software functions can be built into the hardware.
A less obvious, but possibly more significant, advantage is gained from the separateness itself. The system can be more easily protected. A fully-developed datacomputer on which there is only maintenance activity can provide a very carefully controlled environment. First, it can be made as physically secure as required. Second, it needs to execute only system software developed at CCA; all user programs are in a high-level language (datalanguage) which is effectively interpreted by the system. Hence, only datacomputer system software processes the data, and the system is not very vulnerable to capture by a hostile program. Thus, since there is the potential to develop data privacy and integrity services that are not available on general-purpose systems, one can expect less difficulty in developing privacy controls (including physical ones) for the datacomputer than for the systems it serves. 2.2.2 Mass Storage Hardware The datacomputer will store most of its data on mass storage devices, which have distinctive access characteristics. Two examples of such hardware are Precision Instruments' Unicon 690 and Ampex Corporation's TBM system. They are quite different from disks, and differ significantly from one another. However, almost all users will be ignorant of the characteristics of these devices; many will not even know that the data they use is at the datacomputer. Finally, as the development of the system progresses, data may be invisibly shunted from one datacomputer to another, and as a result be stored in a physical format quite different from that originally used. In such an environment, it is clear that requests for data should be stated in logical, not physical terms. 2.3 Network Environment The network environment provides additional requirements for datacomputer design. 2.3.1 Remote Use Since the datacomputer is to be accessed remotely, the requirement for effective data selection techniques and good mechanisms for the expression of selection criteria is amplified. This is because of the narrow path through which network users communicate with the datacomputer. Presently, a typical process-to-process transfer rate over the Arpanet is 30 kilobits per second. While this can be increased through optimization of software and protocols, and through additional
expenditure for hardware and communications lines, it seems safe to assume that it will not soon approach local transfer rates (measured in the megabits per second). A typical request calls for either transfer of part of a file to a remote site, or for selective update to a file already stored at the datacomputer. In both of these situations, good mechanisms for specifying the parts of the data to be transmitted or changed will reduce the amount of data ordinarily transferred. This is extremely important because with the low per bit cost of storing data at the datacomputer, transmission costs will be a significant part of the total cost of datacomputer usage. 2.3.2 Interprocess Use of the Datacomputer System Effective use of the network requires that groups of processes, remote from one another, be capable of cooperating to accomplish a given task or provide a given service. For example, to solve a given problem which involves array manipulation, data retrieval, interaction with a user at a terminal, and the generalized services of a language like PL/I, it may be most economical to have four cooperating processes. One of these could execute at the ILLIAC IV, one at the datacomputer, one at MULTICS, and one at a TIP. While there is overhead in setting up these four processes and in having them communicate, each is doing its job on a system specialized for that job. In many cases, the result of using the specialized system is a gain of several orders of magnitude in economy or efficiency (for example, online storage at the datacomputer has a capital cost two orders of magnitude lower than online costs on conventional systems). As a result, there is considerable incentive to consider solutions involving cooperating processes on specialized systems. To summarize: the datacomputer must be prepared to function as a component of small networks of specialized processes, in order that it can be used effectively in a network in which there are many specialized nodes. 2.3.3 Common Network Data Handling A large network can support enough data management hardware to construct more than one datacomputer. While this hardware can be combined into one even larger datacomputer, there are advantages to configuring it as two (or possibly more) systems. Each system should be large enough to obtain economies of scale in data storage and to support the data management software. Important data bases can be duplicated, with a copy at each datacomputer; if one datacomputer fails, or is cut off by
network failure, the data is still available. Even if duplicating the file is not warranted, the description can be kept at the different datacomputers so that applications which need to store data constantly can be guaranteed that at least one datacomputer is available to receive input. These kinds of failure protection involve cooperation between a pair of datacomputers; in some sense, they require that the two datacomputers function as a single system. Given a system of datacomputers (which one can think of as a small network of datacomputers), it is obviously possible to experiment with providing additional services on the datacomputer-network level. For example, all requests could be addressed simply to the datacomputer-network; the datacomputer-network could then determine where each referenced file was stored (i.e., which datacomputer), and how best to satisfy the request. Here, two kinds of cooperation in the network environment have been mentioned: cooperation among processes to solve a given problem, and cooperation among datacomputers to provide global optimizations in the network-level data handling problem. These are only two examples, especially interesting because they can be implemented in the near term. In the network, much more general kinds of cooperation are possible, if a little farther in the future. For example, eventually, one might want the datacomputer(s) to be part of a network-wide data management system, in which data, directories, services, and hardware were generally distributed about the network. The entire system could function as a whole under the right circumstances. Most requests would use the data and services of only a few nodes. Within this network-wide system, there would be more than one data management system, but all systems would be interfaced through a common language. Because the datacomputers represent the largest data management resource in the network, they would certainly play an important role in any network-wide system. The language of the datacomputer (datalanguage) is certainly a convenient choice for the common language of such a system. Thus a final, albeit futuristic, requirement imposed by the network on the design of the datacomputer system, is that it be a suitable major component for network-wide data management systems. If feasible, one would like datalanguage to be a suitable candidate for the common language of a network-wide group of cooperating data management systems. 2.4 Different Modes of Datacomputer Usage Within this network environment, the datacomputer will play several roles. In this section four such roles are described. Each of them imposes constraints on the design of datalanguage. We can analyze them in terms of four overlapping advantages which the datacomputer provides:
1. Generalized data management services 2. Large file handling 3. Shared access 4. Economic volume storage Of course, the primary reason for using the datacomputer will be the data management services which it provides. However, for some applications size will be the dominating factor in that the datacomputer will provide for online access to files which are so large that previously only offline storage and processing were possible. The ability to share data between different network sites with widely different hardware is another feature provided only by the datacomputer. Economies of scale make the datacomputer a viable substitute for tapes in such applications as operating system backup. Naturally, a combination of the above factors will be at work in most datacomputer applications. The following subsections describe some possible modes of interaction with the datacomputer. 2.4.1 Support of Large Shared Databases This is the most significant application of the datacomputer, in nearly every sense. Projects are already underway which will put databases of over one hundred billion bits online on the Arpanet datacomputer. Among these are a database which will ultimately include 10 years of weather observations from 5000 weather stations located all over the world. As online databases, these are unprecedented in size. They will be of international interest and be shared by users operating on a wide variety of hardware and in a wide variety of languages. Because these databases are online in an international network, and because they are expected to be of considerable interest to researchers in the related fields, it seems obvious that there will be extremely broad patterns of use. A strong requirement, then, is a flexible and general approach to handling them. This requirement of providing different users of a database with different views of the data is an overriding concern of the datalanguage design effort. It is discussed separately in Section 2.5. 2.4.2 Extensions of Local Data management Systems We imagine local data handling systems (data management systems, applications-oriented packages, text-handling systems, etc.) wanting to take advantage of the datacomputer. They may do so because of the
economics of storage, because of the data management services, or because they want to take advantage of data already stored at the datacomputer. In any case, such systems have some distinctive properties as datacomputer users: (1) most would use local data as well as datacomputer data, (2) many would be concerned with the translation of local requests into datalanguage. For example, a system which does simple data retrieval and statistical analysis for non-programming social scientists might want to use a census database stored at the datacomputer. Such a system may perform a range of data retrieval functions, and may need sophisticated interaction with the datacomputer. Its usage patterns would make quite a contrast with those of a single application program whose sole use of the datacomputer involves printing a specific report based on a single known file. This social-science system would also use some local databases, which it keeps at its own site because they are small and more efficiently accessed locally. One would like it to be convenient to think of data the same way, whether it is stored locally or at the datacomputer. Certainly at the lower levels of the local software, there will have to be differences in interfacing; it would be nice, however, if local concepts and operations could easily be translated into datalanguage. 2.4.3 File Level Use of the Datacomputer In this mode of use, other computer systems take advantage of the online storage capacity of the datacomputer. To these systems, datacomputer storage represents a new class of storage: cheaper and safer than tape, nearly as accessible as local disk. Perhaps they even automatically move files between local online storage and the datacomputer, giving users the impression that everything is stored locally online. The distinctive feature of this mode of use is that the operations are on whole files. A system operating in this mode uses only the ability to store, retrieve, append, rename, do directory listings and the like. An obvious way to make such file level handling easily available to the network community is to make use of the File Transfer Protocol (see Network Information Center document #17759 -- File Transfer Protocol) already in use for host to host file transfer. Although such "whole file" usage of the datacomputer would be motivated primarily by economic advantages of scale, data sharing at the file level could also be a concern. For example, the source files of common network software might reside at the datacomputer. These files have
little or no structure, but their common use dictates that they be available in a common, always accessible place. It is taking advantage of the economics of the datacomputer, more than anything else, since most of these services are available on any file system. This mode of use is mentioned here because it may account for a large percentage of datalanguage requests. It requires only capabilities which would be present in datalanguage in any case; the only special requirement is to make sure it is easy and simple to accomplish these tasks. 2.4.4 Use of Datacomputer for File Archiving This is another economics-oriented application. The basic idea is to store on the datacomputer everything that you intend to read rarely, if ever. This could include backup files, audit trails, and the like. An interesting idea related to archiving is incremental archiving. A typical practice, with regard to backing up data stored online in a time-sharing system, is to write out all the pages which are different than they were in the last dump. It is then possible to recover by restoring the last full dump, and then restoring all incremental dumps up to the version desired. This system offers a lower cost for dumping and storage, and a higher cost for recovery; it is appropriate when the probability of needing a recovery is low. Datalanguage, then, should be designed to permit convenient incremental archiving. As in the case of the previous application (file system), archiving is important as a design consideration because of its expected frequency and economics, not because it necessarily requires any extra generality at the language level. It may dictate that specialized mechanisms for archiving be built into the system. 2.5 Data Sharing Controlled sharing of data is a central concern of the project. Three major sub-problems in data sharing are: (1) concurrent use, (2) independent concepts of the same database, and (3) varying representations of the same database. Concurrent use of a resource by multiple independent processes is commonly implemented for data on the file level in systems in which files are regarded as disjoint, unrelated objects. It is sometimes implemented on the page level. Considerable work on this problem has already been done within the
datacomputer project. When this work is complete, it will have some impact on the language design; by and large however, we do not consider this aspect of concurrent use to be a language problem. Other aspects of the concurrent use problem, however, may require more conscious participation by the user. They relate to the semantics of collections of data objects, when such collections span the boundaries of files known to the internal operating system. Here the question of what constitutes an update conflict is more complex. Related questions arise in backup and recovery. If two files are related, then perhaps it is meaningless to recover an earlier state of one without recovering the corresponding state of the other. These problems are yet to be investigated. Another problem in data sharing is that not all users of a database should have the same concept of that database. Examples: (1) for privacy reasons, some users should be aware of only part of the database (e.g., scientists doing statistical studies on medical files do not need access to name and address), (2) for program-data independence, payroll programs should access only data of concern in writing paychecks, even though skill inventories may be stored in the same database, (3) for global control of efficiency, simplicity in application programming, and program-data independence each application program should "see" a data organization that is best for its job. To further analyze example (3), consider a database which contains information about students, teachers, subjects and also indicates which students have which teachers for which subjects. Depending on the problem to be solved, an application program may have a strong requirement for one of the following organizations: (1) entries of the form (student,teacher,subject) with no concern about redundancy. In this organization an object of any of the three types may occur many times. (2) entries of the form (student, (teacher,subject), (teacher,subject), . . . (teacher,subject)) (3) entries of the form (teacher, subject,(student...student), subject,(student...student), subject,(student.. .student)) and other organizations are certainly possible. One approach to this problem is to choose an organization for stored data, and then have application programs write requests which organize
output in the form they want. The application programmer applies his ingenuity in stating the request so that the process of reorganization is combined with the process of retrieval, and the result is relatively efficient. There are important, practical situations in which this approach is adequate; in fact there are situations in which it is desirable. In particular, if efficiency or cost is an overriding consideration, it may be necessary for every application programmer to be aware of all the data access and organization factors. This may be the case for a massive file, in which each retrieval must be tuned to the access strategy and organization; any other mode of operation would result in unacceptable costs or response times. However, dependence between application programs and data organization or access strategy is not a good policy in general. In a widely-shared database, it can mean enormous cost in the event of database reorganization, changes to access software, or even changes in the storage medium. Such a change may require reprogramming in hundreds of application programs distributed throughout the network. As a result, we see a need for a language which supports a spectrum of operating modes, including: (1) application program is completely independent of storage structure, access technique, and reorganization strategy, (2) application program parametrically controls these, (3) application program entirely controls them. For a widely-shared database, mode (1) would be the preferred policy, except when (a) the application programmer could do a better job than the system in making decisions, and (b) the need for this increment of efficiency outweighed the benefits of program-data independence. In evaluating this question for a particular application, it is important to realize the role of global efficiency analysis. When there are many users of a database, in some sense the best mode of operation is that which minimizes the total cost of processing all requests and the total cost of storing the data. When applications come and go, as real-world needs change, then the advantages of centralized control are more likely to outweigh the advantages of optimization for a particular application program. The third major sub-problem arises in connection with item level representations. Because of the environment in which it executes, each application program has a preferred set of formatting concepts, length indicators, padding and alignment conventions, word sizes, character representations, and so on. Once again it is better policy for the application program to be concerned only with the representations it wants and not with the stored data representation. However, there will be cases in which efficiency for a given request overrides all other factors.
At this level of representation, there is at least one additional consideration: potential loss of information when conversion takes place. Whoever initiates a type conversion (and this will sometimes be the datacomputer and sometimes the application program) must also be responsible for seeing that the intent of the request is preserved. Since the datacomputer must always be responsible for the consistency and the meaning of a shared database, there are some conflicts to be resolved here. To summarize, it seems that the result of wide sharing of databases is that a larger system must be considered in choosing a data management policy for a particular database. This larger system, in the case of the datacomputer, consists of a network of geographically distributed applications programs, a centralized database, and a centralized data management system. The requirement for datalanguage is to provide flexibility in the management of this larger system. In particular, it must be possible to control when and where conversions, data re- organizations, and access strategies are made. 2.6 Need for High Level Communication All of the above considerations point to the need for high level communication between the datacomputer and its users. The complex and distinct nature of datacomputer hardware make it imperative that requests be put to the datacomputer so that it can make major decisions regarding the access strategies to be used. At the same time, the large amounts of data stored and the demand of some users for extremely high transmission bandwidths make it necessary to provide for user control of some storage and transmission schemes. The fact that databases will be used by applications which desire different views of the same data and with different constraints means that the datacomputer must be capable of mapping one users request onto another users data. Interprocess use of the datacomputer means that datasharing must be completely controllable to avoid the need for human intervention. Extensive facilities for ensuring data integrity and controlling access must be provided. 2.6.1 Data Description Basic to all these needs is the requirement that the data stored at the datacomputer be completely described in both functional and physical parameters. A high level description of the data is especially important to provide the sharing and control of data. The datacomputer must be able to map between different hardware and different applications. In its most trivial form this means being able to convert between floating point number representations on different machines. On
the other extreme it means being able to provide matrix data for the ILLIAC IV as well as being able to provide answers to queries from a natural language program, both addressed to the same weather data base. Data descriptions must provide the ability to specify the bit level representations and the logical properties and relationships of data. 2.6.2 Data integrity and Access Control In the environment we have been describing, the problems of maintaining data integrity and controlling use of data assume extreme importance. Shared use of datacomputer files depends on the ability of the datacomputer to guarantee that the restrictions on data-access are strictly enforced. Since different users will have different descriptions, the access control mechanism must be associated with the descriptions themselves. One can control access to data by controlling access to its various descriptors. A user can be constrained to access a given data base only through one specific description which limits the data he can access. In a system where the updaters of a database may be unknown to each other, and possibly have different views of the data, only the datacomputer can assure data integrity. For this reason, all restrictions on possible values of data objects, and on possible or necessary relationships between objects must be stated in the data description. 2.6.3 Optimization The decisions regarding data access strategy must ordinarily be made at the datacomputer, where knowledge of the physical considerations is available. These decisions cannot be made intelligently unless the requests for data access are made at a high level. For example, compare the following two situations: (1) a request calls for output of _all_ weather observations made in California exhibiting certain wind and pressure conditions, (2) a series of requests is sent, each one retrieving California weather observations; when a request finds an observation with the required wind and pressure conditions, it transmits this observation to a remote system. Both sessions achieve the same result: the transmission of a certain set of observations to a remote site for processing. In the first session, however, the datacomputer receives, at the outset, a description of the data that is needed; in the second, it processes a series of requests, each one of which is a surprise. In the first case, a smart datacomputer has the option of retrieving all of the needed data in one access to the mass storage device. It can then buffer this data on disk until the user is ready to accept it. In
the second case, the datacomputer lacks the information it needs to make such an optimization. The language should permit and encourage users to provide the information needed to do optimization. The cost of not doing it is much higher with mass storage devices and large files than it is in conventional systems. 2.7 Application Oriented Concerns In the above sections we have described a number of features which the datacomputer system must provide. In this section we focus on what is necessary to make these features readily available to users of the datacomputer. 2.7.1 Datacomputer-user Interaction An application interacts with the datacomputer in a _session_. A session consists of a series of requests. Each session involves connecting to the datacomputer via the network, establishing identities, and setting up transmission paths for both data and datalanguage. Datalanguage is transmitted in character mode (using network standard ASCII) over the datalanguage connection. Error and status messages are sent over this connection to the application program. The data connection (called a PORT) is viewed as a bit stream and is given its own description. These descriptions are similar to those given for stored data. At a minimum this description must contain enough information for the datacomputer to parse the incoming bit stream. It also may contain data validation information as well. To store data at the datacomputer, the stored data must also have a description. The user supplies the mapping between the descriptions of the stored and transmitted data.
_____________________________________ | | / / | ______ ___________ | \ \ | | |---| | | / / | | | | DATA | | \ \ | | | |DESCRIPTION| _______ | DATALANGUAGE ___________ | | | |___________| | |<-------------------->| | | |STORED| |________| USER | | PATH |APPLICATION| | | DATA |__________________|REQUEST| | | PROGRAM | | | | |_______|<----!--------------->|___________| | | | ___________ | ! DATA PATH | | | | | | ! / / | | | | PORT |-----! \ \ | | | |DESCRIPTION| | / / | |______| |___________| | \ \ |_____________________________________| / / NETWORK Figure 2-1 A Model of Datacomputer/User Interaction 2.7.2 Application Features for Data Sharing In using data stored at the datacomputer, users may supply a description of the data which is customized to the application. This description is mapped onto the description of the stored data. These descriptions may be at different levels. That is, one may merely rearrange the order of certain items, while another could call for a total restructuring of the stored representation. So that each user may be able to build upon the descriptions of another, data entities should be given named types. These type definitions are of course to be stored along with the data they describe. In addition, certain functions are so closely tied to the data (in fact may be the data in the virtual description case -- see section 3), that they must also reside in the datacomputer and their tie with the data items should be maintained by the datacomputer. For example, one user can describe a data base as made up of structures containing data of the types _latitude_ and _longitude_. He could also describe functions for comparing data of this type. Other users, not concerned with the structure of the _latitude_ component itself, but interested in using this information simply to extract other fields of interest can then use the commonly provided definitions and functions. Furthermore, by adopting this strategy as many users as possible can be made insensitive to changes in the file which are tangential to their main interests. For example, _latitudes_ could be changed from binary representation to a character form and if use of that field were restricted to its definitions and associated functions, existing
application systems would be unaffected. Conversion functions could be defined to eliminate the impact on currently operating programs. The ability of such definitional facilities means that groups of users can develop common functions and descriptions for dealing with shared data and that conventions for use of shared data can be enforced by the datacomputer. These facilities are discussed under _extensibility_ in Section 3. ___________________________________________ _______________ | ____________ | | ___________ | | |APPLICATION | | | |APPLICATION| | | _| DATA |_|____|_| PROGRAM | | | | |DESCRIPTIONS| | | |___________| | | | |____________| | |_______________| | | ^ | HOST 1 | ______ | | | | | | | _____|______ | | | | | | DATA | | | | | | | FUNCTIONS | | | | | | |____________| | _______________ | | | ___________ | ____________ | | ___________ | | | | | STORED |__| | | | | |APPLICATION| | | | |__| DATA |____| |_|____|_| PROGRAM | | | |STORED| |DESCRIPTION|__ | | | | |___________| | | | DATA | |___________| | |____________| | | | | | | ^ | ____________ | | ___________ | | | | | | | | | | |APPLICATION| | | | | _____|_____ | | |_|____|_| PROGRAM | | | | | | DATA | |_| | | | |___________| | | | | | FUNCTIONS | |____________| | |_______________| | |______| |___________| | HOST 2 |___________________________________________| DATACOMPUTER Figure 2-2 Multiple User Interaction with the Datacomputer 2.7.3 Communication Model We intend that datalanguage, while at a high level conceptually, will be at a low level syntactically. Datalanguage provides a set of primitive functions, and a set of commonly used higher level functions (see section 4 on the datalanguage model). In addition, users can define their own functions so that they can communicate with the datacomputer at a level as conceptually close to the application as possible.
There are two reasons for datalanguage being at a low level syntactically. First, it is undesirable to have programs composing requests into an elaborate format only to be decomposed by the datacomputer. Second, by choosing a specific high level syntax, the datacomputer would be imposing a set of conventions and terminology which would not necessarily correspond to those of most users. DATACOMPUTER ENVIRONMENT | OUTSIDE ENVIRONMENT | _______ | |____ | __|GENERAL|____ | | DMS |____ | | |_______| _________ ________ _________ | | | | HIGHER | | |__| _______ ________ |PRIMITIVE|___| LEVEL |___|LOW-LEVEL|_____|COBOL | | COBOL | |LANGUAGE | |LANGUAGE| | SYNTAX |__ |SERVER |___|PROGRAM | |_________| |________| |_________| | |_______| |________| | | _______ |__|ON LINE| | | QUERY |_______ |_______| | | ___|____ |TERMINAL| | | USERS | |________| | APPLICATION APPLICATIONS | SERVERS Figure 2-3 Datacomputer/User Working Environment 2.8 Summary In this section we have presented the major considerations which have influenced the current datalanguage design effort. The datacomputer has much in common with most large-scale shared data management systems, but also has a number of overriding concerns unique to the datacomputer concept. The most important of these are the existence of a separate box containing both hardware and software, the control of an extremely large storage device, and embedding in a computer network environment. Data sharing in such an environment is a central concern of the design. Both extensive data description facilities and high level communication
between user and datacomputer are necessary for data integrity and for datacomputer optimization of user requests. In addition, the expected use of the datacomputer involves satisfying several conflicting constraints for different modes of operation. One way of satisfying various user needs is to provide datalanguage features so that users may develop their own application packages within datalanguage.
3. Principal Language Concepts This section discusses the principal facilities of datalanguage. Specific details of the language are not presented, however, the discussion includes the motivation behind the inclusion of the various language features and also defines, in an informal way, the terms we use. 3.1 Basic Data Items Basic data are the atomic level of all data constructions; they cannot be decomposed. All higher level data structures are fundamentally composed of basic data items. Many types of basic data items will be provided. The type of an item determines what operations can be performed on the item and the meaning of those operations. Datalanguage will provide those primitive types of data items which are commonly used in computing systems to model the real world. The following basic types of data will be available in datalanguage: _fixed_point_numbers_, _floating_point_numbers_, _characters_, _booleans_, and _bits_. These types of items are "understood" by the datacomputer system to the extent that operations are based on the type of an item. Datalanguage will also include an _uninterpreted_ type of item, for data which will only be moved (including transmitted) from one place to another. This type of data will only be understood in the trivial sense that the datacomputer can determine if two items of the uninterpreted type are identical. Standard operations on the basic types of items will be available. Operations will be included so that the datacomputer user can describe a wide range of data management functions. They are not included with the intent of encouraging use of the datacomputer for the solving of highly computational problems. 3.2 Data Aggregates Data aggregates are compositions of basic data items and possibly other data aggregates. The types of data aggregates which are provided allow for the construction of hierarchical relationships of data. The aggregates which will definitely be available are classified as _structs_, _arrays_, _strings_, _lists_, and _directories_. A struct is a static aggregate of data items (called _components_). A struct is static in the sense that the components of a struct cannot be added or deleted from the struct, they are inextricably bound to the struct. Associated with each component of the struct is a name by which that component may be referenced relative to the struct. The struct aggregate may be used to model what is often thought of as a record,
with each component being a field of that record. A struct can also be used to group components of a record which are more strongly related, conceptually, than other components and may be operated on together. Arrays allow for repetition in data structures. An array, like a struct, is a static aggregate of data items (called _members_). Each member of an array is of the same type. Associated with each member is an index by which that member can be referenced relative to the array. Arrays can he used to model repeating data in a record (repeating groups). The concept of string is actually a hybrid of basic data and data aggregates. Strings are aggregates in that they are compositions (similar to arrays) of more primitive data (e.g., characters). They are, however, generally conceived of as basic in that they are mostly viewed as a unit rather than as a collection of items, where each item has individual importance. Also the meaning of a string is highly dependent on the order of the individual components. In more concrete terms, there are operations which are defined on specific types of strings. For example, the logical operators (_and_, _or_, etc.) are defined to operate on strings of bits. However, there are no operations which are defined on arrays of bits, although there are operations defined on both arrays, in general, and on bits. Strings of characters, bits, and uninterpreted data will be available in datalanguage. Lists are like arrays in that they are collection of similar members. However, lists are dynamic rather than static. Members of a list can be added and deleted from the list. Although, the members of a list are ordered (in fact more than one ordering can be defined on a list), the list is not intended to be referenced via an index, as is the case with an array. Members of a list can be referenced via some method of sequencing through the list. A list member, or set (see discussion under virtual data) of members, can also be referenced, by some method of identification by content. The list structure can be used to model the common notion of a file. Also restrictive use of lists as components of structs provides power with respect to the construction of dynamic hierarchical data relationships below the file level. For example, the members of a list may themselves be, in part, composed of lists, as in a list of families, where each family contains a list of children as well as other information. Directories are dynamic data aggregates which may contain any type of data item. Data items contained in a directory are called _nodes_. Associated with each node of a directory is a name by which that data item can be referenced relative to the directory. As with lists, items may be dynamically added to and deleted from a directory. The primary motivation behind providing the directory capability is to allow the user to group conceptually related data together. Since directories
need not contain only file type information, "auxiliary" data can be kept as part of the directory. For example, "constant" information, like salary range tables for a corporation data base; or user defined operations and data types (see below) can be maintained in a directory along with the data which may use this information. Also directories may themselves be part of a directory, allowing for a hierarchy of data grouping. Directories will also be defined so that system controlled information can be maintained with some of the subordinate items (e.g. time of creation, time of update, privacy locks, etc.). It may also be possible to allow the data user to define and control his own information which would be maintained with the data. At the least, the design of datalanguage will allow for parametric control over the information managed by the system. Directories are the most general and dynamic type of aggregate data. Both the name and description (see below) of directory nodes exist with the nodes themselves, rather than as part of the description of the directory. Also the level of nesting of a directory is dynamic since directories can be dynamically added to directories. Directories are the only aggregate for which this is true. Datalanguage will also provide some specific and useful variations of the above data aggregates. Structs will be available which allow for optional components. In this case the existence of a component would be based on the contents of other components. It may also he possible to allow for the existence to be based on information found at a higher level of data hierarchy. Similarly, components with _unresolved_ type will be provided. That is the component may be one of a fixed number of types. The type of the component would be based on the contents of other components of the struct. It is also desirable to allow the type or existence of a component to be based on information other than the contents of other components. For instance, the type of one component might be based on the type of another component. In general, we would like for datalanguage to allow for the attributes (see below) of one item to be a function of the attributes of other items. We would also like to provide mixed lists. Mixed lists are lists which contain more than one type of member. In this case the members would have to be self defining. That is, the type of all member would have to be "alike" to the degree that information which defines the type of that member could be found. Similar to components whose type is unresolved are Arrays with unresolved length. In this case, information defining the length of the array must be carried with the array or perhaps with other components of an aggregate which encompasses the array.
In all of the above cases the type of an item is unresolved to some degree and information which totally resolves the type is carried with the item. It is possible that in some or perhaps all of these cases the datacomputer system could be responsible for the maintenance of this information, making it invisible to the data user. 3.3 General Relational Capabilities The data aggregates described above allow for the modeling of various relationships among data. All relationships which can be constructed are hierarchical. Two approaches can he taken to provide the capability of modeling non- hierarchical relationships. New types of data aggregates can be introduced which will broaden the range of data relationships expressible in datalanguage. Or, a basic data type of "pointer" can be introduced which will serve as a primitive out of which relations can be represented. Pointer would be a data type which establishes some kind of correspondence from one item to another. That is, it would be a method of finding one item, given another . Providing the ability to have items of type pointer does not necessitate the introduction of the concept of address which we deem to be a dangerous step. For example, an item defined to point to a record in a personnel file could contain a social security number which is contained in each record of the file and uniquely identifies that record. In general a pointer is an item of information which can be used to uniquely identify another item. While the pointer approach provides the greater degree of flexibility, it does this at the price of relegating much of the work to the user as well as severely limiting the amount of control the datacomputer system has over the data. A hybrid solution is possible, where some new aggregate data types are provided as well as a restricted form of pointer data type. While the approach to be taken is still being studied, the datalanguage design will include some method of expressing non-hierarchical data structures. 3.4 Ordering of Data Lists are generally viewed as ordered. It is possible, however, that a list can be used to model a dynamic collection of similar items which are not seen as ordered. The unordered case is important, in that, given this information the datacomputer can be more efficient since new members can be added wherever it is convenient. There are a number of ways a list can be ordered. For instance, the ordering of a list can be based on the contents of its members. In the
simplest case this involves the contents of a basic data item. For example, a list of structs containing information on employees of a company may be ordered on the component which contains the employee's social security number. More complex ordering criteria are possible. For example, the same list could be ordered alphabetically with respect to the employee's last name. In this case the ordering relation is a function of two items, the last and first names. The user might also want to define his own ordering scheme, even for orderings based on basic data items. An ordering could be based on an employee's job title which might even utilize auxiliary data (i.e. data external to the list). It is also possible to maintain a list in order of insertion. In the most general case, the user could dynamically define his ordering by specification of where an item is to be placed as part of his insertion requests. In all of the above cases, data could be maintained in ascending or descending order. In addition to maintenance of a list in some order, it is possible to define one or more orderings "imposed" on a list. These orderings must be based on the contents of a list's members. This situation is similar to the concept of virtual data (see below) in that the list is not physically maintained in a given order, but retrieved as if it were. Orderings of this type can be dynamically formed (see discussion of set under virtual data). Imposed orderings can be accomplished via the maintenance of auxiliary structures (see discussion under internal representation) or by utilization of a sorting strategy on retrievals. Much work has been done with regard to effective implementation of the maintenance and imposition of orderings on lists. This work is described in working paper number 2. 3.5 Data Integrity An important feature of any data management system is the ability to have the system insure the integrity of the data. Data needs to be protected against erroneous manipulation by people and against system failure. Datalanguage will provide automatic validity checks. Many flavors need to be provided so that appropriate trade-offs can be made between the degree of insurance and the cost of validation. The datalanguage user will be able to request constant validation: where validity checks are made whenever the data is updated; validation on access: where validity checks are performed when data is referenced but before it is retrieved; regularly scheduled validation: where the data is checked at regular intervals; background validation: where the system will run checks in its spare time; and validation on demand. Constant validation and validation on access are actually special cases of the more general concept of event triggered validation. In this case the user specifies
an event which will cause data validation procedures to be invoked. This feature can be used to accomplish such things as validation following a "batch" of updates. Also, some mechanism for specifying combinations of these types would be useful. In order for some of the data validation techniques to be effective, it may be necessary to keep some data validation "bookkeeping" information with the data. For example, information which can be used to determine whether an item has been checked since it was last updated might be used to cause validation on access if there has not been a recent background validation. The datacomputer may provide for optional automatic maintenance of such special kinds of information. In order for the datacomputer system to insure data validity, the user must define what valid is. Two types of validation can be requested. In the first case the user can tell the datacomputer that a specific data item may only assume one of a specific set of values. For example, the color component of a struct may only assume the values 'red', 'green', or 'blue'. The other case is where some relation must hold between members of an aggregate. For example, if the sex component of a struct is 'male' then the number of pregnancies component must be 0. Data validation is only half of the data integrity picture. Data integrity involves methods of restoring damaged data. This requires maintenance of redundant information. Features will be provided which will make the datacomputer system responsible for the maintenance of redundant data and possibly even automatic restoration of damaged data. In section 2 we discussed possible uses of the datacomputer for file backup. All features which are provided for this purpose will also be available as methods of maintaining backup information for restoration of files residing at the datacomputer. 3.6 Privacy Datalanguage will have to provide extensive privacy and protection capabilities. In its simplest form a privacy lock is provided at the file level. The lock is opened with a password key. Associated with this key is a set of privileges (reading, updating, etc.). Two degrees of generality are sought. Privacy should be available at all levels of data. Therefore, groups of related data, including groups of files could be made private by creating private directories. Also, specific fields of records could be made private by having private components of a struct where other components of the struct are visible to a wider (or different) class of users. We would also like the user to be able to define his own mechanism. In this way, very personalized, complex, and hence secure mechanisms can be defined. Also features such as 'everyone can see his own salary' might be possible.
3.7 Conversion Many types of data are related in that some or all of the possible values of one type of data have an "obvious" translation to the values of another. For example, the character '6' has a natural translation to the integer 6, or the six character string 'abc ' (three trailing blanks) has a natural translation to the four character string 'abc ' (one trailing blank). Datalanguage will provide conversion capabilities for the standard, commonly called for, translations. These conversions can be explicitly invoked by the user or implicitly invoked when data of one type is needed for an operation but data of another type is provided. In the case of implicit invocation of conversion of data the user will have control over whether conversion takes place for a given data item. More generally we would like to provide a facility whereby the user could specify conditions which determine when an item is to be converted. Also, the user should be able to define his own conversion operations, either for a conversion between types which is not provided by the datacomputer system or to override the standard conversion operation for some or all items of a given type. 3.8 Virtual and Derived Data Often, information important to users of data is embedded in that data rather than explicitly maintained. For example, the dollar value of an individual's interest in a company in a file of stock holders. Since the value of the company changes frequently, it is not feasible to maintain this information with each record. It is useful to be able to use the file as if information of this type was part of each record. When referencing the dollar value field of a record, the datacomputer system would automatically use information in the record, such as percentage of ownership in the company, possibly in conjunction with information which is not part of the record but is maintained elsewhere, such as company assets, to compute the dollar value. In this way the data user need not be concerned with the fact that this information is not actually maintained in the record. The _set_, which is a specific type of virtual container in datalanguage, deserves special mention. A set is a virtual list. For example, suppose there is a real list of people representing some population sample. By real (or actual) data we mean data which is physically stored at the datacomputer. A set could be defined to contain all members of this list who are automobile owners. The set concept provides a powerful feature for viewing data as belonging to more than one collection without physical duplication. Sets are also useful, in that, they can be dynamically formed. Given an actual list, sets based on that list can be created without having been previously described.
As mentioned above, virtual data can be very economical. These economies may become most important with respect to the use of sets. Savings are found not only in regard to storage requirements, but also in regard to processing efficiency. Processing time can be reduced as a result of calculations being performed only when the data is accessed. The ability to obtain efficient operation by optimization becomes greater when virtual data is defined in terms of other virtual data. For sets, large savings may be realized by straight forward "optimization" of the nested calculations. The above ideas are made more clear by example. Having created a set of automobile owners, A, a set of home owners, HA, can be defined based on A. The members of HA can be produced very efficiently, in one step, by retrieving people who are both automobile owners and home owners. This is more efficient than actually producing the set, A and then using it to create HA. This is true when one or both pieces of information (automobile ownership and home ownership) are indexed (see discussion under internal representation) as well as when neither is indexed. The same gains are achieved when operations on virtual data are requested. For example, if a set, H, had been defined as the set of homeowners based on the original list of people, the set, HA, could have been defined as the intersection (see discussion on operators) of A and H. In this case too, HA can be calculated in one step. Use of sets allows the user to request data manipulations in a form close to his conceptual view, leaving the problem of effective processing of his request to the datacomputer. Another use of virtual data is to accomplish data sharing. An item could be defined, virtually, as the contents of another item. If no restriction is placed on what this item can be, we have the ability to define two paths of access to the same data. Hence, data can be made subordinate to two or more aggregate structures. Stated another way, there are two or more paths of access to the data. This capability can be used to model data which is part of more than one data relationship. For example, two files could have the same records without maintaining duplicate copies. It will also be possible, via data sharing to look at data in different ways. Shared data might behave differently depending on how (and ultimately by whom) it is accessed. Although, the ability to have multiple paths to the same data and the ability to have data which is calculated on access are both part of the general virtual data capability, datalanguage will probably provide these as separate features, since they have different usage characteristics. Derived data is similar to virtual data in that it is redundant data which can be calculated from other information. Unlike virtual data it