AU2005200028A1 - Method and Apparatus for Maintaining Data Integrity Across Distributed Computer Systems - Google Patents

Description

Jan 2005 12:58 Smoorenburg Patent Trad 03 9712 0159 p.4

O

ct Method and Apparatus for n Maintaining Data Integrity Across SDistributed Computer Systems 00 N1

O

O 10 o i 1. Title of Invention: Method and Apparatus for Maintaining Data Integrity Across Distributed Computer Systems.

A. Inventors; Anders Vinberg, et al.

lB. Assinee: Computer Associates Think, Inc.

2. Cross-Referenes to Related Annlcatilons. If Any.

rA This application is a Continuation of U.S. Provisional Patent Application having Serial Number 60/131,019 filed on April 26, 1999, which is a Continuation-In-Part ofU.S. Serial No.

09/408.213 filed September 17, 1999, which is a continuation of U.S. Serial No. 08/829,919 filed July 15 1997, which is a continuation of U.S. Provisional Application Serial No. 60/021,980 filed July 18, 1996. Each of these related applications are incorporated by reference herein.

3. Statement as to Rghts to Inventions Made Under Federally Sonsored Research and Develoument. If Any.

This patent is not based upon any federally sponsored research and development.

4. Background and Summary Applicants' system is in the field of software-implemented methods, systems and articles ofmanufactur for maintaining.data integrity across distributed computer systems.

Several different technologies presently exist to support information processing in distributed environments. Each such technology has been designed to meet a specific purpose.

Remote Procedure Call systems, for example, permit a program running on one computer to invoke a function on another computer. Object Request Brokers provide a similar service, but with some minor variations that follow the conventions of object technology. Database access systems let a program retrieve data from a database on another computer. Messaging systems let one program communicate with another on a remote computer, sometimes storing the message if necessary and forwarding it when communication can be established. Publish and subscribe systems permit one program to broadcast a message, and only those systems that have subscribed to that message receive it. Several other technologies exist in this area.

Jan 2005 12:58 Smoorenburg Patent Trad 03 9712 0159 0 0 C3 In many ces, the communications techology provides the same services when ti commumcating with another program on the same computer, or one on another computer; in S some cases, even when commuuicating with a service within the same program. In other cases, different techniques must be used when communicating in the different configurations.

00 S5 However, the currnt state of te art imposes some practical problems. No existing service meets all ofthe requirements of modem distributed applications. The different services Sare rarely integrated, which means that a program that needs some combination of services has to tf do a lot of work to combine them. In practic, it is very difficult to do this, without introducing Ssubtle crrors that undermine the integrity ofthe data. In addition, the closed nature of many of these services often make it impractical for a developer to combine them. For example, Microsoft Corporation provides an ORB service called COM that is based on conventional, conection-based communication. Microsoft also provides stor-and-frward messaging system called MSMQ. However, COM does not support the use of MSMQ for communication, to permit asynchronous invocation of services State Shinia Tchnoloev Obet Reuestl Brokers There ae a number of systems in the present art that provide for invocation of services in a remote server. Thse are often called Remote Procedure Call (RPC) services. When they are based on an object model, they ar called Object Request Brokes. Such systems are fundamentally flawed in that they maintain the state ofthe objects an the server. When constructing a distributed system where it is desirable for client-side programs to reference the individual properties that in aggegte constitute the state of a server-ide object, developers generally choose between two options, neither of which is attractive.

The object server can expose the properties individually, using property retrieval methods ofthe type getCurentBalance and property seting methds like setCurrntBalance. But this can be very inefflcient: to retrieve a complete picture of an object's state, the client program would have to make a large number of requests. Modem network systems and database management systems are not designed to handle large numbers of small requests very efficintly: the overhead and latency of both networks and databases would make this very costly.

Jan 2005 12:58 Smoorenburg Patent Trad 03 9712 0159 p.G 0 0 Ce The object server can expose a gae&te method that retunsa a data structure that contains Sthe entire state. This is more eficient, since the entire state is shipped to the client in one S conversation, but it breaks the object model Firt, if the state is encoded as a typical sruct of common non-object languages, we have a breakdown of the programming model, intermixing 00 5 object and non-object technology. If the state is encoded as an object, we have two different 0 types of objects with very different characteistics: the state is a local, client-side object with no o methods and no relationship with the server; the original service is an object with methods but no n properties. To change the properties of the server object, the application has to make the changes o to the local state object and then ship it back to the server by invoking a method like setStatefthtate). While the technique certainly works, it is not a clean or easily maintained programming model.

In addition, after the client-side state has been modified but not yet written back to the server, we have two inconsistent versiona of the state, and processing logic would get different results depending on which version it acceaes. Because of these limitations in shipping state, it is desirable to extend Object Request Brokers with services that handle state more efficiently.

Database Access Svytema There are a number of systems that provide remote acaess to database servers. Some of these systems include automatic cache management When a record has been retrieved from the server, the appli ction can retrieve values from the record again without requiring a re-fetch, and changes to the recads are maintained in the cache and written back to the server all at once when a Commit operation is executed. Sme such systems are based on object technology, in that they present the retrieved data in the form of objects in the application's programming language.

Such systems have a serious limitation in that they retrieve the objects to the client, but they cannot then invoke methods ofthe object on the serve. Invoking methods of an object on the server would raise a difficulty for such systems once the object's state has been shipped to the client, that's where the object is maintained, its state may be modified on the client, and executing methods on the server may not be meaningfuL It should be noted that this problem also occurs with ordinary relational (SQL) databases, which commonly provide support for executing stored procedures. For example, if a record is Jan 2005 12:59 Smoorenburg Patent Trad 03 9712 0159 p.7 0 0 ci S retrieved to the client, then changes ae made to that record in the client-side cache, and if those cI changes have not yet been written back to the server, and you now invoke a srver-side stored procedure is invoked, the stored procedure would operate on the basis of incorrect data.

Because of those limitations in supporting distributed proccasing, it is desirable to extend 00 C 5 database access systems with services that manage distributed processing more consistently.

SCaching with Store Forward Technolov c Cache Manaaement t Cache management is a well-known technology: many systems, from database access otool to Web browsers, provide local caching of information to improve response time and reduce network traffic A read cache is used to keep copies ofinftmmaton that has been retrieved: ifthe application requests the same information again, it may be fetched from the cache. The cache may be transient, with information surviving only during the session, or it may be persistnt, keeping information on disk between sessions or even when the computer is turned off. Of course, if the information is changed on the seer, the cache may become stale. In some situations such as web browsers, such staleness is acceptable, and responsibility for updating the information from the serverrests on the user. In other cases, this is not acceptable, either because the information is more dynamic or because the application is more important Asynohmrous event notification of serv-side changes is a proven technique for maintaining synchronicity among the clcments of a distributed application. An application program can work with objects persistently stored in a database, and use caching for its well-known performance benefits. If another application elsewhere in the network changes a value of an object in the database, the system will send an event notification to the application, updating the value of the object The value is updated in the cache, and an event notification is sent to the application so it can update the value in its calculation or on-screen.

A write cache is used to temporarily hold changes made to the data. When a client-side applicatio makes changes to objects in its cache, those change are held in the client-side write cache. Eventually, the changes are written through to the database server. As long as the client and server are connected, the changes ae written through when a Commit operation is done in the applicaton. Depending on the strategy ofthe cache manager and the concurrency control Jan 2005 12:59 Smoorenburg Patent L Trad 03 9712 0159 p.8 0 0 ci manager, changes may be written through earier, but at a minimum the write-through is ir) completed at the Commit time.

o With a classical cache management system, both event notification (synchronizing chages from serve to client) and cache write-through (synchronizing changes from client to 00 c 5 server) operate effectively only as long as the client computer is connected to the database server.

o Such systens, however, cannot handle a sitnation when the comnection has been lost. If the database server is not accessible at Commit time, changes cannot be written through and are lost.

In Similarly, any changes that occur in the database while the systems are disconnected would be Slost, since no notifications can be sent to the client While an application can certainly respond to a failure exception by going into a pending state, waiting for me reestabliahment of the connection so the commit operation can be completed, this is an unattractive solution for several reasons. First, it places the burden of handling such problems on the application developer. Correct handling of such outages is dificult, and it is unlikely that all application developers would have the skill or the budget to handle it corectly.

Second, the application is essentially stopped during this wait; with an uncompleted transaction pending, no other database operatios can be done because they would become part ofthe same transaction, which violates the semantics ofthe application.

Further, if the application is shut down, intentionally or unintentionally, the pending state of the application is lost, and al the changes are likewise lost.

The systems may be disconnected for a number of rasons. There may be unplanned outges: network links may go down temporarily due to hardware or software failures, congestion or radio link interference. Such muplanned outages are more common today than in the past, because more systems operate in widely distributed configurationa, communicating over unreliable dial-up or wireless links. There may also be planned outages: a laptop computer, for example, may be only intermittently connected, with a sale representative using the machine to quote prices to prospective clients, and only occasionally connecting to headquarters to download price changes in summary, while existing cache management systems are usul it would be desirable to improve their behavior in the ftce of commnicaton outages.

Jan 2005 13:00 Smoorenburg Patent Trad 03 9712 0159 P.9 0 0 I It might appear that the issue of data integrity would be moot if applicatios used o convea onal, pessimistic concunercy control, by locking objects in the database. If an application holds exclusive locks on objects, other applications cannot update them, so no 00 5 notifications need be sent, and none need be queued. There are at least two practical argments 0 againtthis.

SFirst, pessimistic concurrency control is not practical in a far-flung distributed V) nviroment, certainly not in one with intrmittcat connection. An organization cannot pennit traveling salesmen to hold locks on objects in a database in headquarters-that would for example, prevent headquarters from changing prices. Experience suggest that the only practical coacurncy control model in mach widely distributed environments optimistic, in which remote applications do not hold locks in the database and instead rely on event notification.

Second, regardless of the locking regimn, changes may be made on the server by method invocations initiated by the same application. Such side effects are then propagated out to the remote application using event notification. In some cases, with long-running methods, the connection may have been broken by the time the method is completed, and hence the event notifications need to be queued in a storesnd-iorward system.

While this cenario does not appear likely in a traditional transaction processing application, where server-side methods ae shot-running, today there ae other application types that might have this need For example, an application may keep track of the archival status of files on a disk, and the method invoked may be a backup job; after the completion of the backup job, the modified archival status fags should be sent to the application, and this may need to be queued since there is no need to intempt the backup job just because a network link is temporarily interrupted.

Stoe-and-Forward M aina SYvtcms Store-and-forward is another well-known technique, where messages that are sent to a computer location are stored in a queue temporarily if the destination computer is not available, and delivered as soon a a connection can be established.

Peristence by Reachability Thnoloav Jan 2005 13:00 Smoorenburg Patent Trad 03 9712 0159 In some systems, object databases operate nder a convention that when an object of a potaetially persistent class is created in an application, it is still transient The object becomes o perslatent only when explicitly saved through the execution of some specific method or statement.

00 5 In such systems, objects may also have rferences to one another, These references may o be direct, so that an object has a property that contains a direct pointer or an address or path to Sanother object Altematively, they may be ndIrect, so them is a third object that acts as the association or link between the two objects.

Such systems have at least one potential problem a persistent object may have a S 10 dangling reftiace, a pointer to an object that was never saved and therefore does not exist when an application tries to recreate the object structure.

The common solution for this problaem is mtaccatic persistence throuagh reachability also known as "transitive persistence". Systems that use this technique automatically navigate the efmrncs, finding all objects that are reachable frm the persistent objects, and saving those as well.

However, such systems implement such persistence through reachability only within a single database More complex application systems that accommodate objects from several databses, and that support relationships among objects in separate databases, do not provide automatic rmanagement of prsistence.

Dufliate Obict asolutia Technoo In any system that retrieves data rm a databas4 there is the possibility of retrieving the same data twiee. This is true in the simplest programs that rad data from a file, and in programs that use ordinary relational tables. The possibility fbr double retrieval creates the possibility for an insidious progrn enmr, known as the last pdate. Consider this example written in pseudocode: find Cesiumr based ODUDP P,~t~ tind anothiiem bpped01 mift.

addio b4som pupeFt7 tvwe add 200W the sImU pmWpftw e seoop ;i save tie Mat etm asave the second ten Jan 2005 13:01 Smoorenburg Patent Trad 03 9712 0159 p.11 0 0 If the two first statements were coincidentally to find the same item, we would expect to t have both change applied to the property of the same item, so the property was increased by 0 300, but in Act, that would not happen The program has two copies of the original property.

Let's say that the original value was 1000, for example. The third statement of the program 0 5 would make the property 1100. The fourth stalement would make the property 1200. The fifth Sstatement would write 1100 to the database. The last statement would write 1200 to the 0 database. In effect, the addition of 100 has been lost.

n It should be noted that tranaction management or concurrency control does not solve this S problem, since the error occurs even when all these operations occur within the same transaction context. Concurrency control prevents sepaate programs from interfering, but it does not eiminate the possibility of rors in programming logic.

It could be argued that this is a straightforward erro and one that the programmer should have tested for, noticing th tthe two original read operations indeed refbrred to the same object This may be hard to do, however, because the object retrieval may be very indirect. We may have initially found two separate people and then we locate the separate departments they work in, and then we locate the manages ofthemanugrs of those dpartments. It may not be obvious that we have now gotten to the same person through two different paths. Similarly, we may have rerieved an object in one part of the program, and then in a completely unrelated part of the program, perhaps written by a different programmer, we execute a query that retrieves several objects, one of which is the same one we already fetched.

Due to the compleity of the lost update problem, no existing database systems provide a solution. However, it is possible to solve the problem and reduce the possibility of lost updates with applicants' aystem.

Obiect Databases While thepotential problem occurs in al databases, indeed in all persistent stores, it appears more disturbing with an object database with a close language binding. Because such an object database appear to be at a higher level, because it presents the objects of the database as a vast ocean of objects in which the application can seamlesly navigate, errors such as lost updates due to object proxy duplication are more rritating. Simply, developers who use object dtabases expect more than users of the sipler relational databases.

Jan 2005 13:01 Smoorenburg Patent L Trad 03 9712 0159 p.12 Dynamic concurmncv Contrl Technology In many cases, applicaion programs require the classical attributes ofconcurency control including atomicity, consistency, isolation and durability of operations performed on data retrieved from data sources. Many applications need to access both transactional and nontransactional data sources, and the disclosed system is designed to support all these providers.

Database systems have traditionally relied on locking to guarntee isolation of concurntly running transactions. The classical two-phase locking approach requires that a trsacdon lock a database resource and keep the lock until it is committed or aborted. This works well for applications that use a large number ofshort transactions.

Two-phase locking is less suitable fir modem web-based applications that are characterized by longer tranactions, lower transaction rates, and iddle tier data caching. A long rnning transaction holding a lock on a popular database resource, e.g. the number of books in stock, could prevent all other transactions from running, thus paralyzing the entire web site.

Thefre recent years have se increased interest in alternative concurrency control mechanisms. In particular, the optimistic concuracy control mechanism has been implemented in a number of database management systems and application servers.

Optimistic transactions consist of two distinct phases: a long-running read phase followed by a short write phase, also known as the commit phase. During the read phase, objects ae retrieved without locking and placed into the private transaction cache where they can be modified without affcting other transactions. Objects are written back to the shared store during the commit phase. Instead of locking, an optimistic transaction relics on the assumption that no other transaction has modified the objects while it was running. This assumption is validated before changes made by the transaction are saved in the database. It is believed that optimistic concurrency control outperforms other methods in systems with low to moderate contention.

The majority of today's e-commerc applications fit this profile.

Barlier implementation ofth, optimistic concurency control mcchanism were available as sub-componnt of larger database management systems. Very often, only data stored in these systems could be accessed in an optimistic fashion, without locking. This situation was in conflict with the trend towards infrmation portals and tnparent data accas that emerged as a Jan 2005 13:02 Smaoorenburg Patent Trad 03 9712 0159 p. 13 result of the Increased use of the Internet Web sites are often built around, data o stored in legacy data source such as relational and mainframe based databases.

Many of the modem application servers follow the traditional "star' 00 architecture, as illustrated in Figure 1. The web server and the application server S 5 processes are in the center of the star. They are connected to a number of web o browsers and to several information providers. The application server is responsible for bringing data from the information providers to the web server O clients. Data caching and optimistic transaction processing is also done In the C middle tier where the application server is located.

This architecture is suitable for applications that have only web-based, or "thin", clients and for the ones that access only a limited number of back end information providers. At the same time, it is not optimal for applications with a mixture of both 'thin" and 'fat" clients. In such a setting, a 'fat" client would need to access data residing in the cache of a remote application server, not much of an improvement compared to the traditional client/server architecture. In addition, bringing raw data from a large number of information providers to a single central location may have negative scalability implications when the data needs to be modified before It can be made available to the clients.

Accordingly, there is a need for a method and apparatus which more reliably maintains data integrity among distributed computer systems in a network.

In this regard, the present invention provides in one aspect a method for maintaining the Integrity of data stored throughout a distributed computer system, the method including: identifying an object to be persisted; determining a database provider for storing the object; and storing the object with the database provider.

The present invention provides in another aspect a method for maintaining the integrity of data stored throughout a distributed computer system, the method Including the steps of: storing a location identifier associated with a method of an object, the location identifier representing a location for processing the method; invoking the method; and Jan 2005 13:02 Smoorenburg Patent Trad 03 9712 0159 p.14 0

O

processing the method at the location represented by the location 0 identifier.

Other aspects and preferred aspects are disclosed in the specification 00 and/or defined in the appended claims, forming a part of the description of the 0 5 invention.

o Systems employing the disclosed technology, enable a network of Sdistributed computer systems to maintain the integrity of data stored across the distributed computer systems. Use of the disclosed technology accomplishes this N and other objects, features and advantages using several techniques, including: State Shipping with Remote Function Invocation; Caching with Store Forward Capability; Persistence by Reachability; Duplicate Object Resolution; Distributed Methods; and Dynamic Concurrency controls.

Brief Description of the Drawings Jan 2005 13:02 Smoorenburg Patent Trad 03 9712 0159 0 0 (C 5 6. Detaned Descrition o Applicants' system combines several known technologies in new ways, and adds new o technologies, to address these problems of existing systems. Applicants' system addresses t several sp ific problems, ach through a particular combination of technologies. The system S author also uniquely combines these servioes to provide a single infatmructure that supports service providers (methods without perastnt torng), in addition to classical databases (persistent storage without methods) and object database (persistat storage with methods).

Integrating services of different types without introducing integrity problems is quite dificut. Inded, constructing a distributed application system is difficult in any case, because errors in program logic may be apparent only when the traffic patterns combine in an unfortunate way. Applicants' system introdues several methods that addes these problems, thus reducing the difficulty ofbuilding a distributed application.

Constructing a distnributed application system with good perfnoance is also quite difficult The way the entire application system is purtitioned acoss the sveral computer systes, and the way the communication is stutured, often requires careful fino-tuning. This is a difficout task for an applicaton developer. Applicants' system rduces the burden n the developer by automating some ofthe perfoman tuniue g and by making it easier to alter the partitioning and to tune the communications without xtmnsive modification ofthe application's soaur code. Applicants' system removes many of the setting ftom the application and eables a system manager to optnize the behavior of h application system in a particular confiuration, and to modi4 its behavior in response to changes in available technology, business rquirements, distribution structue and load, without modifying the source code and with animal risk of introducing sorm.

For the purposes of this specification oertain tens will have the following defnitions in addition to their canrmo meanings a "provider which is a softwre system that provides Jan 2005 13:03 Smoorenburg Patent L Trad 03 9712 0159 p.

16 0 information or services. When the distinction is sigfficant, the specificaton refer to an n '"information provider, a provider whose main fAmotion is the delivery of information, or to a 0 "service provider", a provider whose main fimtion is delivering a processing service of ome sort. Information providers include classicl, relational and object databases, directories, fle C 5 systems, montoring agents, the Internet, and other progams that provide data. Service providers S include business applications, management tools, backup agents, and other programs.

O The distinction between service and infomnation providers however, is rarely rigorous; S for example, most modem databases provide support for invoking stored procedures, and the S Internet may be used to place an order as well as rtriving catalog data.

A "consumer" is a program that retrieves data, makes changes to data, stores data or invokes services. A consumer may be an application program of some sort, or a simple inteac ive tool such as a browser that allows a human being to interact with the information or services.

Similarly, the distinction between providers and consumers is not rigorous either. A single software component may be both a consumer and a provider at the same time. A provider may respond to requests, but in meeting those requests may acts as consumer, requesting other services from other providers.

In addition, the information flow between a consumer and a provider is not always structured as a classical request/response. A provider may send event notifliations or other messges to a consumer, or to other providers.

The system provides services that allow software and hardware systems to communicate, whether they are located in the same process on one computer, in different processes on one computer, or on different computers; whether they act as consumers or providers, of information or services; whether they send or respond to requests, or send or react to events.

State Shninfr combined with Remote Funetion Invocation Applicants' system combines state shipping with remote method invocation. When an object is accessed, its state is shipped t the client and stmed in the client-side cache, in the manner ofpreent-art database acoes systems. This ombination is accomplished under a rigorous object model, and the objects arc eposed to the application as native objects in the application language, through a language binding. The state of the objects may be accessed Jan 2005 13:03 Smoorenburg Patent Trad 03 9712 0159 p.17 0 0 S directly, and these acceses are resolved locally from the cache. The state maybe updated n directly, and the changes are held locally in the cache, and written back to the server later 0 under a lazy-write algorithm. The lazy-write algorithm may decide to do the write-through at various times, depending on the concurency control model used and on optimization decisions, c 5 but at the vy latest it is written when a Commit operation is invoked. Server-sid methods are exposed through this language binding in the foam ofstandard methods of the application 0 programming language.

SPro- and Post-Method Svnchronization SSince thse methods execute on the serve r on some other computer, in the case of [0 distributed methods the state should be synchrozed betwem the client and srver or other affected computers. Hence, when a server-side method is invoked, the cache manager writes through to the server all changes made to the state of the objects in the client application before the sverve-side method is actually invoked. Of course, such synchronization is not necessary when executing client-side methods.

It is possible to modify the logic of the state synchonization service of Applicants' system to optimize the amount of infimation written through. Some of the state changes may not be relevant to the method, and hence would not need to be written through at this time.

However, in the general case the state synchronization service cannot determine this, since the methods may be implemented in a multitude of languages and may be arbitrarily complex; therefore, to be on the saf side, it should writ through all property changes. Of course, such manual control would be a possible variation ofApplicants' system.

After the srver-side method has been invoked, the may be a need to synchroniz the client-side cache with the database. The method may have had side effects, difying the state of the object it belongs to, or that of other objects that ar also in the client's cache. Therefore, after a method invocation, the state synchronization service automatically synchronizes all objects in the cache that have been modified inthe saver.

Methods that access the database purely through the present Infstructure pose no problem. The infraucture track all activity going on, detenmine what hanges have been made, and can easily send notification of all such changes up to the cache. If the methods access the database through direct techniques that the infastrctur cannot track, the Jan 2005 13:04 Smoorenburg Patent Trad 03 9712 0159 P.18 0 0 ixnkastructurc may be able to rely on event notification from the database. Most dtabase S systems permit a program to registe fr notification on changes, through "triggers" or other 0 techniques, and the infastrctur can use such notifications m the basis for cache 00 synchrouLzatio C 5 If the linfiastructme determines tat neither technique is available it should take a o psaasimitic approach, invalidating and refreshing the sentire cache. Note that in this case, no data is lost in the cache, since all pending changes will be written through before the method S invocation.

O ANoication vent Notification E Any such "ide-ffect" changes to objects that occur as a consequence of the method invocation are not only used to refrah the cache, but also passed back to the client-side application as regular change notifications, to permit the application to take the new value into account in its computation or to display it in the user interfaeo. Thee is no fundamental difference between a change made by another program mad a change caused by a side effect of a method: the applicatio needs to be notified ofboth.

Transaction Mnanent Under classical pessimistic concurrency control, the application holds locks on records that it has read. In such cases, the entire discussion about pro- and post-method synchronization does not involve transaction management at a. The sequence of steps in the application looks like this: i eg V..

All ofthen operations work widn the samne pessimistic transaction context. The only unusual offect is that the application has to be prepared to deal with change notifications appearing due to method side effects. While under ordinary cirouawtances, an application operating under peslmistic cncurrency control has not traditionally been concemed with Jan 2005 13:05 05 Jn 205 1z05 Smoorenburg Patent Trad 03 5712 0155 .1 P. IS chang notification on objects it controls, in today's comnplex, multi-ftheaded applications, some tndata held in one window or thread may be modified by another Window or thread, 3o in reality soy piece of cods has to be prepared to accept change mioificatons on any piece of data that it shareswt any other entity.

00 Ni 5 When operating in an optimistic concimeacy control regime, however, the application o holds no lcks on object. in the databas. Amy timne cbanges we written to the database, they are o validated and immediately cozumiftLd Since-any pending change. arc written to the database before invcking the metod, te method invocation wamaticufly commits penading transactions.

00 amnd padhiechange to SaWrW L S validd tern aginst aunisfoftedabe sand an wAxcapn to the ant .hwlemisead it canithnito 'A -14 Applicants' system alibs aohroption,- dimscusdin detail ith rspec to Dynamnic Concurrency Control. The trasaction. management service can switch over to pessimistic concununcry control when the method is invoked. These execution options may be selected as part cithe configuration settings of the 'yatn.

CSA~gfMlUS SS Fno Give the iitatloa oftonvwentionul sysems, it may be desirable to have an infrliuReO tha cMM SaPt itransaction commit even when the connection is broken, and can mnaintain tho transaction in at puauitent store-snd-tbrwsrd queue, sending it through when die connection is reestablished.

Applicant' system uniquely ocatbines these two capabilities, cache management and soeand fr-wcd. With stome-avdeibrward meassaging &~eding thbe cache nmaement system,, all notifications. of serve-side changes that occur during an outage are stored in a queue, and as soon Jan 2005 13:05 Smoorenburg Patent L Trad 03 9712 0159 0 0 c as the connection is restablished the changes ae propagated and reflected in the cache and the W application.

SSimilarly, notifications of client-side changes ar also stored if necessary, and forwarded to the erver as soon as the connection is reestablished. Since such asynchronous write-though 0 5 to the erver may have problematic interaction with concurrency control mechanisms, it may be o disabled by the application program or by an administrative policy.

o A lication Tranmarncv n An advantage of the interation of the present invention is that the application need not 0 o be aware of the intenmittent synchronization process. The application may be written as if it had continuous and reliable access to the database, with the language bindings providing immediate access to avast pool of objects. The cache improves the performance of the application, the store-and-forward queuing nmresw that all changs are propagated to the application, all without any change in the style ofprogramming in the application.

Resilience To anlanned Outaae This integration of stom and-forward queuing with cache management increases the resilience of the entire system to short-term unplanned outages. Such an unplanned outage may be very short, especially with dial-up or wireless linkages. For example, a cellphone connection may be lost while a vehicle trvels under a bridge, and re-established within seconds or minutes, The appliation and cache manager may continue to run during this outage. Yet, notifications from the sver may be lost.

In practice, most networkprotocols would detect the lost connection and immediately signal the failure to the database server, which would then consider the whole operation lost.

However, with the store-and-forward technique, the database server continues with its operation in the assurance that messages will get through eventually, and no failure messages are sent to the database.

Continuallv Rmin Apploations It is becoming increasingly common, for an application to be continually running while the connection to a database server is intermittent. Vehicles, ships and airplanes may have contnually running applications dealing with inventory, dispatch or routing ultra-lightweight portable computing devices, including so-called Personal Digital Assistants, may run continually.

Jan 2005 13:06 Smoorenburg Patent Trad 03 9712 0159 p.21 0 0 c In these caues the cache may be kept alive continuously, and change notifications must be V queued up while the connection is down.

persistent Cachine The value of this integration of caching and store-and-forward is increased further when 00 5 the cache is made persistent Without a persistent cache, the benefits discussed above accrue o only while he application, or at least the cache manager, is kept running. With a conventional, S non-pesistent cache, the values in the cache ar discarded as soon as the cache manager is shut Sdown. When it is restarted, the cache manager has to fetch the objects afresh from the database, 0 and since those values of necessity reflect the current value, there is no need to queue up the change notifications that occured in the meantime.

However, in many cases it is highly beneficial to use apersistent cache that is persisted onto local storage when the application or cache manager is shut down, and than recreated from the local store when the cache manager is restarted. In some cases, persisting the cache is the responsibility of the cache manager. In others cases, for example when operating on a laptop, it is the responsibility of the operating system.

Whatever component is responsible for persisting the cache, the situation is logically equivalent to a long-running cache, and atore-and-forward queuing of change notifications is used to maintain the currecy of the information in the cache Applicants' system may be contrasted with replicated database which can be configured to provide some similar benefits. Database replication schemes however, rely on database homogeneity, which means that an identical, or at least structurally and semantically similar, database be available at every node. On small computers, such as laptops or palmtops, this is not practical In any case, Applicants' system provides these benefits with merely a persistent or non-persistnt cache, which is a much useless burdensome technology than a database.

Perristence by Reteh abflht Applicants' system extends the concept of persistence through reachability across several information providers. Whenever the infrastcture decides that an object needs to be persisted, through the persistence-though-reachability algorithm or because the application program explicitly requested it to be persisted it determines to which database provider the object is to be Jan 2005 13:06 Smoorenburg Patent Trad 03 9712 0159 p.22 0 0 Spersisted. This may be determined through any one of a number of teahiques, including for example: O The lass that the object belongs to may be associated with a schema defined in a particular database, and the object is to be stored in the same database 0 5 The class that te object belong to may hae a particular database store specified, o regardless of where the schema came from 0 S* The class that the object belongs to, or the particular object, may have been associated .t with a particular stoe dynamically 1 The system may decide which store to use based on some criteria of its own C 10 In these cases, the two objects that have a mrachability relationship may be persisted to different databases. The relationship between them may be implemented through a reference that can accmmodate differet stores.

In these oIrumstances, the persistence by reachability algorithm relates in substantially the same way regardless ofwhere the objects are to be stored. As long as the objects are held in the cache memory or in th application program's own memory, the relationships between the objects are known to the persistence infastructure. The persistence by reachability aspect of Applicants' system navigates these references and makes its determinations which objects should bepesisted in the conventional manner, by identifying all objects that are reachable from each persistent object, and maring them to be persisted. It then uses one of the techniques listed above to determine to which database the object should be persisted, and then effects the pesistence in a conventional manner.

Duolicate Obtect Resontion Applicants' system solves the problem of duplicate object resolution by automatically detecting when an operation would retrieve from the database an object that has already been activated or fetched flm the sever to the client. Since the system maintains in its cache all the objects that have been activated, whenever an object is fetched the system detennines if the newly fetched object would duplicate an existing object in the cache, and if so, discards the new copy and instead ues the existing copy in the cache.

This technique applies whether the object is fetched from the database directly, by name; by following an interobject relationship, an association or pointer from one object to another; by Jan 2005 13:07 Smoorenburg Patent L Trad 03 9712 0159 p.23 0 0 ci executing a quey that retumn a number ofobjects that match some specified criteria; by lookup It in a list of ecently visited objects; orby any other technique. How the object is arrived at does not matter; when it arrives, the system detects that it is a duplicate and discards the new copy.

Performance Improvement Variations 00 Ci 5 Of course, in many caes the system can detect that an object to be fetched already exists o in the cache and eliminate the request to the database server altogether, thus reducing network traffic and database load, and improving response time. In other cases, such as when executing a t query, utilizing the database cannot be avoided. In any case, these variations only affect o porformance, and do not alter the fundamental operation of the system.

Distributed Methoda An object database stores objects, and those objects have both properties and methods.

The schema defines which methods exist for a given class of object.

In a conventional object database, the methods are implemented in the same object database. The definition ofthe method simply give its name and the class to which it belongs, and no further location specification is required since its location is implicit in the location of the object.

There are circumstances however, in which it might be necessary or desirable to implement a method somewhere else. For example: The database may be less than completaly capable. For example, if objects ar stored in a relational database, the ability to execute methods may be limited. In such a case, the methods may be Implemented as separate programs, perhaps executing on the sme machine where the database i, perhaps executing on another one.

Given an object database with complete method execution capability, it may be desirable to replicate data to a less capable stor. Replication is often done for reasons of improvd availability, reduced network trafi and improved response time. However, if the replicated database does not provide support for method execution, the system may have to go back and execute methods in the original object database.

The objects that are stored in the database may represnt physical pieces of equipment or software elsewhere in the system. Such an object,-a router, for example-may have J a an 2 0 05 13: 0 Smoorenburg Patent Trad 03 9712 0159 p.24 0 0 capabilities implemented as programs that operate in the object itself; management Sservices may provide functions that are implemented elsewhere; and the application may provide some mctions that involve a graphical user interface and hence operate 00 on the user's workstation. In all these cases, these services may be viewed as methods of the object, even though they may be implemented all over the network.

The methods of a class may be implemented in legacy programs or third-party programs over which the developer has little control. The methods may be o implemented in different programming languages, databases or operating systems.

0 Nonetheless, in order to simplify the use of these methods in application development, it isdesirable to present these various methods as parts of one cohesive whole. It is ofcourse possible to build a piece of software that ties together these sparate method implementations, but building such a connecting service is a lot of work, and any change in configuration requires changes to the code. It is desirable to define and maintain the relationships etmaUlly, through administration ofschemas, regardless of type and location of the methods.

Therefore, Applicants' system permits the schema to include definitions of methods that specify the location of the method, not just its name. During execution, the system permits an application to access an object in the database, and simply invoke a method on that object as if it were a standard, built-in method; the system will take care of dispatching the method invocation request to the appropriate location.

Method hrmlementation The definition ofa method should also specify how the method is implemented. It may be for example, a C or C-f+ DLL, invoked with a normal subroutine call. It may be a Java program, an executable program, or even a BAT fila The schema identifies which technology is used, and how the method should be invoked.

The method should also specify how the object is identified. When a method is implemented as an extemal program, it of oprse does not have the context of the original object The method schema specifies how the invocation is to be done, in such a way that the context can be preserved.

Class and nstance Level Methods Jan 2005 13:08 Smoorenburg Patent Trad 03 9712 0159 0 0 ci The technique for handling etenmal implementations of methods works equally well for Sinstance-level methods that are attached to an object as for class-level methods that are attached to the class. When invoking an external method that is configured as an instance level method, 00 the identity of the originating object is passed in to the method. The way this is done for C 5 example, the iddeatfy could be passed is specified in the schaa, as part ofthe method definition o as an argument in the call, as a command line parameter when launching an executable, or 0 whatever is appropriate in the specifi configuration.

Pro- and Post-Method Smconization 0 As previously discussed, a system that combines remote method invocation with state shipping should take car of ynchronizing the state between the liet and the server before and after a method execution. In the case ofdistributed methods, this state synchronization should include the computer on which the method is to be executed.

This maybe done in at least the following ways: The state may be synchronized to the database that serves as the persistent store of the object. The method can then retrieve the state fiom the database before xecuting.

During the execution of the method, the system writes any persistent state changes to the database. The client-ide cache manager is notified of these state changes and uses them as the basis for post-method state synchronization Note that if the method is running on a computer that already has a copy of the object in its cache, it may not need to fetch the object from the database.

The state may be synchronized to the cache of the server that is to execute the method.

Dynamic Concurreney Control Applicants' system employs a unique dynamic concurrency control mechanism tha is based on a combination of locking (pessimistic) and non-locking (optimistic) concurrency control strategies.

Bach instance of the runtime system contains its own object cache and its own transaction manager. It may be run on a client machine where it can provide local data access for non-web based applications. Several instances of the runtime can be grouped together in the middle tier.

There they can be accessed by the web clients and by other runtime instances. Automatic load Jan 2005 13:09 Smoorenburg Patent Trad 03 9712 0159 p.26 balancing cnsures improved scalablity. A supate rUntimn instance can be placed close to a legacey database and rid to perfonrm objeot-to-relatonal mapping& Objects are deliverd to the middle tier in an assembled form, which reduces the number of anetwork mesages requited to 00 aess the objects.

SN architeture nessfully resolves many non-trivial issucs asaociatcd with transparent o inftnmation access in a distributed environment.

N~ Distributd Validaton 0 All of the exiting application servers and transparent data access frameworks perform .0 validation of optimistic transactions in the middle tier, the place where the shared object cache and the private transaction caches are located. Validation is often done by comparing date stamps of objects in the private transaction cache against date stamps of the same objects in the shared cache. Validation fails if the stamps are diffbreL Some systems perform validation by comparing befbre and after images ofobjects in different eaches of the application server awlmrnmet An optimistic transaction may access data via a number of instances Of the nmtime systm. For example, a transaction started by the application "Appl" shown in Figure 2 passes though at least three runtime instances when accessing data in the information provider "DB 1".

In a prfcrred embodiment, eachof the runtime instances maintains the read and the write ets of the transaction. When the transaction commits, its rmad set is compared with the write sets of transactions that have already commted. The transaction is aborted if a non-empty intersection of the sts is fbmd in at least one of the runtime instances.

Provia Specifc Validation In some cases, the rules for how a transaction should be validated may be specific to the application. For eample, consider a database of employoe records. If one application has changed a phone number of a record, other applications can be considered valid, unless they happen to modify the same phone number. But if the employee name is changed, the identity of any transaction that attempts to modify any part of the record is questionable, and any such transaction should be invalidated. This knowledge, that a phone number is insignificant but a name is significant, may be specific to the application.

Jan 2005 13:08 Smoorenburg Patent Trad 03 9712 0159 p.27 (-r Applicants' system allows the provider to take over the responsibility for transaction tr validation by registering itself as capable of doing this, using the standard interface between providers and the rntime. If the provider claims this responsibility, the runtime system asks the provider to validate. The provider ifai to validate the nmtime it relies on its own, more N 5 conservative technique. In a configuration with several providers, while others may do their o own, application-pecific validation, some may defer the validation to the runtime system.

C Parallel Validation o The read and write acts of a distributed tranaction are partitioned across multiple o instances of the runtime system. When the transaction commits, its validation is performed in parallel in all of the involved rntime instances. In an ideal case, when a transaction accesses N infnnmation sources via N different instances of the runtime running on N different CPUs, the validation can be up to N times faster compared to the time it would take if only one central application server was involved. Given the distributed nature of the enterprise wide data access, this case will be increasingly common.

Early Abort Prior to Validaon While an optimistic tranaction is pending, other programs may modify the data that is used by the transaction-this is a fundamental principle of optimistic concurrency control. The conflicting modifications to the data ar detected during the validation phase, using one of a number of techniques. This also is common in optimistic systems using Applicants' system However, the event notification system can send notifications of any such conflicting changes to the runtime system, allowing it to abort the pending transaction early. This early abort feature reduces wasted effort by the human operator, who might otherwise expend effort entering data into a transaction that is eventually invalidated. This early abort feature also reduces wasted load on computers, networks and databases.

Early Abort Durine Ilibuted Validation A long-running optimistic transaction need not wait until the end of its read phase in order to discover that it has been invalidated by another transaction. Once a committing transaction passes the validation phase, it compares its write set against the read sets of transactions that are curently active, and sends invalidation events to all transactions for which the ntersection of the sets is not empty. The invalidation events are propagated via the Jan 2005 13:09 Smoorenburg Patent Trad 03 9712 0159 p.28

O

O

distributed event manager that among other features is designed to provide guaranteed ad intf order event delivery.

C-ybrid Transactions Systems supporting optimistic transaction maagement do so on a per-transaction basis.

00 c 5 An application indicates the mode of a transaction when the transaction is started. A new 0 S transaction can be started in either pessimistic or optimistic mode, 0 The mode of a transaction can be different not only on a per-transaction but also on a pero provider basis. For example, an application may access two data stores: a private lightweight 0 database used as the local peristent object cache, and a remote departental database containing shared data. Applicants' system allows the application to access the local database in the pessimistic mode, thus avoiding the unnecessary validation overhead. The shared database can still be accessed in the on-locking optimistic mode.

This feature bccomea particularly important when the invention is used to access nontransactional providers such as LDAP, ADSI, MAPI and the NT file system. Optimistic concum cy control may provide isolation fbr transactions accessing such providers when mdializability is required. It can also be disabled when it is not necessary.

Uniaue wraoer API Modem database systems provide varying derees of support for optimistic transaction management. Unlike other transparent data access framework that insist on using their own algoithms for managing various phases of an optimistic transaction, Applicants' system is capable of leveraging unique capabilities of different information providers through development of customized wrappers. The wrapper API contains basic function of the optimistic concurrmncy control that can be re-defined by the wrapper developers. This is not the case with the ODBC, JDBC, and OLE DB APIs used by other systems.

Dynamic Transactions In Applicants' system, the state of an object values of its properties) is shipped to the point where the object is used. The application "App shown in Figure 2, for example, is accesing objects in the cache of the local runtime instance. This is a clear improvement compared to other systems that either leave the object on the serv side or in the middle tier. Of Jan 2005 13:09 Smoorenburg Patent Trad 03 9712 0159 p.

2 9 0 Scourse, when a server side object method needs to be invoked the state of the object is on the client mand it may be required on the server in order to run the method.

SApplicants' system accounts for this by enabling the transaction manager switch a transaction from an optimistic to the pasimistic mode dynamically while the transaction is 0 0 rmming. The mode switch is performed transparently for the application at the moment when the Sapplication invokes a server side method. Only the provider that is used to run the method is 0 afected. All other providers can continue runing in the optimistic mode.

it Note that the mode switch is only necessary if the provider itself does not support a form o of optimistic concurrency control. Otherwise, the changes can be saved in the provider as part of the distributed optimistic transaction. For example, the mode of a distributed transaction is never changed when the state of an object is propagated among various instances of the runtime system. This feature is valuable for supporting plug-in service providers, such as the system's own catalog administrator.

Adaptive transactions IS Optimistic transactions are not suitable for applications with high contention on data items accessed by the clients. In such an application, a large number of optimistic transactions would not be able to pass the validation phase causing an unacceptably high rollback rate. While starting all transactions in tho pessimistic mode is a straightforward solution to the problem generally used by other systems. It is too limiting in cases when the contention level changes over time. For example, in a real-world situation contention may be high between 8 am. and 6 and low during the rest of the time.

Applicants' system provides a solution to the problem is facilitated by the neural network agent technology. One example of this technology is more fully disclosed in U.S. Patent Application Serial No. 09/084,620, which is incorporated herein by reference. Transactions are always started without explicit definition of their mode. When the contention is low, they are nm in optimistic mode. When the rollback rate grows above a certain limit, the default mode is automatically changed to pessimistic. The rollback rate can be consistently kept within acceptable limits once the Neutral Agent accumulates sufficient knowledge to allow reliable prediction.

Sassion.Wide Event Notification Prior to Commit Jan 2005 13:10 Smoorenburg Patent Trad 03 9712 0159 0 0 SThe infrastructure of the Applicants' system extensively uses event notification, usually I from providers to consumers. For example, a consumer may be notified that objects bold in the client-ide cache under optimistic conurrency management have been changed in the database, and hence the current transaction may need to be invalidated. Applications and providers can 00 C 5 also send events to one another in a generic way.

SIn addition, it is common that the different elements that make up a client-side application Sneed to communicate among one another. For example, a GUI application may consist of Smultiple windows, some running in the same thread, some ruming in diffbrent threads in the o same process, some running in different processes.

Such different application elements may, for example, be presenting the same data item in different views or different contexts. Of coure, once one application element makes a change and commits it to the provider, all applications interested in that obj ect are notified of the change.

However, within one client-side application, event notification must occur long before the change is committed. Of course, if all the application elments are part of the same process, they will share data through the common session and common cache. This leaves open considerations: elements should be able to cooperate when they are in different processes? The user sees the different windows as part of the same application, and is not aware of or interested in onfiguration details, sch as threads and processes. To the user, one application should present one set of data.

Regardless of how the data sharing is physically done, the various application elements such as GUIs, need to be notified when some data changes. This burden should not be born by the developer, becase as the nmnber of elements in the application grows, the number of notification relationships grows combinatorically.

Whenever a new component is added to the application, other components that could potentially access the same data as the new component should be changed accordingly. This would be an expensive and effectively unmaintainable architecture for a developer. Instead, the inastructure should provide a notification service that automatically sends notifications among components when some data changes, without waiting for a commit point, and without regard to thread and process Jan 2005 13:10 Smoorenburg Patent Trad 03 9712 0159 p.31 0 ci c boundaries. Of course, session boundaries are honored as it is mndamental to tr transaction management, that sessions are isolated from each other until changes are Scommitted.

To address these issues, Applicants' system provides two facilities: S 5 1. When an application component connets to the infastructur, it can join an existing Ssession.

O 2. Change event notifications are propagated within the session before the commit point, lt and without involving the provider.

0 Summary of Prefaerd Imllementation Hints The object infrastructure is designed to offer significant advantages of reduced development effort, reduced programming errors and improved performance by transferring to the infrastructure many of the responsibilities that would ordinarily reat on the application programmer. This applies also to the work of developing graphical user interfaces (GUIs).

Applicants' system enables a genric user intrface to be built based on prsenting the state of objects in forms, navigating associations between them, and presenting methods as entries on menus or as buttons.

Applicants' system provides away for the database designer or application developer to guide the workings of the infrastructur through the definition of hints.

Isolatin the Aplication From Imrlementation The hints of the present invention not only provide direct advantages in improving the workings of the infrastructure or reducing the effort of the developer but also permit the developer to tune the behavior of an application without dropping down to a lower level of abstraction. By maintaining the programming conventions at a high level of abstraction, most of the advanced functionality of the system remains functional Examples of such high-level functionality that might be undermined by explicit tuning at a lower level of abstraction include: Cache management Read-ahead Write-behind Duplicate object elimination Jan 2005 13:11 Smoorenburg Patent Trad 03 9712 0159 p.32 0 0 Through the use of hints, a developer or database designer can tune the behavior of the system without any explicit tuning-dependent statements in the application program.

SAdvantageously, the system may be re-tuned in response to changes in the available hardware, distributed configuration, load patterns or other factors, without requiring any changes to the 0 0 5 application.

O Adapt to Exteral Obiect Model o The general object management system of Applicants' system is capable of n accommodating different kinds of infomation and service provides. Its general object model can o be adapted to any system that meets the most minimal characteristics of an object system.

Using Applicants' system, a wrapper, an interfacing module for a provider, is written specifically to a provider, and can expose any number of special hints that describe the characteristics of the provider. Applications that are aware of the meaning of these hints can take full advantage of the advanced characteristics of the provider. Applications that do not know anything about the provider operate in the normal way, based on the standard object model.

End of Comments.

Examples of Hit Performance Tuning Hints: Many providers may specify hints that describe likely optimal ways to process information, given the semantics of the information: A projection list-a list of the properties that are most likely to be used by the application, and hence recommended to be included in any retrieval; or Frames used for read-ahead-a list of associations or other relationships to be used in grouping objects for read-ahead.

GUI Hints: A provider can specify hints that suggest how an automatic GUI generator should display the infomation for example: Grouping of properties onto tabs of a form; Whether an associated object should be displayed in a general navigation structure, such as a tree or network browser, or as a special page or tab on a form What default property should be used when an object is displayed; What property should be used as the name, caption, "tool tip" text, long description, help text, icon, 3-D representation, sound or video of an object; Jan 2005 13:11 Smoorenburg Patent Trad 03 9712 0159 p.

3 3 0 ci C What entries should appear on a contet menu for an object; How a GUI should handle navigation, such as entry of an object; O How auto-layout telmniques should rrange the contents of an object, under some specific association type; 00 5 How numeric values should be formatted on display; O What values are accepted on data entry; or o How a drag-and-drop operation is to be interpreted.

t1 Source of the Hints 1 Such hints may be provided by either the provider or by the consumer. Without any 10 hints, the system will work in a standard way. Hints from either the provider or the consumer will guide the workings ofvarious syste services. If the various hints are in conflict, it is up to the service that uses these hints to decide what to do.

Open and Extensible Hints are read and interpreted directly by the various services of the infrastructure, or by the provider, orby the application itself Since hints are an offioial part of the inastructure's data model, any component can create hints, and any component can use them. Further, since the entire infrastructure is open and extensible, any service may be replaced by another, perhaps one that uses another strategy for interpreting the hints.

Hints Never Carse a Failure The object infastructure defines the semantics of the operations it offers. The hints may give guidance to the infrastructure in how to optimize the operation, but they should not be allowed to be restrictive.

Semantically, hints are a form ofside-band oommunication. They are not to be confused with regular object properties. Applicants' system maintains a distinction between hints and regular object properties. Consequently, the definition ofnew hints, or the deletion of some hints, is not like a change in the schema, and does not require a recompilation of the program.

Amplication Procesaing Hints Since the hints are handled as a part of the data or metadata normally processed in the system, any component may define and use them. This means that the various parts of a distributed application can use hints to communicate about how processing is to be done. Using Jan 2005 13:12 Smoorenburg Patent Trad 03 9712 0159 p.34 0 0 an agreed-upon messag convention, a provider can direct a GUI component to bring up related objects and analyze the structure of the relationship.

O Such special codes utilized in existing application systems, often requires complicated special techniques ofsideband communication. These special codes have to be maintained. For 00 5 example, if the system is to be switched over from SNA to TCP/IP, the main database access 0 protocol has to be poited, and the sideband messaging protocol has to be ported as well. With o the application processing hint facility of Applicants' system, the special hints are logically in a t sideband, outside of the normal database access traffic, but technically they are part of the o mainstream protocol and require no special ca or maintseance.

c to Undo Undo Soort in the Cache Manauer In Applicants' system, the system maintains a cache of all the objects that have been fetched fin' the database. Any time a change is made to the database, the new values are written into the cache.

An undo management facility is added to the cache manager. This service records every change in an undo queue. For typical changes to a property value, the undo queue simply records the previous value. For lifetime changes such as creation or destruction of objects, the undo queue records the equivalent operation that would revert that operation: delete for create, create for delete. In practice, the operations that would revert a change may be very complex: for example, if a class is deleted, all instance of that class, all subclasses, all instance of all subclasses, all methods and other facilities of those classes are also deleted. The undo manager stores the inverse operations required whenever possible. In some cases an undo operation may not be possible or practical, and in those cases the appropriate notation is stored in the undo queue.

The undo manager also provides undo and redo functions that the application can directly invoke, as well as support services usefdl to a GUI, such as listing the sequence of operations in the undo queue in human-readable form, Undo and Tansaction Managsment It should be noted that this undo capability cannot be effectively provided by the rollback capability of database transaction management systems. Database rollback is by far too Jan 2005 13:12 Smoorenburg Patent Trad 03 9712 0159

O

O

c expensive an operation to be connected to a trial-and-eor user operation available through a t mouse click. In addition, no database transaction system provides redo, and very few provide multiple undo.

Once a set of operations have been committed in a transaction, it may be problematic to 00 5 undo the individual steps of th transaction For example, if the userwants to transfer money Sfom one account to another, and this is done as one transaction consisting of one subtract and o one add operation, it would be problematic to undo the individual steps of that operation.

I' Therefore, typically a Commit operation conolidates the undo queue, collapsing all the o operations in the current transaction into one undo-able operation. In some cases, a specific counter-operation may be known for a specific operation for a particular method may be specified, in the database schema, the coresponding counter-mcthod- In others, the transaction may not be undoable at all. However, the details of bow a transaction is handled may be modified in any particular implementation ofthe system.

Non-Stack Undo Model The common approach to undo operations is a stack model: actions must be undone in the reverse order they were made, and redone in the order they were made. In principle, however, there is nothing to prevent individual actions to be undone in arbitrary sequence.

Random access undo operations are in principle possible, and in some cases quite useful Such random access undo is often rejected in the current state of the art, because in many cases it may not be possible. For example, if operation 6 in the queue creates an object, and operation 11 modifies a property of that object, then undoing operation 6 alone would not be possible because it would invalidate operation 11. But such a simplistic interpretation is not the only one possible. It is possible topermit random-access undo, if the inter-dependencies of the actions is maintained. In this example, undoing action 6 would automatically undo action 11, but if action 7 has no relationship with action 11 or with any other operation in the queue it can be undone by itself Such random-acces undo is quite useful in many ofthe complex applications that are increasingly common. In classical online transaction procsing, the sequence of operations is often quite simplistic and the benefit of random-access undo is limited, but in a more complex, knowledgeintensive application it is quite useful.

Jan 2005 13:13 Smoorenburg Patent Trad 03 9712 0159 p.36 0 0 ci e Scehema Tolerance ^t In a theoretical, ideal situation, all related objects would be stored in a single database, o managed under a single schema. Existing systems, however employ several distributed databases in these cases, a consistent schema is assumed, but in reality schema consistency may be hard to achieve. For technical, economic, practical and admisttive reasons, information 0 may be stored in databases under inconsistent schemas. For example, after a business is O acquired, much of its information may need to be kept in existing databases that don't match the n acquirer's schema. For example, a U.S. company that acquires a Venezuelan or Russian 0 Sorganization will have to deal with the diffrnt defitions of a name: given name, family name, mother's family name in Venezuela, given name, patronymic, family name in Russia.

Applicants' system accommodates such schema inconsistencies and permits the application from navigating seamlessly between the different databases. Consider this example: For eachm prI staff"'. .en If the initial object, the boss, is stored in database A under a particular schema, the employees that make up the staffmay be in another database, B, under a different schema.

Indeed, some employees may be stored in A and some in B. The system hides this distinction, permitting the application to retrieve the objects and access their properties and methods regardles of where they are stored: 019kbos r bW or each p If the distinction between the two databases is significant, the application can interrogate the object, find out its location, and adapt itself to the specific schema, for example retrieving the patronymic when one adsts.

Assocations Amonu Entites Jan 2005 13:13 Smoorcnburg Patent Trad 03 9712 0159 p.37 0 0 Applicants' system abstracts the concept of relationships among entities as an t association. Such an association is defined as part of the schema, and an application program can o the services of the invention to navigate the association. Because the definition is abstract, removed fom the physical implementation, a provider can use whatever technique it wants for S 5 maintaining such relationships, such as foreign keys in a relational database, object references in o an object database, methods that execute queries, as long as they are exposed in a manner o consistent with the semantics of an association.

Once the rlationship is presented as an asociation, the application can easily navigate o this relationship. The associated objects simply appear as related objects in the application programming language. The application code might look like this pseudo-code:

L

naxt enp a An association is an abstract concept that provides a bi-directional reference between two entities.

In most cases, the entities that ar associated ae objects, but the concept is not limited to objects. Associations can dist between any pair of entities that have a globally unique name, such as g. a file with a path, a web page with a URL, an object with a UUID. For example, association from a property of a class can be assigned to a validation method or visualization service for that property.

An association can also have additional properties of its own, such as the inception date.

A program, such as perhaps a database administration tool used by a human can define types of associations, and create individual associations among the named entities. The program can also navigate using these associations, retrieving either the associated entities or the associations themselves.

The associated entities m automatically fetched as they are needed: the association is automatically de-rferenced as the application navigates over the set of associated entities.

This automatic de-referencing is made efficient through automatic optimization and selftuning techniques, for example: Jan 2005 13:13 Smoorenburg Patent Trad 03 9712 0159 p.38 ci The system does not fetch all of th objects innnediately, on the getAssociatedItems lt statement. A lazy fetch is used, fetching objects from the database as they are needed The lazy fetch is not done for each object as it is needed. The lazy fetch manager fetches objects in batches.

00 c 5 The system is self-tuning, adapting the size of the batch to the observed performance o of the environment. Rapid turnaround causes batch size to be increased. Slow o trnaround causes batch size to be decreased. More complex tuning algorithms can of ^t course be inserted here.

0 o The performance tuming subsystem also permits the developer to provide hints, guiding the system in its tuning.

Association Registrv The various implementations that may bo used for associations, such as objects with references, tables with foreign keys, methods or queries, may be used directly for navigation, without any further specification or metadata. For example, consider a class called Containment used to implement an association. It holds two references to the container object and the contained item: C A A- i Such a class can be directly used for navigation by programing something ie this: Collection cont ens t.orMtAmsociatedlt r' "CIn almser,"co1ial. co n i The last statement tells the system to look for objects of the class Containment, using the property named container to find references to the curent object (referenced by cont), and then find all objects referenced by the property contained item.

The class Containment is not special in anyway, it is just a regular class. (This is similar to the way you do joins in SQL, using a regular table without any special characteriatics.) However, the present invention makes this concept more general by allowing the Containment clas to be registered in the association registry. This tells the infrastructure that this class is normally used as an association, but it does not limit the use of the class in any other way.

Jan 2005 13z14 Smoorenburg Patent Trad 03 9712 0159 p.39 Similar registration can be made for other types of associations, including but not limited l to: A pair of direct object-to-object rferences (for a 1-to-I association); Collections of object- 0 to-obect rfer ces (for 1-to-N and N-to-N associations) and Search methods that retrieve Logically associated objects.

00 C 5 The association registry offers numerous benefits, including for example: o Consumer applications do not need to know the implementation model used by a o specific provider, they simply reference the abstract concept of associations, and the Sinfrastructure translates that reference into the appropriate operations given the 0 o specific implementation of the association Separating the application from the implementation of the association makes it more resilient to changes in the provider's data model, and more able to accommodate replacement of a database provider with another one ofvery different type General tools that do not know anything about the application-level schema of the provider can find out which types of associations exist, using the association registry for introspection; this is particularly usful for graphical browsers When generating the Layer 2 model, the Layer 2 generator can create virtual properties based on the registered association types; in the example above, we can register the association through the Containment class under the name contents, and this permits the Layer 2 code generator to create a virtual property contents for the Container class, giving applications an easier way to navigate the associations References to External Obicts The associations between objects may reference information stored in external providers.

For example, the boss in the previous example could be stored database A, and the staff in database B.

The automatic navigation of associations, and the automatic de-referencing of the objects, completely hides this distinction from the application. The application can navigate between these objects, going fom one to the other, without cocemrn for where the objects are located.

Indeed, since the architecture of the system hides the nature of the providers from the application, it is possible for such external references to point to any type of data source; Jan 2005 13:15 Smoorenburg Patent Trad 03 9712 0159

I

0 0 including but not limited to data in a non-object database, a person in a directory, a file or, a page t on the Internet.

Associations Stored Extemallv In many cases, an application may use existing database providers whose schema cannot 0 change.

o In such instances, it may be desirable to maintain relationships between these objects in rC these providers. As a practical example, information may be stored in the databases of a o commercial financial management system, which references people or departments, and it may o be desirable to link those references to the corresponding entries in a directory. This linking would allow an application to be built that reacts to data in the database and makes decisions based on the organizational data in the directory, perhaps sending email.

However, if nether the database or the directory can be changed, it may not be possible to store the association and do a straightforward join between the databases.

Applicants' system permits the storage of such associations in other databases. Thus, two databases that don't frence each other, and cannot be made to reference each other, can nonetheless be linked.

In Applicants' system, the application navigates the associations in the standard way, the only thing it has to do is specify where the association is stored when it names the association to be used for the navigation. Even this minor inconvenience can be eliminated by registering the association in a registry. This way, the association may be mentioned by name in the standard way, and the system looks it up in the rgistry, finds where it is actually stored, goes to the external association store to get the link, and then follows the link to the other object.

External Loopback Associations This same technique can be used to store associations that reference objects within a single database, in those situations when the schema cannot be changed to accommodate direct storage of the associations.

Langeuae Bindlnes Applicants' system exposes its services through an Application Programming Interface (API) tat is available in common programming languages including Java, and any language that supports COM, including Visual Basic.

Jan 2005 13:15 Smoorenburg Patent Trad 03 9712 0159 p.41 0 c Laver I Lanymae Bindine Every system for accessing information providers or service providers uses a core data 0 model In SQL-based systems such as ODBC and JDBC, the data model is tables with atomic elements in the cells, extended with stored procedures. In ORBs, the data model is interfaces, 00 essentially a specification for a procedure cal. The choice of data model is an essential part of o every design, because it embodies some tradeoffs: S* If tm data model-is primitive, it limits the capabilities of the system SQL) l0 If the data model is rich, it demands a lot of capabilities of the providers it works with 0 and makes it cumbersome to comct primitive providers Java RMI) C 10 I If the data model is very specific, it cannot be adapted to accommodate different designs Microsoft WMI) If the data model is very general, it provides a low level of services and leaves all semantic intelligence to the application LISP) The present object model begins with something called an item, then proceeds to define collections of items, introduces the notion of item names, and, finally, named collections and item ownership. In this way, a compact and elegant foundation is built for a self-descriptive data fonmat that can be used to store, move and reference data. This data model is called Layer 0.

This format is built upon to define higher level concepts such as classes and objects, properties and methods (Layer and then on top of that build specific classes such as computers and employees (Layer 2).

The unique advantages of this design are: The basic data model (Layer 0) is simple enough that it can accommodate any existing system, without imposing any requirements on an object model.

Because the higherlevel model (Layer 1) is based on the simple Layer Omodel, it can provide a more modem object model and richer semantics for those providers that present such capabilities, without demanding it of simpler providers.

The implementation of the Layer 1 binding is based on dynamic processing, and is therefore tolerant of inconsistecies and change. If a provider changes its data model, for example if a database schema is modified, the Layer 0 model automatically accommodates the change, and the Layer I model does so as well. Applications that 195tU1.2 Jan 2005 13:18 Smoorenburg Patent Trad 03 9712 0159 p.42 0 0 c use the provider do not require change or ecompilation, and do not fail if they do not f have to. Specifically, added capabilities can be simply ignored if the application is o not interested, but if the application uses the services of Applicants' system to do introspection it can find out about the new capabilities and exploit them. Capabilities 00 5 that are removed have no impact on applications that do not use those capabilities.

SEven if an application attempts to use a capability that has been removed, it simply Sgets an exception message and can attempt to recover.

ID Layer 2 Proxy Bindin 0 Problem: Resilience vs. AsuMMa When constructing and maintaining a large application system, especially one that include previously existing elements or elements that come from other application systems, maintaining the consistency of th entire system is challenging. The trm Configuration Management (or simpler, Version Management) rafers to the activity of ensuring that the various elements of the system are compatible, so they can inleoperate, and consistent, so they work from the same assumptions. In existing systems, configuration management is commonly viewed as a build-time activity.

In Applicants' system, configuration management is a runtime activity: the elements of the system should be to communicate, to negotiate and agree on a common version; each element should be resilient and should be able to continue to operate even when a corresponden is inconsistent.

In existing systems, such resilience is sometimes achieved through dynamic interfaces, often called late binding. A late-bound system, such as the IDispatch-based interfaces of COM, is able to adapt to whatever interface an element exposes.

However, making a software system entirely late-bound is often unattractive, because it eliminates the possibility ofvalidating the consistency of the configuration at compile time, with early binding. Compile-time validation permits a level of assurance that is never achievable with late-bound systems since ran-time testing can never be shown to be exhaustive.

Applicants' system combines the resilience oflate binding and the assurance of early binding.

Problem Resilience vs. Convenience Jan 2005 13:16 Smoorenburg Patent Trad 03 9712 0159 p.

4 3 0 ci SDynamic interfaces allow a consumer to adapt itself the specification of a provider, by interrogating the provider through introspection and reflection services. Such services are Scommonly provided in component architectures such as COM, and in database access services such as ODBC and embedded SQL.

00 5 The benefits of thse system come at a high price. Programming to these interfaces is o quite cumbersome and err-prone. To lessen the burden on application developers, we would 0 0 prefer an interface that exposes the data model of the business application is preferable, A integrating it directly in the programming environment. Ths is particularly attractive with S modem object-oriented languages such as Java and Visual Basic, which directly support C 10 relatively rich semantic models.

The Solution In existing systems the benefits of these two approaches are never available at the same time. Early-bound systems such as COM do not provide the resilience or flexibility of late binding. Plexible systems such as dynamic SQL do not provide the assurance and convenience of early binding.

Applicants' system combines these benefits through a high-level language binding. The Layer 2 proxy binding exposes objects available in information providers in the form of native classes in a specified programming language. A developer defines the schema in the database, and then uses that schema in the application language: the proxy binding is data provider centric.

This type of binding provides several advantages over the Layer I binding, without sacrificing its core advantages.

The flexibility of the Layer 1 binding, and its ability to dynamically accommodate any provider through introspection entails a heavy burden on the developer. The Layer 2 binding presents the developer with the semantic constructs of the application domain, reducing the development and maintenance effort.

The dynamic adaptability of Layer 1, namely the capability to accommodate changes to provider schema, makes it difficult to validate consistency between consumers and providers at compile time. The Layer 2 binding generates classes in the target programming language or infratructur, and these classes can be used for compiletime validation.

Jan 2005 13:16 Smoorenburg Patent Trad 03 9712 0159 p.44

O

O

At th same time since the Layer 2 binding is implemented on top of Layer it is f ~capable of dynamic accommodation just like any Layer 1 application. Hence, an o existing application over which the developer has no control, because of remote deployment, administrative barriers or cost considcramtio can continue to operate 0 5 with the resilience of Layer 1. Applications that deserve compile-time validation can Sbe run through the standard Layer 2 validation process.

O

O

O

0 i Unified Proxy and Persistence Btndin Problem: Limits of Pemlsatnce Bindin=s Unifiad Bindings In Applicants' system, the Layer 2 persistence and proxy bindings are unified into an integrated whole, thus reducing the risk of inconsistncy; A developer can define the initial schema either in the application program or in the database tool A utility migrates the schema definition from one environment to the other. If the original definition was the application program, a database schema is constructed and installed in the database. If the initial definition was the database, source code is generated for use in the application program. The developer can then refin the definition in either location; for example, by adding indexing and clustering definitions, modifying the properties and their types and attributes, adding or modifying server client side methods. The utilities of Applicants' system keep the two environments synchronized, translating the definition from one language to another.

Aoects of the Invention This unification is achieved through these haracteristics of the invention: The syntax of the application program is identical for the two language bindings (the proxy and persistence models); The database schema is identical for the two binding models; Applications in both models use the same runtime support library; Application in both models have the same semantics; The code generator used in the proxy model generates code that is compatible with the persistence model preprocssor; The schema generator and code generator used in the persistence model generate schema and code that are compatible with the proxy system; and Jan 2005 13:17 Smoorenburg Patent t Trad 03 9712 0159 41 Both code preprocessors and code generators accept and preserve o user extensions.

Benefits 00 This unification offers at least two very Important advantages: S 5 First, unified bindings support iterative development. Accordingly, a Sdeveloper carl use the two models iteratively, moving back and forth between the V' two ways of working. Second, unified bindings provide pedagogical unity, limiting o an application to a single source code model making the system easier to learn.

C While the best mode for carrying out the invention has been described in detail, those familiar with the art to which the Invention relates will recognise various alternative designs and embodiments for practicing the invention. These altemative embodiments are within the scope of the present invention.

"Comprises/comprising" when used in this specification is taken to specify the presence of stated features, integers, steps or components but does not preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.

Claims (6)

1. A method for maintaining the integrity of data stored throughout a 00 distributed computer system, the mefthd Including: o Identifying an object to be persisted; o determining a database provider for storing the object;, and storing the object with the database provider.

2. A method as claimed in claim 1, wherein the step of determining includes the step of selecting the database provider based on an associated with a schema defined in a database maintained by the database provider.

3. A method as claimed in claim 1. wherein the step of determining Includes the step of selecting the database provider based on a database associated with the class to which the object belongdi.

4. A method as claimed In claim 1, wherein the step of determining Includes the step of selecting the database provider based on a dynamic association of a database to the class to which the object belongs. A M~ethod for maintaining the integrity of data stored throughout a distributed computer system, the method Including the steps of: storing a location ideniier associated with a method of an object, the location Identifier representing a location for processing the method; invoking the method; and processing the method at the location represented by the location identifier.

Jan 2005 13:17 Smoorenburg Patent Trad 03 9712 0159 p. 47 0 0 43

6. A method for maintaining the integrity of data stored throughout a o distributed computer system, substantially as herein disclosed with reference to figure 2 of the accompanying representations. 00 DATED this 5 day of January 2005 o COMPUTER ASSOCIATES THINK, INC 0 SMOORENBURG PATENT TRADE MARK ATTORNEYS cPO BOX 9 KANGAROO GROUND VIC 3097 AUSTRALIA