Complex procedures are the norm in today's interconnected computing world. A “complex” procedure is one whose successful completion requires the successful completion of a number of separate steps. If any one of these separate, constituent steps fails, then the procedure as a whole also fails.
Advances in communications allow various steps of a complex procedure to be performed on various computing devices. In some cases, the use of multiple computing devices is inherent in the procedure itself, as when a client requests a resource from a resource server. This common scenario becomes even more complex if the resource server asks an authentication server to verify the client's credentials before fulfilling the request. The client's request will fail if any one of the devices, the client, the resource server, or the authentication server, fails to perform its part of the transaction. The client's request can also fail because of a communications failure in the networks connecting these three devices.
In other cases, the complex procedure can be performed on a single computing device, but multiple devices are invoked to speed the procedure. For example, an intensive mathematical computation is broken into steps, and the steps are distributed to individual servers. As one hedge against possible failure, the same computational step can be distributed to a number of servers. In any case, a failure of one step causes the entire computation to fail or, in the case of redundant servers, can slow down the production of the final result.
The potential for trouble in a multi-step procedure increases when the procedure involves multiple databases. Here, a failure can not only prevent a client's database request from being fulfilled, but can also leave the databases in inconsistent states, i.e., “unsynchronized.” For a simplified example, consider a computing environment with two resource servers and a directory server that directs client resource requests to the appropriate resource server. Moving a resource from one resource server to another (in order to, for instance, balance the load of requests between the resource servers) involves the updating of both of the resource servers and of the directory server. A failure in the multi-step resource movement procedure could leave the directory server directing client requests to a resource server that no longer has, or does not yet have, the appropriate resource.
Techniques have been developed to mitigate failures in multi-step procedures that, like the situation given above, involve multiple databases. If all of the computing devices involved in the procedure use the same type of database, then a well known “two-phase commit” process can be invoked. The two-phase commit is designed to keep the databases synchronized at their pre-procedure state if an error occurs at any time during the procedure. In the first phase, each of the databases involved receives an update command. A transaction monitoring system then issues a “pre-commit” message to each database. If a database can successfully perform the update, then it temporarily stores the update and acknowledges the pre-commit command. If the transaction monitor receives acknowledgements from all of the databases involved, then it issues to them a “commit” message. Upon receiving the commit, each database makes the temporary change permanent. The procedure has been successfully performed, and the databases are now synchronized in their post-procedure state. If, on the other hand, the transaction monitor does not receive all of the expected acknowledgements, then the multi-step database update procedure has failed, and the temporary changes at each database are discarded. While the procedure has failed, the databases are synchronized in their pre-procedure state. Because of this synchronization, it is possible to either safely retry the multi-step database update procedure or to safely abandon the attempt.
Useful as it is, the two-phase commit only applies to a limited scope of procedures. It only works well if all of the servers involved use the same type of database. (It can be implemented across different database types, but at a significant increase in cost and complexity.) It also does not work where the multi-step procedure calls for changes to data structures other than databases.
Another useful technique for managing errors during a multi-step procedure is the “rollback.” Here, for each step of the multi-step procedure a method is developed for “rolling back,” or undoing, the results of that step. When a step fails, its partial results are rolled back. The results of previous, successfully performed steps can also be rolled back. This continues with all of the involved devices until they are all in their pre-procedure state. Then, just as in the case of a failed two-phase commit, it is safe to either retry the multi-step procedure or to safely abandon the attempt.
The rollback procedure, though in theory more widely applicable than the two-phase commit, has its own serious drawbacks. First, just like any other step in the multi-step procedure, each rollback step can itself fail. To handle this, a method of rolling back each rollback step is developed. This illustrates the second drawback: Adding rollback steps to a multi-step procedure complicates an already complicated scenario. This additional complication increases both the development and the processing costs of the multi-step procedure and may actually decrease the overall probability of the procedure's success.
In view of the foregoing, the present invention provides a rollback-free method for performing multi-step procedures in the presence of possible failure. As the procedure proceeds from its initial state through transitions to its final state, its constituent transitions are monitored for failure. If a failure is detected, then the procedure is “sidetracked” into a recovery coordination state. From the recovery coordination state, the procedure is logically taken back to its initial state and retried. Each of the transitions leading up to the failed transition can be either re-executed or simply skipped, relying upon the results achieved before the failure. In this manner, the procedure is shepherded, without rollbacks, through its transitions until it successfully reaches its final state.
In some embodiments, to ensure that the transitions in the multi-step procedure need not be rolled back when a failure is detected, the transitions are designed to be “idempotent,” that is, the transition can be performed any number of times with the same result.
In a particular embodiment, a multi-step procedure is developed for moving a resource from one resource server to another. The steps of the procedure are designed so that all throughout the procedure, both of the resource servers and a directory server are kept synchronized. This allows client requests to proceed without interruption even during the resource movement.
As the resource movement example illustrates, the methods of the present invention do not depend upon having a unified database structure, or indeed any database structure at all, on the servers involved in the multi-step procedure.
While the appended claims set forth the features of the present invention with particularity, the invention, together with its objects and advantages, may be best understood from the following detailed description taken in conjunction with the accompanying drawings of which:
a and 5b are state/transition diagrams of rollback-free procedures according to embodiments of the present invention;
a and 7b together form a flowchart of an exemplary procedure according to the present invention for moving a resource from one resource server to another without rollbacks; and
a and 8b together form a state diagram of the servers in the example of
Turning to the drawings, wherein like reference numerals refer to like elements, the present invention is illustrated as being implemented in a suitable computing environment. The following description is based on embodiments of the invention and should not be taken as limiting the invention with regard to alternative embodiments that are not explicitly described herein.
In the description that follows, the present invention is described with reference to acts and symbolic representations of operations that are performed by one or more computing devices, unless indicated otherwise. As such, it will be understood that such acts and operations, which are at times referred to as being computer-executed, include the manipulation by the processing unit of the computing device of electrical signals representing data in a structured form. This manipulation transforms the data or maintains them at locations in the memory system of the computing device, which reconfigures or otherwise alters the operation of the device in a manner well understood by those skilled in the art. The data structures where data are maintained are physical locations of the memory that have particular properties defined by the format of the data. However, while the invention is being described in the foregoing context, it is not meant to be limiting as those of skill in the art will appreciate that various of the acts and operations described hereinafter may also be implemented in hardware.
The present invention provides a rollback-free method for performing multi-step procedures in the presence of possible failure.
The computing environment of
Any of the devices of
The server 102 and the client 104 of
To clearly distinguish the present invention from the prior art,
It may happen, on the other hand, that one or both of the databases is unsuccessful in trying to move to its requested state. For example, instead of successfully performing the transition 310, the database may revert in transition 316 to its pre-procedure, initial state 306.
In state 318, a monitoring agent queries all of the databases involved in the procedure to find out whether they have successfully made their transitions. If all of them have, then each database is told to mark its change as permanent in state 320. The update procedure has been successfully performed, and all of the databases are in their requested, post-procedure states 308. If, however, at least one database could not comply with the request, then each database is told to “back out” of the request, that is, to discard the change it marked as temporary. Here, the update procedure has failed, but, because of the back-out enabled by the two-phase commit technique, the databases are still synchronized, albeit in their pre-procedure initial states 302. Because the databases are synchronized, the procedure can be retried or can be abandoned with safety.
This two-phase commit technique is readily applicable to the environment of
Unfortunately, the rollback technique of
Contrast the prior art techniques of
b shows a slight variation on the rollback-free technique of
In order for the rollback-free techniques of the present invention to work properly, the procedure should be tolerant of having its transitions performed multiple times without causing a loss of data integrity. This aspect is discussed below, especially in relation to step 610 of
An embodiment of the rollback-free technique illustrated in
Step 602 is included for the sake of completeness. It corresponds to setting the initial state 400 of
The procedure is actually performed in step 604 which corresponds to the states 404, 408, 412, and 416 and their failure-free transitions of
If a failure is detected during the performance of the procedure, however, the procedure stops, and control passes to step 608 (which corresponds to the recovery coordination state 502 of
Avoiding rollbacks, the rollback-free failure recovery technique begins in step 610 to retry the procedure from the beginning. If some of the transitions specified by the procedure were successfully completed before the failure occurred, then those transitions can be skipped. Alternatively, those successful transitions can be re-executed. This is an important point: The procedure should be designed so that, as much as possible, its transitions can be re-executed without causing a loss of data integrity. In developing a procedure that has this property, it helps to consider the database concept of “idempotence.” A database update or a transition in a procedure is said to be idempotent if it can be invoked multiple times and produce the same result every time. For example, consider the case of a variable X with an initial value of 1. Two functions are defined that alter the value of X: INCREMENT(X) gives X a final value 1 more than its previous value, while ASSIGN(2, X) gives X the final value of 2. Clearly, INCREMENT(X) produces a new value for X every time it is run, while ASSIGN(2, X) always leaves X with the final value of 2. ASSIGN(2, X) is idempotent, and INCREMENT(X) is not. Because the techniques of the present invention do not rollback previously created intermediate results, the procedures should be developed with idempotence in mind. Strict idempotence is not always necessary, but the procedure should be able to be performed even in the face of inconsistent data caused by previously created intermediate results. As another simple example, the procedure should be able to intelligently ignore an error message saying that the requested transition has already been performed. An example of such a procedure is discussed below in relation to
In any case, the procedure begins again at its initial state (400 of
a, 7b, 8a, and 8b illustrate a procedure developed to work with the rollback-free failure recovery techniques of the present invention.
Step 602 of
In this particular embodiment, each of the resource servers 108 and 110 has a Boolean moveaway flag. In step 704, the first resource server 108′ moveaway flag is set to FALSE to indicate that the resource-moving procedure has not yet begun. Note that these steps 700, 702, and 704 are not really part of the resource-moving procedure: They are shown to illustrate the state of the actors before the procedure actually begins.
In general, the procedure is actually performed in step 604 of
In step 710 of
In step 712, the moveaway flag of the second resource server 110 is set to FALSE indicating that the just arrived resource is expected to stay a while. In some embodiments, the resource takes up server memory. In these embodiments, the first resource server 108 can delete its copy of the resource in step 714 in order to free up its memory. Once the procedure completes successfully, only the second resource server 110 will ever receive a request for the resource.
Finally, in step 716, the directory server 102 sets its current resource server flag to indicate that the resource is now on the second resource server 110 and sets its target resource server flag so that it does not indicate any resource server. The latter flag is used to indicate that the resource is not currently being moved. In the illustrated embodiment, these two flags are set in one indivisible (or “atomic”) operation so that their values remain consistent at all times. Fortunately, some operating systems provide the ability to change exactly two variables in one operation. The atomic operation of step 716 either succeeds with respect to both variables or fails with respect to both so that their values remain consistent. Some operating systems do not provide the ability to change two variables in one atomic operation. If that is the case, then the current resource server flag and the target resource server flag are simply set in consecutive steps.
If a failure is detected at any time during the resource-moving procedure of
In view of the many possible embodiments to which the principles of the present invention may be applied, it should be recognized that the embodiments described herein with respect to the drawing figures are meant to be illustrative only and should not be taken as limiting the scope of the invention. For example, those of skill in the art will recognize that the illustrated embodiments can be extended to cover other, non-illustrated multi-step procedures without departing from the spirit of the invention. Although the invention is described in terms of software modules or components, those skilled in the art will recognize that such may be equivalently replaced by hardware components. Therefore, the invention as described herein contemplates all such embodiments as may come within the scope of the following claims and equivalents thereof.
This application is a continuation of U.S. patent application Ser. No. 10/659,453 entitled “METHOD AND SYSTEM FOR ROLLBACK-FREE FAILURE RECOVERY OF MULTI-STEP PROCEDURES,” filed on Sep. 10, 2003, which application is incorporated herein by reference in its entirety.
Number | Name | Date | Kind |
---|---|---|---|
4663706 | Allen et al. | May 1987 | A |
4807224 | Naron et al. | Feb 1989 | A |
4878167 | Kapulka et al. | Oct 1989 | A |
5151987 | Abraham et al. | Sep 1992 | A |
5440545 | Buchholz et al. | Aug 1995 | A |
5774479 | Lee et al. | Jun 1998 | A |
5842224 | Fenner | Nov 1998 | A |
5884327 | Cotner et al. | Mar 1999 | A |
5889772 | Fischer et al. | Mar 1999 | A |
6199055 | Kara et al. | Mar 2001 | B1 |
6247012 | Kitamura et al. | Jun 2001 | B1 |
6321374 | Choy | Nov 2001 | B1 |
6377959 | Carlson | Apr 2002 | B1 |
6463456 | Kan et al. | Oct 2002 | B1 |
6502088 | Gajda et al. | Dec 2002 | B1 |
6535997 | Janson et al. | Mar 2003 | B1 |
6728747 | Jenkins et al. | Apr 2004 | B1 |
6735620 | Blackmore et al. | May 2004 | B1 |
6965571 | Webber | Nov 2005 | B2 |
7178060 | Bamford | Feb 2007 | B2 |
20040083276 | Shiraga | Apr 2004 | A1 |
20050055606 | Kutan et al. | Mar 2005 | A1 |
Number | Date | Country | |
---|---|---|---|
20070101192 A1 | May 2007 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 10659453 | Sep 2003 | US |
Child | 11614466 | US |