JP2013033441A - Database management method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize a database management technique of a multiple master node type which can efficiently utilize resources.SOLUTION: A management method performs the steps of: updating a database of a higher master node if a transaction minimal value is equal to or more than a cluster minimal value and aborting lower database update information if the transaction minimal value is smaller than the cluster minimal value, in the higher master node which receives the lower database update information; generating an update record of the higher master node as a transaction log; distributing the transaction log to the lower master node; updating the own database on the basis of the received transaction log; and discarding a transaction with the transaction minimal value which is smaller than the cluster minimal value.

Description

  The present invention relates to a database management method, and more particularly to a database management method in which a plurality of master nodes are hierarchically connected by a network.

  Even for a write-once database such as PostgreSQL, in order to make effective use of resources, it is desirable to completely delete the tuple after a certain period of time has elapsed after it has been deleted.

  However, since a transaction is managed from the start to the end with a single snapshot, even if a specific tuple is deleted by a certain transaction, this does not immediately refer to the tuple from all transactions. Not exclusively. Therefore, in principle, the tuple is not completely erased unless all the series of transactions that refer to the tuple are committed. Even if the tuple is deleted, it must be visible on other snapshots currently being executed.

  Therefore, PostgreSQL uses a transaction ID assigned to each transaction and adopts a mechanism to determine how far to completely erase while referring to the transaction ID that has already been committed as seen from the snapshot of each transaction. Was. That is, the minimum value of the transaction IDs that are considered to be being executed in the snapshot of each transaction is calculated, and the deletion is performed in accordance with the ID of the minimum value.

JP 2006-293910 A

  Although the operations described above are operations on a database having a single node configuration, when the present invention is applied to a multi-master node, the present inventors have found that the following problems that cannot be expected in a single node system occur. It was issued.

  That is, the multi-master node system is composed of a huge number of nodes constructed hierarchically and has a huge number of transactions that are simultaneously executed in parallel on the nodes. For this reason, when there is a variation in processing capability between nodes, the processing delay of one node affects the overall processing capability.

  For this reason, in the case of a database having a multi-master node configuration, the value of the minimum value ID that can be completely erased indefinitely does not increase, so that the effectiveness of the complete erase process of the entire node cannot be ensured.

For this reason, a method of forcibly setting the minimum value ID to a value of a certain size and performing the complete erasure process is conceivable.
By the way, in this kind of multi-master node format database, the present inventor has performed a shadow copy of the database of the lower master node and a heap tuple map developed in its own memory in order to perform the update efficiently and without contradiction. Has been proposed as a writing set to unify database update processing for the entire node (Japanese Patent Application No. 2010-239713).

  By the way, while the write set is generated and transmitted to the upper node in this lower node, a complete erasure command for the tuple that is the update target of the write set is transmitted from the upper node to the lower node. In this case, the upper node that subsequently received the write set may update the tuple that should have been completely erased based on the contents of the write set.

  Therefore, even in a multi-master node format database, it is necessary to perform a transaction discarding process without destroying the consistency of the entire database.

  In order to solve the above-described problems, the present invention employs the following means.

  A first aspect of the present invention is a write-once database management method that hierarchically has upper and lower master nodes capable of record update, and in a session of any lower master node, Sending the shadow copy of the database of the lower master node, the heap tuple map expanded on its own memory, and the transaction minimum value of the snapshot referenced by the executing transaction as a write set; In the upper master node, the transaction minimum value in the write set received from the lower master node is compared with the cluster minimum value held by the upper master node, and if the transaction minimum value is equal to or greater than the cluster minimum value, Heap data in the write set The map is compared with its own database to verify whether there is an update in the database of the tuple registered as a target. When the update is made, the write set is aborted, and when the update is not made, the shadow copy Updating its own database using the command, generating the update record as a transaction log, aborting the write set if the transaction minimum value is smaller than the cluster minimum value, and the transaction Distributing the log to a lower master node including the transmission source lower master node, updating the own database based on the transaction log received in the lower master node, and the upper master node; A method for managing a write-once database having hierarchical master nodes, comprising: notifying a lower master node of a cluster minimum value held by a node and discarding a transaction having a transaction minimum value smaller than the cluster minimum value. It is.

  Further, according to the second aspect of the present invention, the lower master node compares the transaction minimum values between a plurality of snapshots in which transactions being executed are recorded in time series, and sets the smallest value as the node minimum value. The node minimum value is notified to the upper master node as a tree minimum value, and the upper master node is selectively clustered from its own node minimum value and the tree minimum value notified from the one or more lower master nodes. The write-once database management method according to the first embodiment, wherein the minimum value is determined and held.

The third form of the present invention has a middle master node in a hierarchy between the upper master node and the lower master node, and the lower master node records the transactions being executed in time series. The transaction minimum values between the plurality of snapshots are compared, the smallest value is set as the node minimum value, and the node minimum value is notified to the intermediate master node as the tree minimum value. The node minimum value is compared with the tree minimum value notified from the one or more lower master nodes, and the lowest value is notified to the upper master node as the tree minimum value of the intermediate master node. Is a cluster minimum value selectively from its own node minimum value and the tree minimum value notified from the one or more intermediate master nodes. It is a management method of a write-once database of the first aspect, wherein the determining and holding.

  The fourth aspect of the present invention is the write-once database management method according to the third aspect, wherein the intermediate master node further has a tree structure of two or more hierarchies.

  Further, in the fifth aspect of the present invention, the notification to the lower master node of the cluster minimum value held by the upper master node is performed asynchronously with the notification to the lower master node of the transaction log. This is a type database management method.

  According to this method, the transaction ID can be efficiently discarded by notifying the lower master node of the cluster minimum value asynchronously with the replication management process.

  According to a sixth aspect of the present invention, the write-down type database according to the first aspect, wherein the notification of the cluster minimum value held by the upper master node to the lower master node is performed by including the cluster minimum value in the transaction log. It is a management method.

  According to this method, since the notification order can be forced by including the cluster minimum value in the replication management process, it is possible to discard the transaction ID without any contradiction.

  According to a seventh aspect of the present invention, the write set is generated at a middle master node instead of a lower master node, and at least a transaction minimum value of a snapshot referenced by a transaction being executed at the middle master node. Is a write-once database management method according to the third or fourth embodiment.

  An eighth embodiment of the present invention is a method for managing a write-once database in an upper master node having nodes hierarchically, and in any lower master node session, a shadow copy of the lower master node database, Transmitting the heap tuple map expanded in the memory of the master node and the transaction minimum value of the snapshot referenced by the executing transaction to the upper master node as a writing set, and the writing received in the above If the transaction minimum value in the set is compared with the cluster minimum value held by the upper master node and the transaction minimum value is equal to or greater than the cluster minimum value, the heap tuple map in the write set and its own database As a target Verifying whether there is an update in the database of recorded tuples, aborting the write set if an update has been made, updating the database using the shadow copy if no update has been made, and When the transaction minimum value is smaller than the cluster minimum value, the step of aborting the write set and the cluster minimum value held by the upper master node are notified to the lower master node, and the transaction smaller than the minimum value is transmitted. And a step of discarding a transaction having a minimum value, and a method for managing a write-once database in an upper master node having hierarchically lower master nodes.

A ninth aspect of the present invention is a database management method for a lower master node in a write-once database that hierarchically has upper and lower master nodes capable of record update, and updates to the database at any lower master node When an instruction occurs, the database processing unit of the lower master node sends a shadow copy of the database expanded on its own memory to the upper master node, a heap tuple map, and a snap referenced by the transaction being executed. When the transaction minimum value of a shot is generated as a writing set, the transaction master value and the upper master node in the writing set received from the lower master node are held by the upper master node that receives the shot transaction minimum value. Compared to the cluster minimum value, the transaction minimum When the database is equal to or larger than the cluster minimum value, the shadow copy is used to update the database of the upper master node, and this update record is generated as a transaction log. A step of transmitting the writing set to a higher-order master node for distribution to a node, and a transaction log processing unit of the lower-order master node, when receiving the transaction log, based on the transaction log And a lower master node receiving a cluster minimum value held by the upper master node and discarding a transaction having a transaction minimum value smaller than the minimum value. It is a management method of soil database.

  In the tenth aspect of the present invention, the lower master node compares the transaction minimum values between a plurality of snapshots in which transactions being executed in time series are recorded, and sets the smallest value as the node minimum value. The value is notified to the upper master node as a tree minimum value, and the upper master node selectively selects a cluster minimum value from its own node minimum value and the tree minimum value notified from the one or more lower master nodes. The write-once database management method according to the eighth aspect or the ninth aspect.

  In an eleventh aspect of the present invention, a middle master node is provided in a hierarchy between the upper master node and the lower master node, and the lower master node is a plurality of time-sequential transactions recorded therein. The transaction minimum values between the snapshots of the snapshots are compared, the smallest value is set as the node minimum value, and this node minimum value is notified to the intermediate master node as the tree minimum value. The value is compared with the tree minimum value notified from the one or more lower master nodes, and the smallest value is notified to the upper master node as the tree minimum value of the intermediate master node. The cluster minimum value is selectively determined from its own node minimum value and the tree minimum value notified from the one or more intermediate master nodes. - a eighth form or how to manage the write-once database ninth embodiment according to retain.

  A twelfth aspect of the present invention is the write-once database management method according to the eleventh aspect, wherein the intermediate master node further has a tree structure of two or more layers.

  In a thirteenth aspect of the present invention, the notification to the lower master node of the cluster minimum value held by the upper master node is performed asynchronously with the notification to the lower master node of the transaction log. This is a database management method.

  In a fourteenth aspect of the present invention, the write-down database management according to the eighth aspect is performed, wherein the notification of the cluster minimum value held by the upper master node to the lower master node is performed by including the cluster minimum value in the transaction log. Is the method.

  In a fifteenth aspect of the present invention, the write set is generated at a middle master node instead of a lower master node, and at least a transaction minimum value of a snapshot referenced by a transaction being executed at the middle master node is written. The write-once database management method according to the eleventh or twelfth embodiment.

A sixteenth aspect of the present invention is a write-once database management method that hierarchically has an upper level that can be updated and one or more lower level master nodes, and in each lower level master node, a transaction being executed is Comparing the transaction minimum values of the snapshots to be referenced with each other and transmitting a node minimum value which is the smallest value to the upper master node; and the upper master node receiving the node minimum value from each of the lower master nodes, Reading the node minimum value of the node and comparing it with the node minimum value received in the above, determining and updating the cluster minimum value held by the upper master node within the range of the respective node minimum values, and the upper master The cluster minimum value updated in the node is notified to each subordinate master node, and the transaction is smaller than the minimum value. A step of discarding a transaction with Deployment minimum, a management method of a write-once database having a master node in a hierarchical consisting.

  A seventeenth aspect of the present invention is a method for managing a write-once database that has hierarchically the master nodes according to the sixteenth aspect, which are two layers of the upper master node and the lower master node.

  According to an eighteenth aspect of the present invention, there are at least three or more hierarchies including the upper master node, a middle master node of one or more hierarchies, and a lower master node, and the intermediate master node also has its own node minimum value. The upper master node executes the step of transmitting to the upper master node, and the upper master node receives the node minimum value from each of the lower master node and the middle master node, reads out its own node minimum value, and received the node A write-once database that hierarchically has master nodes according to the sixteenth mode for executing a step of determining / updating a cluster minimum value held by a higher-order master node within a range of each node minimum value in comparison with the minimum value It is a management method.

  According to a nineteenth aspect of the present invention, there is provided a write-once database management method having hierarchically upper and lower master nodes capable of updating records, and in any one of the lower master nodes, Compare the shadow copy of the database of the lower master node, the heap tuple map expanded in its own memory, the transaction minimum value of the snapshot referenced by the transaction being executed, and the transaction minimum value in the node. A node minimum value that is the smallest value is transmitted as a writing set, and the upper master node compares the node minimum value read from each writing set with the node minimum value held by itself. Therefore, the upper master node maintains a value within the range of the minimum value of each node. Determining and updating the cluster minimum value to be performed, and comparing the transaction minimum value in the write set received from the lower master node with the cluster minimum value held by the upper master node in the upper master node, If the value is equal to or greater than the cluster minimum value, the heap tuple map in the write set is compared with its own database to verify whether there is an update in the database of the tuple registered as the target. When the update is not performed, the write set is aborted. When the update is not performed, the shadow copy is used to update its own database, and this update record is generated as a transaction log. If it is smaller than the minimum value, the step of aborting the write set, the step of distributing the transaction log to a lower level master node including the lower level master node of the transmission source, and receiving at the lower level master node Updating its own database based on the transaction log, and notifying the lower master node of the cluster minimum value held by the upper master node, and discarding a transaction having a transaction minimum value smaller than the minimum value. Is a method for managing a write-once database having a hierarchical master node.

In a twentieth aspect of the present invention, there is an intermediate master node having one or more hierarchies between the upper master node and the lower master node. In any of the intermediate master node sessions, The shadow copy of the database on the intermediate master node,
Compares the heap tuple map expanded in its own memory, the transaction minimum value of the snapshot referenced by the running transaction, and the transaction minimum value in the node, and writes the node minimum value which is the smallest value A write-once database management method that hierarchically includes master nodes according to the nineteenth aspect, further including a step of transmitting as a set.

  The twenty-first aspect of the present invention is a write-once database management method that hierarchically has upper and lower master nodes that can be updated, and refers to a transaction being executed in a session of any lower master node. Transmitting the lower database update information including the transaction minimum value of the snapshot to the upper master node; and in the upper master node receiving the lower database update information, the transaction minimum value and the cluster minimum value held by the upper master node; If the transaction minimum value is equal to or greater than the cluster minimum value, the step of updating the database of the upper master node based on the lower database update information, and the transaction minimum value is greater than the cluster minimum value If small A step of aborting the subordinate database update information, a step of generating an update record of the upper master node as a transaction log, and a step of distributing the transaction log to a lower master node including the lower master node of the transmission source, The lower master node updates its own database based on the transaction log received above, and notifies the lower master node of the cluster minimum value held by the upper master node, and a transaction smaller than the minimum value. And a step of discarding a transaction having a minimum value, and a method for managing a write-once database in an upper master node having hierarchically lower master nodes.

  In the twenty-second aspect of the present invention, the lower database update information includes a shadow copy of a database of a lower master node, a heap tuple map developed in its own memory, and a minimum transaction of a snapshot referenced by an executing transaction. When the upper master node receives the write set, the transaction minimum value is equal to the cluster minimum value by comparing the transaction minimum value with the cluster minimum value held by the upper master node. If it is larger, the heap tuple map in the writing set is compared with its own database to check whether there is an update in the database of the tuple registered as a target. Abort the set and have not been updated 21. The write-once type according to the twenty-first aspect, which includes the steps of updating the own database using the shadow copy, and aborting the write set when the transaction minimum value is smaller than the cluster minimum value. This is a database management method.

  In the twenty-third mode of the present invention, in addition to the above, there is a middle master node of one or more layers between the upper and lower master nodes. In any of the middle master node sessions, the upper master node On the other hand, in the upper master node that has received the intermediate database update information, the step of transmitting the intermediate database update information including the minimum transaction value to be updated of the database of the intermediate master node to the upper master node; If the transaction minimum value is equal to or greater than the cluster minimum value by comparing the transaction minimum value with the cluster minimum value held by the upper master node, the database of the upper master node is determined based on the intermediate database update information. And updating the device according to the twenty-first or twenty-second embodiment. It is a management method of the type database.

  According to the present invention, even in a multi-master node type database, complete erasure of the entire tree can be realized with a larger minimum value ID level without destroying the consistency of the entire database.

Explanatory drawing which shows the database structure of the multi-node format of the present invention Block diagram showing the configuration of the database controller Diagram showing the contents of a writing set Explanatory drawing showing processing of write set in the upper master node Explanatory diagram showing the structure of transaction log data Diagram showing the relationship between multi-node database tree structure, transaction minimum value, tree minimum value and node minimum value Explanatory diagram of update of cluster minimum value and transaction discard processing using that cluster minimum value Figure showing snapshot contents The figure which shows the modification of writing set

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows the structure of the hierarchical master node of this embodiment. As shown in the figure, a node configuration having a middle-order master node (MS201, MS202... MS20n) and a lower-order master node (MS301, MS302... MS30n) hierarchically under the upper master node (MS101). It has become. Each node (information processing apparatus) has a database. The upper master node (MS101) has a slave, but other lower master nodes may have slaves. In the case of such a master / slave configuration, the update management technique described in PCT / JP2010 / 054311 by the present applicant (unpublished prior application by the present applicant) may be applied to update the database between the two. it can.

FIG. 2 is a functional block diagram of the lower master node (MS201), but the upper master node (MS101) has the same function.
As shown in the figure, when a database update command is input from the client (CL), the database processing unit (11b) performs lower-level database update information on the back-end memory (BEM) constructed on the memory (MM). Is generated. As shown in FIG. 4, this writing set includes a heap tuple map (HTM), a shadow copy (SC), and a transaction minimum value (Xmin). Here, it is assumed that an update command for deleting (DELETE) row number 4 from the master database (101a) and rewriting (UPDATE) row number 5 to a new value (sc1) is input.

At this time, the database processing unit 11b does not write directly to the master database (11a) while referring to the master database (11a), but stores the write set generated in the back-end memory (BEM) as a communication module ( 11d) to the higher master.
Such processing is the same in the upper master node (MS101) and the lower master nodes (MS201, MS202... MS20n, MS301, MS302... MS30n).

  FIG. 8 shows the relationship between a transaction and its snapshot in each node (upper node, middle node or lower node). A plurality of transactions are executed in parallel at each node, and a plurality of snapshots indicating the execution state of the transactions in time series at a plurality of respective time points are held.

  In this figure, transaction ID = 2, 4, 5, 6 and 9 are being executed in this node. Transaction IDs = 1, 3, 7, 8, and 10 are committed. Transaction ID = 11 is in an unexecuted state. In the snapshot at this time, the maximum value of the transaction being executed or committed is set as the transaction maximum value (here, Xmax = 10), and the minimum value of the transaction being executed is set as the transaction minimum value (here, Xmin = 2). It is recorded as.

Such a snapshot is recorded in the aforementioned back-end memory (BEM).
Among the plurality of transaction minimum values between these snapshots, the smallest value is the node minimum value. In the figure, since the transaction minimum value of the snapshot 701 is 2 (Xmin = 2) and the transaction minimum value of the snapshot 702 is 4 (Xmin = 4), the node minimum value is 2 (Nmin = 2). Become. When there is no lower node in the lower master node, the node minimum value (Nmin = 2) becomes the tree minimum value (Tmin = 2).

  The node minimum value (Nmin) generated by the lower-level master node (for example, MS 301), that is, the tree minimum value (Tmin) is notified to the middle level node (for example, MS 201) as needed asynchronously with the writing set described later. The middle master node (MS201) compares the tree minimum value (Tmin) received from each lower master node (MS301-MS30n) with the node minimum value (Nmin) in its own snapshot to obtain a smaller value. Is updated as its own tree minimum value (Tmin).

Next, the minimum tree value (Tmin) of each middle master node (MS201-MS20n) is notified to the upper master node (MS101).
When the upper master node (MS101) receives the tree minimum value (Tmin) received from each middle master node (MS201-MS20n), it compares the node minimum value (Nmin) in its own snapshot, Any minimum value or a value between the minimum values is adopted to update its own tree minimum value (Tmin). Then, the tree minimum value (Tmin) is compared with the cluster minimum value (Cmin) at that time, and a value between them is determined and a new cluster minimum value (Cmin) is determined and updated.

  Next, referring to FIG. 6, the node minimum value (Nmin) of the lower master nodes E and F, that is, the tree minimum value (Tmin), the tree minimum value (Tmin) of the middle master nodes B, C, D, and E, and the upper master A state where the tree minimum value (Tmin) and the cluster minimum value (Cmin) of the node A are updated will be described.

  Here, as a result of comparing the transaction minimum values (Xmin) of the lower master nodes F, when the node minimum value (Nmin) is 3 and the minimum value (Nmin) of the lower master node G is 4, these lower master nodes F , G have no further lower layer, and these node minimum values (Nmin) become the respective tree minimum values (Tmin).

These minimum tree values (Tmin = 3, Tmin = 4) are notified to the middle master node D, respectively.
The intermediate master node D has 4 as the node minimum value, and these are compared and the lowest value (3 in this case) is set as the tree minimum value of the intermediate node D (Tmin = 3).

On the other hand, in the intermediate master node E, the node minimum value (Nmin) is 5, and since this node has no lower-level master node, the node minimum value (Nmin = 5) is the tree minimum value (Tmin = 5). ) Is set.

  The middle master node C receives the tree minimum value (Tmin = 3) from the middle master node D and the tree minimum value (Tmin = 5) from the middle master node E. On the other hand, the intermediate master node C has 4 as its own node minimum value (Nmin), and compares them to update the smallest value of 3 as its own tree minimum value (Tmin = 3). The minimum tree value is notified to the upper master node A.

  On the other hand, in the figure, there is a middle master node B in parallel with the middle master node C, but since this middle master node B has no master node in the lower hierarchy, its own node minimum value (Nmin = 6) is directly notified to the upper master node A as the tree minimum value (Tmin = 6).

  The upper master node A receives the tree minimum value (Tmin = 6) from the middle master node B and the tree minimum value (Tmin = 3) from the middle master node C, and receives its own node minimum value (Tmin = 3). Compared with Nmin = 7), the smallest value among them is updated as the tree minimum value (Tmin = 3) in the upper master node A.

The master node A sets a numerical value between its own node minimum value (Nmin = 7) and the updated tree minimum value (Tmin = 3) as the cluster minimum value (Cmin).
As a criterion for determining the cluster minimum value (Cmin) at this time, if the cluster minimum value is set to a smaller value (for example, Cmin = 3), the transaction minimum value on the snapshots of all nodes becomes 3 or more. The probability that a write set generated in the lower / middle master node will be aborted when it reaches the upper master node is reduced, and the write set that is wasted is reduced, and the load on each node is reduced. On the other hand, since the standard for complete erasure becomes a small value, it is difficult to achieve efficient operation by complete erasure of the entire node.

  On the other hand, if the cluster minimum value (Cmin) is set to a larger value (for example, Cmin = 7), the efficiency of complete erasure is improved, but the write set generated at the middle / lower master node is the higher master. The probability of being aborted when arriving at a node increases, and the load of the write set at the middle / lower master node increases.

  Such a cluster minimum value (Cmin) setting method is such that a prescribed difference value between a node minimum value (Nmin) and a tree minimum value (Tmin) is determined in advance, and if this prescribed difference value is exceeded, it is compulsory. The reference value may be controlled to be within the specified difference.

  When the cluster minimum value (Cmin) is set in this way, tuples are completely erased based on the cluster minimum value (Cmin). That is, in the upper master node, tuples deleted by a transaction having a transaction ID smaller than the cluster minimum value (Cmin) can be completely deleted. The tuples completely erased by the upper master node can be safely and completely erased by the middle and lower master nodes by discarding the transaction and aborting the write set described below.

In the above description, the middle master node BE or the lower master nodes F and G are notified to the upper master node A while comparing the node minimum values (Nmin) with each other and updating the tree minimum value. Although described, the present invention is not limited to this, and each middle master node BE or lower master nodes F and G may directly notify the upper master node A of its own node minimum value (Nmin). Good. The upper master node A compares the node minimum value (Nmin) notified in this way with the node minimum value (Nmin) held by itself, determines a new value within these ranges, and determines it as the cluster minimum value. It can be updated as a value (Cmin).

In this way, each middle / lower master node BG directly notifies the upper master node A of the node minimum value (Nmin), thereby updating the tree minimum value (Tmin) in each hierarchy. A notification system can be simplified without the need for more complicated processing than the method of notifying A.
Further, when the write set is generated in each of the intermediate master node BE or the lower master nodes F and G, the node minimum value together with the heap tuple map (HTM), shadow copy (SC), and transaction minimum value (Xmin). (Nmin) may be registered (see FIG. 9). By registering the node minimum value (Nmin) during the writing set and notifying the higher master node A in this way, another notification system other than the writing set is not required for the higher master node A, and the system is simplified. Can be

  In this case, the upper master node A reads the node minimum value (Nmin) from each writing set directly transmitted from each middle master node BE and the lower master nodes F and G, and holds the node Compared to the minimum value (Nmin), a new value may be determined within these ranges and updated as the cluster minimum value (Cmin).

  In the upper master node A, after updating the cluster minimum value (Cmin) as described above, the same processing as described above is performed. That is, in the upper master node A, the transaction minimum value (Xmin) in the write set received from the intermediate master node BE or the lower master nodes F and G and the cluster minimum value (Cmin) held by the upper master node A ) And the transaction minimum value (Xmin) is equal to or larger than the cluster minimum value (Cmin), the heap tuple map (HTM) in the write set is compared with its own database, and the target Is updated in the database of tuples registered as, and when the update is made, the write set is aborted, and when the update is not made, the own database is updated using the shadow copy (SC). Along with this transaction record, Produced as a log. On the other hand, when the transaction minimum value (Xmin) is smaller than the cluster minimum value (Cmin), the write set is aborted.

  The minimum cluster value (for example, Cmin = 6) set and updated by the upper master node A is notified to the middle master nodes B, C, D, E and the lower master nodes F, G as shown in FIG. . Each node that has received this cluster minimum value (Cmin = 6) discards all transactions that refer to a snapshot having a transaction minimum value smaller than this value (Cmin = 6). As a result, the minimum transaction value (Xmin), the minimum node value (Nmin), and the minimum tree value (Tmin) based on them are always greater than 6. Therefore, after the notification of the cluster minimum value (Cmin), there is no transaction that refers to the completely deleted tuple.

  6 and 7, the description has been given with the multi-master node having a four-level tree structure including the upper master node A, the two-level middle master nodes B, C, D, E, and the lower-level master nodes E, F. However, the upper master node and the lower master node may have a two-layer structure, or the middle master node may have a tree structure of three or more layers, that is, a total of five or more layers.

Next, generation of a writing set and update using the writing set will be described with reference to FIG.
This figure shows the relationship between the master database (11a) in the lower master node (M301 as an example) and the writing set. The master database (11a) is composed of line numbers, instruction contents, and pointers, and is a write-once database in which line numbers are added each time a new instruction is issued from the client terminal (CL). In the case of the figure,
As described above, the case where line number 4 is deleted (DELETE) and line number 5 is rewritten to new instruction contents (UPDATE to sc1) is shown.

  When such an update command is issued to the master database by a command from the client terminal (CL) in the lower master node (MS301), as described above, the heap tuple map (HTM, A writing set comprising a heap file), a shadow copy (SC), and a transaction minimum value (Xmin) is generated.

  In the heap tuple map (HTM), the original line number (ctid) and the line number (sctid) of the new line are associated and registered. In this way, a heap tuple map (HTM) is additionally generated every time the database is updated. Note that since the line number in which the instruction content (sc1) of line number 5 is written is still unknown at this stage, a new instruction (sc1) is written in sctid.

  On the other hand, for the shadow copy (SC), a shadow copy of the line number to be rewritten with reference to the master database (11a) is generated. At this time, since the newly added line number is unknown at this stage, a new command (sc1) is written in the line number.

  At this stage, the database processing unit (11b) of the lower master node (MS301) uses the line number 4 to which the DELETE instruction is applied by generating the heap tuple map (HTM) and the old line number 5 to which the UPDATE instruction is applied. Since it is already known that the command is deleted, only a new command (sc1) may be written as the shadow copy (SC).

  Further, as described above, as the transaction minimum value (Xmin) added to the writing set, the transaction minimum value (Xmin) recorded on the snapshot referred to by the transaction being executed is written as it is.

  The writing set generated in this way is transmitted from the lower master node (MS301) to the middle master node (for example, MS201) and further to the upper master node (MS101).

  In the upper master node (MS101), the database processing unit 11b (central processing unit (CPU)) reads the transaction minimum value (Xmin) from the write set received above. Then, the transaction minimum value (Xmin) is compared with the cluster minimum value (Cmin) held in the back-end memory (BEM) of the upper master node (MS101). At this time, when the transaction minimum value (Xmin) is equal to or greater than the cluster minimum value (Cmin), the transaction log processing unit (11c) is activated to start generating transaction log data, and then the following processing is performed.

  First, a heap tuple map (HTM) is read and compared with its own master database 11a. Here, it is verified whether or not the contents of the target tuple (here, line numbers 4, 5, and 7) are updated on the database (11a). In FIG. 4, since the line numbers 4-6 are not updated, a deletion pointer is assigned to the line number 4 and a deletion pointer is also assigned to the old number 5 to be rewritten. Then, a new command (sc1) is written in the new line number 7.

  When the heap tuple map (HTM) in the writing set from the lower master node (M301) is compared with its own database, the row is already updated by another writing set in the upper master node (M101). The processing by the writing set is aborted.

  On the other hand, when the transaction minimum value (Xmin) is smaller than the cluster minimum value (Cmin) as compared with the transaction minimum value (Xmin) and the cluster minimum value (Cmin) in the write set, the write set is aborted. .

The reason for aborting the writing set will be described below.
In the writing set, there is a possibility that a tuple deleted by a transaction having a transaction ID smaller than the cluster minimum value (Cmin) is an update target. There is a high possibility that such a tuple has already been discarded based on the setting of the cluster minimum value (Cmin) of the upper master node. In that case, consistency cannot be detected for this tuple using the heap tuple map, resulting in a contradiction in the upper master node. Therefore, the writing set having the transaction minimum value (Xmin) smaller than the cluster minimum value (Cmin) is aborted to prevent the occurrence of the contradiction.

  If the discard of the transaction based on the notification of the cluster minimum value (Cmin) is completed at the middle / lower master node, such a write set is not generated. However, the write set generated in the middle / lower master node while notifying the cluster minimum value (Cmin) from the upper master node to the lower master node has a transaction minimum value smaller than the cluster minimum value (Cmin). Therefore, it is necessary to check when the writing set is applied on the upper master node. Therefore, as described above, the transaction minimum value (Xmin) registered in the writing set is compared with the cluster minimum value (Cmin), and if the transaction minimum value (Xmin) is small, the writing set is aborted. This is the solution.

  In this way, the notification of the cluster minimum value (Cmin) can discard all transactions that refer to the tuples that have been completely erased, and the write set created in the wrong way of the notification (Cmin) notification is also the cluster minimum value. Abort can be performed based on (Cmin). Therefore, tuples can be completely deleted even in the middle and lower nodes without losing the consistency of the database.

  FIG. 5 is an example of transaction log data generated by the transaction log processing unit (11c) when the master database (11a) of the upper master node (MS101) is updated as described above. This transaction log is a file in which at least instructions and transaction contents (line numbers and execution process contents corresponding thereto) are continuously recorded in time series.

  According to the figure, following the transaction start instruction (XB1), a log in which an instruction number and a line number are paired is sequentially generated. For example, line number 4 is deleted first as a DELETE instruction (D1) (D14), then line number 5 is deleted as UPDATE instruction (U1), line number 7 is added (U157), and then these commit instructions ( XC1) is issued. Note that a minimum cluster value (Cmin = 6) may be added to the transaction log data. By including the cluster minimum value (Cmin = 6) in this way, it becomes possible to control the transaction discard order in the intermediate master nodes B, C, D, E or the lower master nodes F, G.

  As described above, when the cluster minimum value (Cmin = 6) is notified to the intermediate master nodes B, C, D, E and the lower master nodes F, G asynchronously with the transaction log data, Efficient transaction discard processing can be performed without waiting for generation / transmission of transaction log data.

  The transaction log data is distributed from the communication module (11d) to the intermediate master node (MS201) and all lower master nodes (MS302... MS30n).

The middle / lower master node that has received the transaction log data replicates the transaction log data to its own database.
Specifically, when the lower level master node (for example, M302) receives the transaction log data shown in FIG. 5 by the communication module 11d, the transaction log processing unit 11c is activated and the transaction log data is replicated to its own master database 11a. To do. As a result, deletion pointers are assigned to the line numbers 4 and 5, and a new line number 7 is added. When the cluster minimum value (Cmin = 6) is added to the transaction log data, a snap having a transaction minimum value (Xmin) smaller than the cluster minimum value (Cmin = 6) is obtained at each master node. All transactions that refer to the shot are discarded.

  As described above, the middle / lower master node uniformly manages the database by replication of the transaction log data transmitted from the upper master node.

As mentioned above, although this invention was demonstrated based on embodiment, this invention is not limited to this. Hereinafter, the modification is demonstrated.
(When the database is updated in the upper master node MS101)
As described with reference to FIG. 2, the processing when the master database update instruction is issued in the lower master node (for example, MS301), heap tuple map (HTM, heap file) and shadow copy on the back-end memory (BEM) (SC) is generated. However, when a master database update command is generated in the upper master node (MS101), it is not necessary to notify the upper node, so a write set is not generated. . That is, in such a case, the upper node (MS101) writes update data directly to the master database (11a) as shown in the left diagram of FIG. 4, and the transaction log data shown in FIG. Generated. The transaction log data is distributed to the lower master node, and the lower master node receiving the transaction log data replicates the transaction log data to its own master database.

(When the search process is executed while the writing set is generated in the lower master node)
When a search is executed on the master database of the subordinate master node in the stage where the subordinate master node (for example, MS 101) is generating a write set as shown in FIG. 3, the write set is generated. There is no problem if the search is for a line number other than the line number, but if the search is for the corresponding line (here, line number 4 and line number 5), these line numbers are already substantial. Since update processing is performed, an accurate search result cannot be obtained even if the corresponding row in the master database is searched.

  In such a case, the database processing unit (11b) first refers to the heap tuple map (HTM) and checks whether or not the search relevant row number has been entered. For example, in the example shown in FIG. 3, when the search target is line number 3, the database processing unit (11 b) determines the heap tuple in the writing set on the back-end memory (BEM) built on the memory (MM). Referring to the map (HTM), it is determined whether or not the corresponding line (line number 3) has been entered. In the example of FIG. 3, the corresponding line is not entered. In that case, the master database 11a is directly accessed to search for the corresponding line (line number 3).

  On the other hand, in the example shown in FIG. 3, when the search target is line number 4, when the database processing unit (11b) refers to the heap tuple map (HTM) in the writing set, the corresponding line (line number 4). Will be detected. In this case, even if the master database 11a is accessed, there is no meaning as a search target because an update command for deleting the remaining corresponding row has already been executed. In this case, the database processing unit (11b) detects that the corresponding line (line number 4) has been deleted with reference to the heap tuple map (HTM). Thus, since the search target line is deleted, the database processing unit (11b) does not set the corresponding line as the search target.

  On the other hand, in the example shown in FIG. 3, when the search target is the line number 5, the shadow copy (SC1) corresponding to the corresponding line (line number 5) is referred to the heap tuple map (HTM) as described above. Detect that it has been created.

At this time, the database processing unit (11b) may refer to the shadow copy (SC1) and search for the contents in which line number 5 is rewritten.
In the above description, the case where the heap tuple map (HTM) is referred to for each row has been described. However, the present invention is not limited to such a processing method. Referring to the heap tuple map (HTM), search for the master database (11a) excluding all of the line numbers entered here, and then search the heap tuple map (HTM) again. With reference to the entry, the corresponding line may be excluded from the search target (in the case of line number 4), or the shadow copy (SC1) may be the search target.

(If there is a conflict when updating with transaction log data from the upper master node in the lower master node)
When replication is performed to the database of the lower master node by the transaction log data distributed from the higher master node, a conflict occurs if an update instruction for the database of the lower master node is executed for the corresponding row. It will be.

  Specifically, this applies when transaction log data related to line number 5 is delivered from the upper master node when line numbers 4 and 5 are updated based on an update command to the lower master node. To do.

  In such a case, even if the write set created in the lower master node is transmitted to the upper master node, the transaction log data relating to the corresponding row has already been distributed in the upper master node. Will be aborted when a conflict is detected at the higher master node. Therefore, the conflict in the lower master node can be ignored. According to this method, even writing sets that are known to be aborted in the upper master node must be transmitted to the upper master node, increasing the burden on the upper master node and the lower master node. However, there is an advantage that the processing can be speeded up because it is not necessary to check each other at the stage where the transaction log data from the upper master node is replicated at the lower master node.

On the other hand, as another method for solving such a conflict in the lower master node, when a heap tuple map (HTM) is generated in the lower master node, it is registered in the memory in the lower master node, and this The memory is used as a shared memory and is accessible from other subordinate master nodes. Such access between other lower-level node masters is performed by the database processing unit 11b of each lower-level master node. According to this method, since the conflict is checked when the heap tuple map (HTM) is generated, replication by transaction log data may be delayed, but the conflict problem can be solved only between the lower master nodes. Do not put a burden on the master node.

As mentioned above, although this invention was demonstrated based on embodiment, it is not limited to these.
For example, the writing set is generated in the lower master node (MS301). However, the present invention is not limited to this, and the writing set may be generated in the middle master node (MS201).

  The present invention can be used for a database management system having a multi-master node structure having a hierarchical structure.

MS101 Upper master node SL Slave MS201, MS202 ... MS20n Middle master node MS301, MS302 ... MS30n Lower master node CL Client terminal 11a Master database 11b Database processing unit 11c Transaction log processing unit 11d Communication module BEM Backend memory HTM Heap tuple map SC Shadow copy Cmin Cluster minimum value Xmin Transaction minimum value

Claims (3)

A method for managing a write-once database having hierarchically upper and lower master nodes that can be updated,
Sending, in a session of any lower master node, lower database update information including the transaction minimum value of the snapshot referred to by the transaction being executed to the upper master node;
The upper master node that has received the lower database update information compares the transaction minimum value with the cluster minimum value held by the upper master node, and if the transaction minimum value is equal to or greater than the cluster minimum value, Updating the database of the upper master node based on the database update information;
If the transaction minimum value is smaller than the cluster minimum value, aborting the lower database update information; and
Generating an update record of the upper master node as a transaction log;
Delivering the transaction log to a subordinate master node including a subordinate master node of the transmission source;
In the lower master node, updating its own database based on the transaction log received above,
Notifying the lower master node of the cluster minimum value held by the upper master node, and discarding a transaction having a transaction minimum value smaller than the minimum value;
A method for managing a write-once database in an upper master node having hierarchically lower master nodes consisting of The lower database update information is a write set consisting of a shadow copy of the database of the lower master node, a heap tuple map expanded in its own memory, and a transaction minimum value of a snapshot referred to by an executing transaction. Yes,
When the upper master node receives the write set, the transaction minimum value is compared with the cluster minimum value held by the upper master node, and if the transaction minimum value is equal to or greater than the cluster minimum value, The heap tuple map is compared with its own database to verify whether there is an update in the database of the tuple registered as a target. When the update is made, the write set is aborted, and no update is made. Sometimes updating the database with the shadow copy;
If the transaction minimum is less than the cluster minimum, aborting the write set;
The write-once database management method according to claim 1, comprising: In addition to the above, there is an intermediate master node of one or more layers between the upper and lower master nodes,
In any one of the sessions of the middle master node, sending to the upper master node the middle database update information including the minimum transaction value that is the target of updating the database of the middle master node. ,
In the upper master node that has received the intermediate database update information, the transaction minimum value is compared with the cluster minimum value held by the upper master node, and if the transaction minimum value is equal to or greater than the cluster minimum value, The method for managing a write-once database according to claim 1 or 2, further comprising the step of updating the database of the upper master node based on the intermediate database update information.

Images