State recovery system, state recovery method, and program

Abstract

This invention provides a state restoration system, method, program, and recording medium that efficiently generate or restore a target state based on dependency analysis between information states. [Solution] The state restoration system includes a dependency generation unit that generates a dependency structure with nodes and state dependencies as edges, and a boundary state determination unit that determines a boundary state set on which a target information state can be generated. The boundary state set is determined as the set with the minimum number of elements among sets that satisfy the condition that closure(F,R) encompasses the target information state set. The system also includes a state generation unit that generates or restores the target information state by applying dependencies in the forward direction using the boundary state set as the initial state. This reduces storage capacity and improves processing efficiency.

Description

Technical Field

[0001] The present invention relates to information processing technology, and particularly to a state restoration technology for efficiently generating or restoring a target state based on an analysis of the dependency relationships between information states. More specifically, based on a dependency structure represented by nodes for information states and edges for the generative dependency relationships between information states, it relates to an information processing technology for determining a minimum set of boundary states that enables the regeneration of a target information state and regenerating or restoring the target state from the set of boundary states. For example, the present invention is applicable to the following fields. · Database state restoration · Cache reconstruction · Distributed system state synchronization · Log compression · Generative AI inference state restoration · Large-scale data processing infrastructure

[0002]

Background Art

[0002] In recent years, with the large-scale and decentralized development of information processing systems, a process of regenerating or restoring a target information state based on past information states has been widely used. ​Furthermore, distributed systems utilize checkpoints and logs for state synchronization. Alternatively, state restoration is being performed. On the other hand, generative AI systems, particularly autoregressive inference such as large-scale language models (LLMs), are different. The system uses intermediate states such as a KV cache to enable the continuation of inference. A method for saving data is commonly used. However, these conventional techniques require a large amount of data to restore the target state. Logs, history, or intermediate states must be saved. • Increased storage capacity • Increased memory consumption • Increased recovery processing time • Increased communication volume in distributed systems Problems such as these may occur. In particular, in large-scale data processing systems and generative AI inference systems, The number of states and dependencies becomes very large, so we need to reduce the amount of state information that needs to be saved. There is a need for technology that can efficiently restore the state while simultaneously addressing the issue. [Overview of the Initiative] [Problems that the invention aims to solve]

[0003] This invention was made in view of the above-mentioned problems, Based on the dependency structure between multiple information states, generate a target information state or By determining the minimum set of boundary states required for reconstruction, Achieves efficient state restoration while reducing the number of information states that need to be saved. The purpose is to provide information processing technology. Furthermore, the present invention is • Reduce the amount of data stored. • Faster recovery processing • Reduced data usage • Improved system operation efficiency The objective is to provide a state restoration technology that enables this.

Means for Solving the Problem

[0004] According to one aspect of the present invention, state acquisition means for acquiring a plurality of information states; dependency generation means for generating a dependency structure representing the dependency relationship between the information states; and boundary determination means for determining a set of boundary states that enables generation or restoration of a target information state based on the dependency structure; and state generation means for generating or restoring the target information state using the set of boundary states. A state restoration system including these components is provided. and The dependency structure is represented as a directed graph structure having information states as nodes and generation dependency relationships between states as edges. The set of boundary states is determined as a set that satisfies the following conditions among the sets of information states F that can generate the target information state set T. closure(F, R) ⊇ T Here, R: Set of dependency relationships closure(F, R): The set of states that can be generated by applying the dependency relationship R in the forward direction from the set F is. Furthermore, the set of boundary states F is determined as the set having the minimum number of elements among the sets that satisfy the above conditions. When there is a cycle in the dependency structure, a reduced graph obtained by reducing the dependency graph in units of strongly connected components is generated, and the boundary state set determination process can be executed on the reduced graph. The determination of the set of boundary states is performed by searching for dependency relationships in the reverse direction from the target state.​​​​​​​ This generates a candidate set, and from that candidate set, a set that satisfies the minimum set condition is selected. This can be achieved by selecting [this option]. This results in the minimum set of states required to generate or restore the target state. Only the data can be saved. [Effects of the Invention]

[0005] According to the present invention, the necessary to generate or restore the target information state Only the smallest set of boundary states can be saved. For example, for the number of history entries N Since the number of required states can be reduced to k (k << N), Storage capacity and recovery time can be significantly reduced. the result, ·Storage capacity reduction • Faster recovery processing • Reduced data usage • Reduced memory consumption This can be achieved. Furthermore, the present invention is • Database restoration • Cache rebuilding • Distributed system state synchronization • Generative AI inference state recovery It can be applied to a wide range of information processing fields, such as those listed above. In particular, in generative AI inference systems, based on inference state dependencies By saving only the minimum necessary inference state, the KV cache storage amount is reduced. This makes it possible to reduce the amount of data used and then resume the inference process. [Brief explanation of the drawing]

[0006] [Figure 1] This figure shows the basic configuration of the state restoration system according to the present invention. [Figure 2]This figure shows an example of a dependency graph structure according to the present invention. [Figure 3] This figure shows the concept of boundary state set determination according to the present invention. [Figure 4] This figure shows the process for obtaining a set of candidate boundary states by reverse search according to the present invention. [Figure 5] This figure shows an example of a dependency relationship graph including a circular dependency according to the present invention. [Figure 6] This figure shows the concept of the strongly coupled component reduction process according to the present invention. [Figure 7] This figure shows an example of a processing configuration when the present invention is implemented as a program. [Figure 8] This figure shows an example configuration in which the present invention is applied to a database state restoration system. [Figure 9] This figure shows an example configuration in which the present invention is applied to a cache reconstruction system. [Figure 10] This figure shows an example configuration in which the present invention is applied to distributed system state synchronization. [Figure 11] This figure shows an example configuration in which the present invention is applied to a generative AI inference state restoration system. [Figure 12] This figure shows an example configuration in which the present invention is applied to a log compression and history minimization system. [Modes for carrying out the invention]

[0007] Embodiments of the present invention will be described with reference to the drawings. This invention utilizes a dependency structure between multiple information states, The minimum set of boundary states required to generate or restore the target information state. This technique determines the (Frontier) and regenerates the target state from that set. In this invention, information states are represented as nodes, and generation dependencies between states are represented as edges. A dependency graph is constructed, and the target state can be generated on that graph. Find the minimum set of states. Specific embodiments of the present invention will be described below. [Examples]

[0008] Example 1: Minimum Boundary State Set Determination Algorithm Based on Dependency Graph This example describes the basic algorithm of the present invention. Figure 1 shows the basic configuration of the state restoration system 100 according to this embodiment. The state restoration system 100 comprises the following functional units. 101 Status acquisition unit 102 Dependency Generation Unit 103 Boundary state determination unit 104 State generation unit (1) Acquisition of information status The state acquisition unit 101 acquires multiple information states that exist within the system. An information state is generated as a result of a computation or information processing. It refers to any unit of information. Examples of information states include the following: • Data record status • Calculation result status • Cache status • Log status • Processing result status • Inference state The acquired information state is stored as a state set S. S = {s1, s2, s3, ..., sn} (2) Generation of dependency structure The dependency generation unit 102 analyzes the generation dependency relationships between information states, Generate a dependency structure R. Dependency structures are represented as directed graphs. G = (S, R) Here S: Information state set R: Dependency set That is the case. edge (si → sj) teeth, "In order to generate state sj, state si is required." It means that. Dependencies can be extracted using methods such as the following: • Processing log analysis • Data flow analysis • Computational graph analysis Transaction history analysis ·Inference dependent analysis Figure 2 shows an example of a dependency graph. In this diagram, to generate the information state s4 This indicates that state s2 or s3 is required. (3) Specifying the target state The user or system specifies the target state set T to be restored. T ⊆ S for example T = {s_k} That is the case. (4) Definition of boundary state set In this invention, the boundary state set F is defined as a set that satisfies the following conditions. closure(F,R) ⊇ T Here closure(F,R) This is achieved by applying the dependency R in the forward direction from the states contained in the set F. Represents a set of generateable states. In other words closure(F,R) = F ∪ {all states that can be generated from F by dependencies} That is the case. Furthermore, the boundary state set F is, |F| = minimum It is determined to be a set such that... Figure 3 illustrates the concept of boundary state set determination. The boundary state set F is such that the closure(F,R) based on the dependency R is Conditions for containing the target state set T closure(F,R) ⊇ T It is defined as a set that satisfies the following conditions: Furthermore, in the present invention, among the sets that satisfy the conditions The set with the minimum number of elements is determined as the boundary state set. For example, if the dependency A → B → C → D If that is the case, Target state T = {D} For the boundary state set F is F = {A} This is the result. In this invention, among the sets that satisfy the conditions, The set with the minimum number of elements is determined as the boundary state set F. (5) Boundary state set determination algorithm The boundary state determination unit 103 determines the boundary state set by following the procedure below. Step 1 Obtain the target state set T. Step 2 In the dependency graph G, Explore dependencies in reverse order from the target state T. Step 3 The set of reachable states C is obtained by reverse search. C = business_closure(T,R) Figure 4 illustrates the concept of reverse search processing. Step 4 Generate a boundary candidate set from the candidate set C. Wi5 Among the sets satisfying closure(F,R) ⊇ T, The set with the minimum number of elements is selected as the boundary state set F. (6) Handling circular dependencies If a cyclic structure exists in the dependency graph, A simple search can potentially result in an infinite loop. Figure 5 shows an example of a dependency relationship graph that includes circular dependencies. Therefore, in this invention, Dependency graphs are represented by Strongly Connected Components (SCCs). Generate a reduced graph by reducing the units. Figure 6 illustrates the concept of strongly connected component reduction. By performing boundary state set determination processing on the reduced graph, Processing can be performed safely even when circular dependencies are involved. (7) State restoration process The state generation unit 104 uses the boundary state set F as the initial state. The target state T is generated by applying dependencies in the forward direction. The process will be executed as follows: Step 1 Obtain the boundary state set F. Step 2 Apply the dependency R in the forward direction. Step 3 Add the generated state to the state set. Step 4 The process is repeated until the target state T is generated. (8) Specific examples Dependencies A → B → C → D If that is the case, Target state T = {D} against Backward search results C = {A, B, C, D} This is the result. Select the set that satisfies the minimum set condition from the candidate set C. F = {A} This is the result. State generation process A → B → C → D The states are regenerated in that order, and the target state D is restored. (9) Amount of calculation The number of nodes in the dependency graph |S| Let the number of edges be |R|. The computational complexity of the reverse search process O(|S| + |R|) This is the result. Also, the number of states to be saved |F| And, Storage capacity of conventional methods O(|S|) against In this invention O(|F|) This is the result. (10) Effects According to this embodiment, Only the information states necessary to generate the target state are included. Since it can be stored as a boundary state set, ·Storage capacity reduction • Faster recovery processing • Reduced system load This can be achieved. Especially in large-scale information processing systems with an enormous number of states It can yield significant results. Furthermore, the process of the present invention can be implemented as a computer program. Figure 7 shows the implementation of the present invention as a program. This is a diagram showing an example of a computer configuration.

[0009] Example 2: Database State Restoration System In this embodiment, the boundary state set determination technique of the present invention is used This section explains how to apply this to database state restoration processing. Figure 8 shows the database state restoration system 200 according to this embodiment. This is a diagram showing an example configuration. The database state restoration system 200 comprises the following functional components. 201 Data Acquisition Unit 202 Update History Analysis Department 203 Dependency Generation Unit 204 Boundary state determination unit 205 Recovery Processing Unit 206 State preservation section (1) System Configuration The data acquisition unit 201 retrieves table data stored in the database. Retrieve update history, transaction log, or event log. The update history analysis unit 202 analyzes the transaction log or update history, Extract dependencies between update operations. The dependency generation unit 203 uses update operations as nodes, A dependency graph is generated, with the dependencies between update operations as edges. The boundary state determination unit 204 restores the target database state. Determine the minimum set of boundary states required. The restoration processing unit 205 starts from the boundary state set The target state is regenerated by applying the update operation in a forward direction. The state storage unit 206 stores only the states included in the boundary state set. (2) Update operation model Update operations in a database are, It can be represented as a structure like the following. UpdateOperation Operation_id timestamp target_record update_value dependency_operation Here operation_id :Update operation identifier timestamp: update execution time target_record: Record to be updated update_value: update value dependency_operation: dependency update operation That is the case. (3) Dependency graph generation The update history analysis unit 202 is: Analyze the dependencies between update operations. For example, consider the following update sequence. Operation A: Update Record R1 Operation B: Depends on the result of Operation A OperationC: Depends on the result of OperationB In this case, the dependencies are OperationA → OperationB → OperationC It is expressed as follows. The dependency generation unit 203 uses these as nodes and edges. Generate a dependency graph G. G = (U, E) U: update operation set E: Dependency set (4) Specify target state Specify the target database state to be restored. For example, when restoring the data state at a specific point in time, The set of update operations corresponding to that point in time is set as the target state T. example T = {OperationC} (5) Determination of boundary state set The boundary state determination unit 204 performs the following processing. Step 1 Perform a reverse search from the target update operation T. Step 2 By traversing dependencies, we obtain the set of update operations C that can be reached. example C = {OperationA, OperationB, OperationC} Step 3 From set C closure(F,R) ⊇ T The smallest set satisfying the condition is determined as the boundary state set F. example F = {OperationA} (6) Restoration process The restoration processing unit 205 is: As the boundary state set F is the initial state Apply the update operation in the forward direction. Specifically, the following processes will be performed. Step 1 Apply Operation A Step 2 Apply OperationB Step 3 Apply OperationC This restores the target state. (7) Conservation strategies In conventional database recovery methods, One of the following needed to be saved. • Full transaction log • Regular snapshots • Checkpoints In contrast, in the present invention, Only the boundary state set F will be saved. the result, Storage quantity Conventional method O(N) against This invention O(|F|) This is the result. Here N: Total number of update operations |F| : Boundary state set size (8) Effects According to this example • Reduce transaction log storage volume • Faster recovery processing • Reduced data storage costs This can be achieved. In particular, in large-scale database systems Because the log volume becomes very large, The effect of reducing storage volume according to this invention is very significant. (9) Applicable databases The present invention is applicable to the following database systems. • Relational database • Distributed databases • NoSQL database Event Stream Database • Log structure database Especially in cloud database environments It can yield significant results.

[0010] Example 3: Cache Reconstruction System In this embodiment, the boundary state set determination technique of the present invention is used This section explains how this can be applied to cache rebuilding processes. Figure 9 shows an example configuration of the cache reconstruction system 300 according to this embodiment. The cache reconstruction system 300 comprises the following functional components: 301 Cache status acquisition unit 302 Cache Generation Relationship Analysis Unit 303 Dependency Graph Generation Unit 304 Boundary State Determination Unit 305 Cache Regeneration Unit 306 Cache Storage Section (1) System Configuration The cache state acquisition unit 301 is located in the cache system. Retrieve cached items. The cache generation relationship analysis unit 302, The data or processing result from which each cache item was generated. Analyze it. The dependency graph generation unit 303 is: Determining generation dependencies between cache items as nodes and edges Generate a dependency graph. The boundary state determination unit 304 is: Determine the minimum set of boundary states required for cache regeneration. The cache regeneration unit 305 is: The cache items are regenerated in a forward direction, starting from the boundary state set. The cache storage unit 306 is Only the boundary state set will be saved. (2) Cache item model Cache items can be represented by the following information structure. CacheEntry cache_id input_data generation_rule critical_cache Here cache_id: Cache identifier input_data: Input data used for cache generation generation_rule: Cache generation processing rules dependency_cache: Dependency cache item That is the case. (3) Cache dependencies Cache items may depend on other cache items. For example, consider the following cache generation sequence. CacheA: Generated from base data CacheB: Created from CacheA CacheC: Generated from CacheB In this case, the dependencies are CacheA → CacheB → CacheC It is expressed as follows. (4) Dependency graph generation The dependency graph generation unit 303 is: Based on the analysis results of the cache generation relationship analysis unit 302, Generate a dependency graph G. G = (C, E) Here C: Cache Set E: Dependency set That is the case. edge (Ci → Cj) teeth "Cache Ci is required to generate cache Cj." It means that. (5) Target cache specification Specifies the cache set T to be regenerated. for example T = {CacheC} That is the case. (6) Determination of boundary state set The boundary state determination unit 304 performs the following processing. Step 1 Explore dependencies in reverse order from the target cache T. Step 2 The cache set that can be reached is obtained by tracing the dependencies. example C = {CacheA, CacheB, CacheC} Step 3 From set C closure(F,R) ⊇ T Among the sets that satisfy The set with the minimum number of elements is determined as the boundary state set F. example F = {CacheA} (7) Cache regeneration The cache regeneration unit 305 is: The cache generation process is performed using the boundary state set F as the initial state. Apply in the forward direction. Specifically, the following processes will be performed. Step 1 CacheA generation Step 2 CacheB generation Step 3 CacheC generation This will regenerate the target cache. (8) Cache preservation strategies In conventional cash systems, Save all cached items, Alternatively, deletion based on TTL (Time To Live) or similar criteria is a more common method. However, in large-scale web services and AI systems Because the cache size becomes enormous, Cache storage capacity becomes an issue. In this invention, Only the boundary state set F will be saved. As a result, the amount of storage Conventional method O(|C|) against This invention O(|F|) This is the result. Here |C|: Total cache size |F|: Boundary state set size (9) Applicable cache systems The present invention is applicable to the following cache systems. · Web cache · CDN cache · API response cache · Search result cache · AI inference cache (10) Effects According to this embodiment · Reduction of cache storage capacity · Improvement of cache regeneration efficiency · Reduction of cache update cost can be achieved. Particularly, in large-scale Web services and AI inference platforms significant effects can be obtained.

[0011] Example 4: Distributed system state synchronization In this example, when the boundary state set determination technology of the present invention is applied to the state synchronization process in a distributed computer system it will be described. FIG. 10 is a diagram showing a configuration example of a distributed system state synchronization system 400 according to this example. is a diagram showing a configuration example. The distributed system state synchronization system 400 is composed of a plurality of computing nodes. Each node includes the following functional units. 401 State acquisition unit 402 State dependency analysis unit 403 Dependency graph generation unit 404 Boundary state determination unit 405 State Sharing Unit 406 State Regeneration Unit (1) Distributed System Configuration In a distributed system, multiple nodes each hold a computational state or a data state. For example Node1 Node2 Node3 Node4 The system is composed of multiple nodes such as these. Each node holds a local state. Node1: State set S1 Node2: State set S2 Node3: State set S3 Node4: State set S4 The state set of the entire system is S = S1 ∪ S2 ∪ S3 ∪ S4 and is represented as such. (2) State Dependence Relationship In a distributed system,[[]] the state of a certain node may depend on the states of other nodes. For example s1: Generated at Node1 s3: Generated at Node2 s6: Generated at Node3 s9: Generated at Node4 Let it be so. The state dependence relationship is s1 → s6 s3 → s6 s6 → s9 If that is the case, To generate state s9 s1 and s3 are required. (3) Dependency graph generation The state dependency analysis unit 402 is, Analyze the state dependencies between nodes. The dependency graph generation unit 403 is: States are represented as nodes and dependencies as edges. Generate a dependency graph G. G = (S, R) Here S: Set of states R: Dependency set That is the case. (4) Specify target state Specify the target state set T to be restored or synchronized. for example T = {s9} That is the case. (5) Determination of boundary state set The boundary state determination unit 404 is: The following process is performed. Step 1 Explore dependencies in reverse order from the target state T. Step 2 Obtain the set of reachable states C. example C = {s1, s3, s6, s9} Step 3 closure(F,R) ⊇ T Among the sets that satisfy The set with the minimum number of elements This is determined as the boundary state set F. example F = {s1, s3} (6) State sharing process The state sharing unit 405 is Only the boundary state set F is shared with other nodes. In the conventional method It was necessary to synchronize all states or logs. In this invention Only the boundary state needs to be shared. (7) State regeneration The state regeneration unit 406 is, Starting from a shared boundary state set F Apply dependencies in the forward direction. The process will be executed as follows: Step 1 Boundary state acquisition Step 2 Dependency application Step 3 Intermediate state generation Step 4 Target state generation (8) Reduce data usage In the conventional method Full state synchronization or log synchronization is required. Data usage O(|S|) This is the result. In contrast, the present invention Data usage O(|F|) This is the result. Here |S| : Total number of states |F| : Boundary state set size (9) Applicable distributed systems The present invention is applicable to the following distributed systems. • Cloud computing infrastructure • Distributed databases • Microservices architecture • Distributed AI processing system • Blockchain network (10) Effects According to this example • Reduced data usage • Faster synchronization processing • Reduced load on distributed systems This can be achieved. Especially in large-scale cloud systems This can result in a significant reduction in data usage.

[0012] Example 5: Generative AI Inference State Restoration System In this embodiment, the boundary state set determination technique of the present invention is used as follows: Applicable to generative AI, particularly autoregressive inference processing such as large-scale language models. Let me explain the situation. In the inference process of generative AI, The generation of each token depends on the preceding token. For example, in the Transformer model... Attention relationships between tokens are dependencies It is expressed. Therefore, the inference state is • Input token • Generated tokens Attention dependency As a dependency graph composed of It can be expressed. Figure 11 shows the generation AI inference state restoration system 500 according to this embodiment. This is a diagram showing an example configuration. The Generative AI Inference State Restoration System 500 comprises the following functional units. 501 Input Token Acquisition Unit 502 Inference Dependency Analysis Unit 503 Dependency Graph Generation Unit 504 Boundary State Determination Unit 505 Inference State Storage Unit 506 Inference State Reconstruction Unit 507 Inference Execution Unit (1) System Configuration The input token acquisition unit 501 is, User input, system prompts, interaction history, Obtain external search results or external data, Generates a sequence of inference input tokens. The inference dependency analysis unit 502 is: The inference state or token generation state is Analyze which past tokens or intermediate states it depends on. The dependency graph generation unit 503 is: The inference state is a node, Generate a dependency graph with inference dependencies as edges. The boundary state determination unit 504 is: Enables the generation of a target inference state. Determine the smallest set of boundary states. The inference state storage unit 505 is, Only the boundary state set will be saved. The inference state restoration unit 506 is: Regenerate the inference state from the saved boundary state. The inference execution unit 507 is: The inference process continues using the recovered inference state. (2) Inference state model In the inference process of generative AI, Input token column x1, x2, x3, …, xn An output token is generated based on this. In this embodiment, Treat the following information as independent information states. • Input token state • Embedding vector state • Intermediate layer expression state • Attention mechanism status Key status Value state • KV cache status • Output token state (3) Inference dependencies The inference dependency analysis unit 502 is: Analyze the dependencies between inference states. Dependencies can be determined, for example, by the following method: • Attention weights in the self-attention mechanism • Attention weight threshold determination • Contribution score Interlayer propagation relationship • Token generation reference history for example, For the generation of token t_j The attention weight of token t_i is If it is above a predetermined threshold, t_i → t_j Set the dependencies. (4) Dependency graph The dependency graph generation unit 503 is: By the set of inference states V and the set of dependencies E The following graph G is generated. G = (V, E) Here V: Set of inference states E: Dependency set That is the case. edge (v_i → v_j) teeth "To generate the inference state v_j" The inference state v_i is required. It means that. (5) Target inference state The target inference state T is, For example, it could be any of the following: • Specific output token generation state • KV cache status at a specific point in time • Intermediate representation state of a specific layer • Dialogue response generation complete state • Inference resumed state for example, Tokens up to t_n have already been generated, If you want to resume the generation process from there, T = {t_n generation state} Set it as follows. (6) Determination of boundary state set The boundary state determination unit 504 performs the following processing. Step 1 Obtain the target inference state T. Step 2 In a dependency graph Perform a search in the reverse direction from the target state. Step 3 Obtain the set of reachable inference states C. Step 4 closure(F,R) ⊇ T Among the sets that satisfy The set with the minimum number of elements This is determined as the boundary state set F. (7) KV cache reduction In conventional generative AI inference, To resume inference It was necessary to store a large KV cache. In this invention, To generate the target inference state Only the minimal inference state set is saved. for example, Token series t1, t2, t3, t4, t5, t6 In contrast, Tokens that contribute to the creation of t6 t2, t3, t4 If that is the only case, Boundary state set F = {state t2, state t3, state t4} It can be done this way. (8) Inference state recovery The inference state restoration unit 506 is: As the boundary state set F is the initial state Apply dependencies in the forward direction. This Intermediate expression • Attention state • KV cache status Regenerate the following, Reconstruct the target inference state. (9) Application to distributed inference This invention It can also be applied to distributed inference systems. Only the boundary state set By sharing among nodes, • Reduced data usage • Reduced memory usage ·Speed ​​up inference resumption This can be achieved. (10) Effects According to this example • Reduced KV cache storage capacity • Improved efficiency in resuming inference ·Reducing distributed inference traffic This can be achieved. Especially in large-scale generative AI systems Significant memory reduction and improved processing efficiency It can be obtained.

[0013] Example 6: Log compression and history minimization system In this embodiment, the boundary state set determination technique of the present invention is used as follows: Log compression and history management processing in information processing systems Let's explain when this applies. Figure 12 shows an example configuration of the log compression system 600 according to this embodiment. The log compression system 600 includes the following functional components: 601 Log acquisition unit 602 Operation Dependency Analysis Unit 603 Dependency Graph Generation Unit 604 Boundary State Determination Unit 605 Minimum history storage unit 606 State regeneration unit (1) System Configuration The log acquisition unit 601 is, Retrieve operation logs or event logs generated by the system. The logs include information such as the following: • Operation identifier ·Operation time • Target of operation ·Operation details Operational dependencies The operation dependency analysis unit 602 is: Analyze the dependencies between operations in the log. The dependency graph generation unit 603 is: The operation log will be used as a node. Generate a dependency graph with operation dependencies as edges. The boundary state determination unit 604 is: The minimum set of operations that makes the target state reproducible Determine it as the boundary state set. The minimum history storage unit 605 is, Only the boundary state set will be included in the history storage. The state regeneration unit 606 is, From the preserved boundary state set Regenerate the necessary operation history. (2) Log model The operation log can be represented with the following structure. LogEntry log_id timestamp Operation calculation type target_object operation_result Latency_Relationship_Relationship Here log_id: log identifier timestamp: execution time Operation type: Operation type target_object: the object to be operated on operation_result: Operation result dependency_log: Dependency log That is the case. (3) Operational dependencies Operation logs may depend on other operation logs. For example, consider the following sequence of operations. LogA: Data generation LogB: Updates the results of LogA Use the results of LogC :LogB in this case LogA → LogB → LogC A dependency exists. (4) Dependency graph The dependency graph generation unit 603 The following graph will be generated. G = (L, E) Here L: Log set E: Dependency set That is the case. (5) Specify target state The target state T is, System state at a specific point in time, or as a system state for a specific version It can be configured. for example T = {LogC} That is the case. (6) Determination of boundary state set The boundary state determination unit 604 The following process will be executed. Step 1 Perform a reverse search from the target log T. Step 2 Obtain the reachable log set C. example C = {LogA, LogB, LogC} Step 3 closure(F,R) ⊇ T Among the sets that satisfy The set with the minimum number of elements This is determined as the boundary state set F. example F = {LogA} (7) History compression In traditional log management, All log history needs to be saved. Storage quantity O(|L|) This is the result. In this invention Only the boundary state set F is preserved. Storage quantity O(|F|) This is the result. Here |L| : Total number of logs |F| : Boundary state set size (8) Regenerate history The state regeneration unit 606 is, Starting from a conserved set of boundary states F Apply dependencies in the forward direction. This Regenerate the necessary history operations, The target state can be restored. (9) Applicable Systems The present invention is applicable to the following history management systems. Event log management system • Audit log system • Event Sourcing System • Software version history management • Operation history management system (10) Effects According to this example • Reduce log storage capacity • Improved efficiency in history management • Improved efficiency in managing audit logs This can be achieved. Especially in large-scale systems Because the number of logs becomes enormous. The effect of reducing storage volume according to this invention is very significant. [Industrial applicability]

[0014] This invention is based on the dependency structure between information states. Determine the minimum set of boundary states that allows the target information state to be regenerated. Generate or restore the target state from the set of boundary states. Regarding information processing technology. This invention can be widely used in the following industrial fields. • Database recovery system • Cache reconstruction system • Distributed system state synchronization • Cloud computing infrastructure • Generative AI Inference Platform • Log compression and history management system • Large-scale data processing infrastructure • Event sourcing system In particular, large-scale cloud environments, AI inference platforms and In a distributed data processing system, Storage capacity reduction Reduce data usage Speed ​​up the restoration process Because you can obtain the effect, It has extremely high applicability in the information processing industry. Therefore, the present invention can be widely used in the information processing industry.

Claims

1. A state acquisition means for acquiring multiple information states, A dependency generation means for generating a dependency structure that represents the generation dependency relationships between the aforementioned information states, Boundary determination means for determining a boundary state set F that can generate or restore a target information state set T based on the dependency structure, A state generation means that generates or restores the target information state set T by applying dependencies in the forward direction with the boundary state set F as the initial state, and Equipped with, The aforementioned boundary state set F is The closure(F,R) is determined as the set with the minimum number of elements among the sets that satisfy the condition that the target information state set T is contained within it. A state restoration system characterized by the following features.

2. The aforementioned dependency structure can be represented as a directed graph structure in which information states are nodes and generation dependencies between information states are edges. The state restoration system according to feature 1.

3. The boundary determination means generates a candidate set by searching for dependencies in reverse from the target information state set, The boundary state set is determined from the candidate set. A state restoration system according to claim 1 or 2, characterized by the features described above.

4. If a circular dependency exists in the dependency structure, a reduced graph is generated by reducing the dependency graph to units of strongly connected components. Determine the boundary state set on the reduced graph. A state restoration system according to any one of claims 1 to 3.

5. The boundary determination means determines the boundary state set using a depth-first search, breadth-first search, or graph optimization algorithm. A state restoration system according to any one of claims 1 to 4.

6. The aforementioned information state includes any of the following: calculation result state, data update state, cache state, log state, or inference state. A state restoration system according to any one of claims 1 to 5, characterized by the following:

7. The state restoration system restores the database state based on the dependencies of the database update operations. A state restoration system according to any one of claims 1 to 6, characterized by the features described above.

8. Regenerate the cache state based on the generation dependencies between cache items. A state restoration system according to any one of claims 1 to 6, characterized by the features described above.

9. Applicable to state synchronization processing in distributed computer systems, Only the boundary state set is shared between nodes. A state restoration system according to any one of claims 1 to 6, characterized by the features described above.

10. The inference state is restored based on the dependencies between token states or intermediate inference states in the generative AI inference process. A state restoration system according to any one of claims 1 to 6, characterized by the features described above.

11. Compress historical information based on the dependencies of operation logs or event logs. A state restoration system according to any one of claims 1 to 6, characterized by the features described above.

12. The process of acquiring multiple information states, A step of generating a dependency structure between the aforementioned information states, A step of determining a boundary state set that enables the generation of a target information state set, A step of generating or restoring the target information state set by applying dependencies in the forward direction using the boundary state set as the initial state; A method for restoring a state, including the following.

13. to the computer Information status acquisition process, Dependency structure generation process, Boundary state set determination process, State generation process based on boundary state set A state restoration program that executes the necessary actions.

14. A computer-readable recording medium on which the program described in claim 13 is recorded.

Images