Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of Claims
Claims 2, 12, and 17 are cancelled. Claims 21 – 23 are new. Claims 1, 3 – 8, 10 – 11, 13 – 16, and 18 – 20 are amended. Claims 1, 3-11, 13-16, 18-23 are pending and examined herein.
Claims 1, 3-11, 13-16, 18-23 are rejected under 35 U.S.C. 103.
Response to Amendment
The amendment filed March 17th, 2026 has been entered. Claims 2, 12, and 17 are cancelled. Claims 21 – 23 are new. Claims 1, 3 – 8, 10 – 11, 13 – 16, and 18 – 20 are amended. Claims 1, 3-11, 13-16, 18-23 are pending and examined herein. Applicant’s amendments to the claims have overcome each and every objection and 112(b) rejection previously set forth in the Non-Final Rejection Office Action mailed December 17th, 2025.
Response to Arguments
Applicant’s arguments, see pages 9 – 11, with respect to 35 U.S.C. § 101 rejection have been fully considered and are persuasive in view of the amendments. The 35 U.S.C. 101 rejection have been withdrawn.
Applicant’s arguments, see pages 11 – 13, with respect to 35 U.S.C. § 103 rejection have been fully considered but are not persuasive. Applicant argues that the cited references do not teach the amended multi-phase training limitations. However, Hilprecht teaches a reinforcement learning based database partitioning advisor trained using an offline phase and an online phase, where the offline phase uses simulated partitioning information and the online phase executes workloads and uses runtime information as reward information to refine the reinforcement learning model. Therefore, Hilprecht teaches or at least suggests the amended multi-phase training operation. Applicant further argues that Levin does not teach identifying performance issues by processing activity data against multiple threshold values of multiple database operational parameters. However, Levin teaches collecting and processing database activity information, including database operational parameters and performance metrics, and detecting database performance issues using rule based or threshold type criteria. Therefore, Levin teaches or at least suggests the claimed identification of performance issues based on activity data and multiple database operational parameters.
Applicant has not presented separate arguments for the patentability of dependent claims 8 and 23 apart from the argument directed to the independent claims. The amended branch limitations of claims 8 and 23 are addressed in the rejection below, which relies on Guo (NPL:”Learning to Branch for Multi-Task Learning”) instead of Zhang.
Accordingly, the rejection under 35 U.S.C. 103 is remained.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1 – 7, 9 – 20 are rejected under 35 U.S.C. 103 as being unpatentable over Hilprecht et al. (NPL:” Learning a Partitioning Advisor for Cloud Databases”) in view of Levin (U.S. Pub. 2009/0240711 A1).
Regarding Claim 1, Hilprecht teaches
executing a multi-phase training operation for a reinforcement learning model, wherein executing the multi-phase training operation comprises: (Pg. 144 of Hilprecht states “We present a two-step learning procedure to efficiently reduce the training time of our DRL agent that first bootstraps a DRL agent offline (with a simple network centric cost model) and then refines the agent online by actually running real workloads” Also, Figure 1 attached in the next page shows multi-phase training operation.)
training the reinforcement learning model in at least one offline training phase using simulated data associated with multiple partitioning actions; and (Pg. 145 2 Overview section of Hilprecht states “We therefore separate the training process into two phases: (1) offline and (2) online training. In the offline training phase, the agent solely interacts with a “simulation” of the customer database.” Pg. 147 3.2 Problem Modeling section of Hilprecht states “We designed the actions to affect at most the partitioning of a single table. More precisely, we support two types of actions: (1) partitioning a table by an attribute or (2) replicate a table. During training, the RL agent can only select one of these actions at each step. This reduces the repartitioning costs during training since similar partitionings are observed successively. In addition, we provide an action for (de-)activating edges as a short-cut to change the partitioning.” Pg. 148 4.1 Phase 1:Offline Training section of Hilprecht states “During the offline training phase, the database partitioning is simulated and the runtimes are estimated using our network-centric cost model cm(P,qi ) approximating computation and network transfer costs of a given query qi for a partitioning strategy P.”)
implementing at least one online training phase comprising (i) executing one or more workloads using at least a portion of the multiple partitioning actions, and (ii) using one or more runtimes corresponding to the one or more executed workloads to generate one or more rewards for refining the trained reinforcement learning model; (Pg. 145 2 Overview section of Hilprecht states “In an optional online training phase, the agent then does not just interact with a simulation but with a real database. However, instead of using the complete database we only use a sample of the data to speed-up this step of the training phase. The benefit of this phase is that it does not depend on the accuracy of our simple network-centric cost model anymore. Instead, we can simply measure the runtimes of queries on the sampled database to compute the rewards of the agent. Consequently, the agent learns the effects of partitionings more accurately. Once the training is completed, we finally use the agent to make actual partitioning decisions. As input, it requires a workload, i.e. which queries were submitted in a certain time window. Based on this workload, the agent suggests partitionings which we deploy on the actual customer database.” Pg. 147 3.2 Problem Modeling sedction of Hilprecht states “Rewards: The overall goal of the learned advisor is to find a partitioning that minimizes the runtime for the workload mix (queries and their frequencies) modeled as part of the input state. This objective has to be minimized by the DRL agent and can be used as a reward. Estimates of the simple network-centric cost model cm(P,qi ) for the queries qi given a partitioning P are used for the offline training and actual runtimes cr (P,qi ) for the online training. Since the DRL agents seeks to maximize the reward, we use negative costs in the reward definition resulting in r = − Ím j=1 fjc(P,qj ).” Hilprecht’s online training phase deploys actual partitionings, executes real workloads on a sampled database, measures actual runtimes, and uses those runtimes as rewards to refine the DRL agent.)
determining one or more of the multiple partitioning actions to be carried out in connection with the at least one database by processing at least a portion of the activity data related to the one or more identified performance issues using the reinforcement learning model; (Pg. 143 “we propose a different route and make the case to use Deep Reinforcement Learning (DRL) to realize a cloud partitioning advisor as a service that can be used for internal and external DBMS solutions.” Pg.145 “Once the training is completed, we finally use the agent to make actual partitioning decisions. As input, it requires a workload, i.e. which queries were submitted in a certain time window.” Pg. 146 3.2 Problem Modeling section of Hilprecht states “In order to formulate the partitioning problem as a DRL problem we model the database and the workload as state and possible changes in the partitioning as actions. Rewards correspond to the gain in performance for a given workload. During training, the agent thus learns the impact of different partitionings on the workload.”)
… based at least in part on the one or more determined partitioning actions; (Pg. 144 of Hilprecht states Fig. 1 where they choose optimized partitioning action and run SQL query
PNG
media_image1.png
261
820
media_image1.png
Greyscale
)
Hilprecht does not explicitly teach that
A computer-implemented method comprising: identifying one or more performance issues associated with at least one database by processing activity data related to the at least one database;
and performing one or more automated actions
wherein the method is performed by at least one processing device comprising a processor coupled to a memory.
However, Levin teaches that
A computer-implemented method comprising: identifying one or more performance issues associated with at least one database by processing activity data related to the at least one database against multiple threshold values of multiple database operational parameters; ([0011] of Levin states “A method and apparatus for enhancing the performance of an environment comprising a database. The method and apparatus optionally collect data from multiple aspects and components of the environment, including hardware, operating system, database installation, Database schema, database data, activity and others, detect performance issues within the collected data, correlate the issues to reveal analysis issues and suggest recommendations.” [0036] of Levin states “Activity data: relates to the actual application and user activities at run time. This includes actual queries and processes being submitted to and executed by the RDBMS. For servers supporting command tracing such as SQL Server, the command tracing will also be collected.” Levin collects activity and performance data from multiple database operational parameters. Levin’s rule engine then evaluates the collected data using configurable and predetermined threshold type criteria. Under BRI, this constitutes processing activity data against multiple threshold values of multiple database operational parameters to identify performance issues. )
and performing one or more automated actions ([0022] of Levin states “For corrections that can be performed automatically, a script is preferably generated for accomplishing the correction. The script can then be executed automatically or manually by the user.” Hilprecht’s advisor outputs specific ALTER TABLE partitioning commands. Levin’s system generates and automatically executes scripts that implement recommended corrections. Combine to perform automated actions by partitioning scripts produced by AI advisor)
wherein the method is performed by at least one processing device comprising a processor coupled to a memory. ([0024] of Levin states “Server 100 comprises a database server engine 102. Server 100 further comprises or is in communication with storage unit 104. Server 100 is a computing platform, such as a such as a mainframe computer, a personal computer, or any other type of computing platform provisioned with a memory device (not shown), a CPU or microprocessor device, and one or more I/O ports (not shown). Storage unit 104 is preferably a mass storage device, for example an optical storage device such as a CD, a DVD, or a laser disk; a magnetic storage device such as a tape or a hard disk; a semiconductor storage device such as Flash device, memory stick, or the like. Database server engine 102 is preferably a software component which accesses the data stored on storage unit 104, maintains the data and its integrity, and provides additional entities such as clients, with access to the data. Server engine 102 is preferably implemented as interconnected sets of computer instructions.”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings of Levin and Hilprecht. Hilprecht teaches a database partitioning advisor that uses a deep reinforcement learning agent to process database and workload information to suggest partitioning decisions to be deployed on the customer database. Levin teaches a database performance analysis and recommendation system that collects activity data, identify performance issues, and recommendation components with script generation and execution to automatically execute scripts that implements recommended actions. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Levin into the system of Hilprecht because both references address automating database tuning in response to workload and performance information. It would have been obvious choice to further automate the DRL agent to effectively partition based on the recommendation for predictable improvement of the partitioning database.
Regarding Claim 3, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein determining the one or more partitioning actions comprises simulating, using at least a portion of the reinforcement learning model, the one or more partitioning actions in connection with one or more different workloads. (Pg. 145 of Hilprecht states “Inthe offline training phase, the agent solely interacts with a “simulation” of the customer database. Since the network is typically the bottleneck of distributed joins, we developed a simple yet generic cost model focused on the network overhead required to answer a query given a certain partitioning. In combination with the metadata (schema and table sizes) about the customer database, we can estimate the query costs given a partitioning in our simulation. These estimates are used as rewards for the agent.”)
Regarding Claim 4, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein determining the one or more partitioning actions comprises recommending, to at least one of at least one user associated with the at least one database and at least one system associated with the at least one database, the one or more partitioning actions and outputting at least one execution template corresponding to at least a portion of the one or more recommended partitioning actions. (Pg. 147 of Hilprecht states “We designed the actions to affect at most the partitioning of a single table. More precisely, we support two types of actions: (1) partitioning a table by an attribute or (2) replicate a table” [0013] of Levin states “The apparatus can further comprise recommendation components, the recommendation components comprising a recommendation issuing component for issuing one or more recommendations for resolving the situation associated with the analysis rules. Within the apparatus, the recommendation components can further comprise a script generation component for generating a script for resolving the situations associated with the analysis rules, or a script execution component for executing a script for resolving the situation associated with the analysis rules.” [0152] of Levin states “Recommendation components 312 comprise recommendation issuing component 384 for issuing one or more recommendation based on the analyzed issues and their prioritization. The recommendations can take the form of a general recommendation to be implemented by a user, such as “upgrade database to higher version”, or a recommendation that can be performed automatically. For such recommendations, a script is optionally generated by script generation component 312, which can be provided to a user. In a preferred implementation, if the user explicitly or implicitly authorizes, the script is automatically executed by script execution component 388.” The script is an execution template that encodes the database commands implementing the recommendation.)
Regarding Claim 5, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein performing the one or more automated actions comprises automatically initiating at least a portion of the one or more determined partitioning actions in connection with the at least one database. (Abstract of Levin states “One or more recommendations are then issued for correcting the root issues hindering performance. Preferably, for one or more recommendations, scripts are generated which are then executed manually or automatically.” Pg. 143 of Hilprecht states “In this paper, we introduce a new learned partitioning advisor based on Deep Reinforcement Learning (DRL) for OLAP-style workloads. The main idea is that a DRL agent learns the cost tradeoffs of different partitioning schemes and can thus automate the partitioning decision.” Levin shows the idea of automated actions automatically initiation some actions in the database. The DRL agent decides which partitioning scheme to apply for the workload. It would have been obvious to have the system automatically initiate some partitioning actions chosen by the DRL advisor via auto executed scripts.)
Regarding Claim 6, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein performing the one or more automated actions comprises automatically training at least a portion of the reinforcement learning model based at least in part on feedback related to the one or more determined partitioning actions. (Pg. 145 of Hilprecht states “In general, DRL agents learn by interacting with an environment by choosing actions and observing rewards which they seek to maximize. In our setup, the environment is the DBMS which the agent manipulates with actions that change the partitioning of individual tables. During the training phase, the agent learns to minimize the runtime of a given workload consisting of a mix of representative queries. In the training phase, the agents thus learns the effects of different partitionings on individual query latencies... We therefore separate the training process into two phases: (1) offline and (2) online training. In the offline training phase, the agent solely interacts with a “simulation” of the customer database…In an optional online training phase, the agent then does not just interact with a simulation but with a real database.” Hilprecht teaches automatically training a DRL agent that changes the partitioning of tables and learns by interacting with an environment by choosing actions and observing rewards, where the environment is the DBMS and the actions change the partitioning. During the training phase, agent also learns to minimize the runtime of a given workload based on feedback related to the partitioning actions. )
Regarding Claim 7, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein determining the one or more partitioning actions comprises processing at least the portion of the activity data using at least one deep learning neural network model. (Pg. 146 of Hilprecht states “There are different ways of realizing the Q-function. For Deep Q-learning [23] (or Deep Reinforcement Learning), a neural network Qθ (s, a) with weights θ is used for the approximation.”)
Regarding Claim 9, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
wherein processing activity data related to the at least one database comprises processing data pertaining to at least one of one or more database log files, one or more query execution patterns, query cost information, and one or more database health parameters. ([0149] of Levin states “The activity data can be retrieved by using one or more performance counters, trace data, log files etc.”)
Regarding Claim 10, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
training at least a portion of the reinforcement learning model using one or more of table name information, data pertaining to one or more attributes used for partitioning, table replication information, schema information, workload information, query information, query frequency information, partition action type information, and runtime information for one or more workloads. (Pg. 146 of Hilprecht states “we can encode the state as a binary vector using an one-hot encoding s(Ti) = ri, ai1, ai2, . . . , ain, where ri encodes whether a table is replicated and the remaining bits indicate whether an attribute is used for partitioning… To explicitly encode co-partitioning we introduce the concept of edges; i.e., if an edge between a pair of join attributes air and ajs of the corresponding tables Ti and Tj is activated, it guarantees co-partitioning. For instance, since the edge e1 in Figure 2b is active the customer and lineorder tables are co-partitioned. The fixed set of possible edges E can easily be extracted from the given schema and workload (i.e., all possible join paths).” Pg. 154 of Hilprecht states “we represent the workload as frequencies of a representative set of queries. Once trained, our learned partitioning advisor can suggest partitionings for any of those workloads… We then retrained the advisor for the additional queries and calculated, with the help of already measured runtimes, how long such an additional training takes on average if part of the workload is not known initially… In addition, incremental training can also make use of the Query Runtime Cache”)
Claims 11, 13 – 15 recite substantially similar subject matter as claims 1, 4 – 6 respectively, and are rejected with the same rationale, mutatis mutandis.
Claims 16, 18 – 20 recite substantially similar subject matter as claims 1, 4 – 6 respectively, and are rejected with the same rationale, mutatis mutandis.
Claims 21 – 22 recite substantially similar subject matter as claims 3, 7 respectively, and are rejected with the same rationale, mutatis mutandis.
Claims 8, 23 are rejected under 35 U.S.C. 103 as being unpatentable over Hilprecht et al. (NPL:” Learning a Partitioning Advisor for Cloud Databases”) in view of Levin (U.S. Pub. 2009/0240711 A1), further in view of Guo et al. (NPL:”Learning to Branch for Multi-Task Learning”)
Regarding Claim 8, the rejection of claim 7 is incorporated herein. Furthermore, the combination of Hilprecht and Levin teaches
implementing … the at least one deep learning neural network model trained to recommend the one or more partitioning actions and (Pg. 145 of Hilprecht states “Once the training is completed, we finally use the agent to make actual partitioning decisions. As input, it requires a workload, i.e. which queries were submitted in a certain time window. Based on this workload, the agent suggests partitionings which we deploy on the actual customer database.”)
implementing … the at least one deep learning neural network model trained to determine at least one execution template corresponding to at least a portion of the one or more recommended partitioning actions. (Within the apparatus, the recommendation components can further comprise a script generation component for generating a script for resolving the situations associated with the analysis rules, or a script execution component for executing a script for resolving the situation associated with the analysis rules.)
The combination of Hilprecht and Levin does not explicitly teach that
… multiple branches of the at least one deep learning neural network model…
However, Guo teaches that
… multiple branches of the at least one deep learning neural network model… (Pg. 1 Abstract section of Guo states “Training multiple tasks jointly in one deep network yields reduced latency during inference and better performance over the single-task counterpart by sharing certain layers of a network… In this work, we present an automated multi-task learning algorithm that learns where to share or branch within a network, designing an effective network topology that is directly optimized for multiple objectives across tasks.” Pg. 2 2 Related work section of Guo states “In the hard sharing setting, all tasks share the same set of backbone parameters, or at least share part of the backbone with branches toward the outputs… Meta Multi-Task Learning (Ruder et al., 2019) uses a shared input layer and two task specific output layers. Instead of choosing between soft sharing or hard sharing approach, a new effort in tackling the multi-task learning problem is to consider the dynamics between different losses across tasks.”)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings of Levin, Hilprecht, and Guo. Hilprecht teaches determining recommended partitioning actions for a database and providing corresponding implementation information, such as an ALTER TABLE statement for the optimized partitioning. Levin teaches a database performance analysis and recommendation system that collects activity data, identify performance issues, and recommendation components with script generation and execution to automatically execute scripts that implements recommended actions. Guo teaches using shared network layers with task specific output branches for related tasks. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Guo into the system of Levin and Hilprecht so that shared database and workload features are learned together, while separate branches are trained for the related tasks of recommending partitioning actions and determining corresponding execution template information. It would have been predictable use of known multi-task branching architecture to perform Hilprecht’s related database partitioning tasks.
Claim 23 recites substantially similar subject matter as claim 8 respectively, and is rejected with the same rationale, mutatis mutandis.
Relevant Prior Art Directed to State of Art
Zubaer et al. (“Understanding and Improving Deep Reinforcement Learning for Data Partitioning”) cited describes using deep reinforcement learning to automatically select database partitioning schemes based on workload performance. The reference is not relied upon in the present rejections.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BYUNGKWON HAN whose telephone number is (571)272-5294. The examiner can normally be reached M-F: 9:00AM-6PM PST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Li B Zhen can be reached at (571)272-3768. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/BYUNGKWON HAN/ Examiner, Art Unit 2121
/Li B. Zhen/ Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights