Abstract: Techniques for generation of synthetic queries from customer data for training of document querying machine learning (ML) models as a service are described. A service may receive one or more documents from a user, generate a set of question and answer pairs from the one or more documents from the user using a machine learning model trained to predict a question from an answer, and store the set of question and answer pairs generated from the one or more documents from the user. The question and answer pairs may be used to train another machine learning model, for example, a document ranking model, a passage ranking model, a question/answer model, or a frequently asked question (FAQ) model.

Abstract: Techniques for partitioning data flow operations between execution on a compute instance and an attached accelerator instance are described. A set of operations supported by the accelerator is obtained. A set of operations associated with the data flow is obtained. An operation in the set of operations associated with the data flow is identified based on the set of operations supported by the accelerator. The accelerator executes the first operation.

Abstract: A technology landscape may be characterized using a performance characterization that includes scores assigned to performance metrics for the technology landscape and using at least one trained machine learning model. In response to a detected calibration trigger, a calibratable performance metric of the performance metrics may be determined. A relationship may be determined between conforming values of the calibratable performance metric during a conforming period for which the at least one trained machine learning model was trained, and non-conforming values of the calibratable performance metric occurring during a calibration period initiated by the calibration trigger. In this way, a score assigned to the calibratable performance metric may be calibrated, based on the relationship.

Abstract: Techniques for making machine learning inference calls for database query processing are described. In some embodiments, a method of making machine learning inference calls for database query processing may include generating a first batch of machine learning requests based at least on a query to be performed on data stored in a database service, wherein the query identifies a machine learning service, sending the first batch of machine learning requests to an input buffer of an asynchronous request handler, the asynchronous request handler to generate a second batch of machine learning requests based on the first batch of machine learning requests, and obtaining a plurality of machine learning responses from an output buffer of the asynchronous request handler, the machine learning responses generated by the machine learning service using a machine learning model in response to receiving the second batch of machine learning requests.

Abstract: Implementations detailed herein include description of a computer-implemented method. In an implementation, the method at least includes provisioning an application instance and portions of at least one accelerator attached to the application instance to execute a machine learning model of an application of the application instance; loading the machine learning model onto the portions of the at least one accelerator; receiving scoring data in the application; and utilizing each of the portions of the attached at least one accelerator to perform inference on the scoring data in parallel and only using one response from the portions of the accelerator.

Abstract: When a restart event is detected within a technology landscape, restart-impacted performance metrics and non-restart-impacted performance metrics may be identified. The non-restart-impacted performance metrics may continue to be included within a performance characterization of the technology landscape. The restart-impacted performance metrics may be monitored, while being excluded from the performance characterization. The restart-impacted performance metric of the restart-impacted performance metrics may be transitioned to a non-restart-impacted performance metric, based on a monitored value of the restart-impacted performance metric following the restart event.

Abstract: Methods and systems enable a database reorganization to occur without a database outage. In one aspect, the method includes creating a shadow copy of a database, the shadow having at least one partition associated with a plurality of first data sets and the database having at least a first partition associated with a plurality of second data sets and a second partition associated with a plurality of third data sets. The method also includes reorganizing the at least one partition of the shadow, taking the first partition offline, replacing the plurality of second data sets with the plurality of first data sets in a schema, and restarting the first partition.

Abstract: Techniques for searching documents are described. An exemplary method includes receiving a document search query; querying at least one index based upon the document search query to identify matching data; fetching the identified matched data; determining one or more of a top ranked passage and top ranked documents from the set of documents based upon one or more invocations of one or more machine learning models based at least on the fetched identified matched data and the document search query; and returning one or more of the top ranked passage and the proper subset of documents.

Abstract: A data manager may include a data opaque interface configured to provide, to an arbitrarily selected page-oriented access method, interface access to page data storage that includes latch-free access to the page data storage. In another aspect, a swap operation may be initiated, of a portion of a first page in cache layer storage to a location in secondary storage, based on initiating a prepending of a partial swap delta record to a page state associated with the first page, the partial swap delta record including a main memory address indicating a storage location of a flush delta record that indicates a location in secondary storage of a missing part of the first page. In another aspect, a page manager may initiate a flush operation of a first page in cache layer storage to a location in secondary storage, based on atomic operations with flush delta records.

Abstract: Methods and systems for training a neural network are provided. In one example, an apparatus comprises a memory that stores instructions; and a hardware processor configured to execute the instructions to: control a neural network processor to perform a loss gradient operation to generate data gradients; after the loss gradient operation completes, control the neural network processor to perform a forward propagation operation to generate intermediate outputs; control the neural network processor to perform a backward propagation operation based on the data gradients and the intermediate outputs to generate weight gradients; receive the weight gradients from the neural network processor; and update weights of a neural network based on the weight gradients.

Abstract: Described is using flash memory (or other secondary storage), RAM-based data structures and mechanisms to access key-value pairs stored in the flash memory using only a low RAM space footprint. A mapping (e.g. hash) function maps key-value pairs to a slot in a RAM-based index. The slot includes a pointer that points to a bucket of records on flash memory that each had keys that mapped to the slot. The bucket of records is arranged as a linear-chained linked list, e.g., with pointers from the most-recently written record to the earliest written record. Also described are compacting non-contiguous records of a bucket onto a single flash page, and garbage collection. Still further described is load balancing to reduce variation in bucket sizes, using a bloom filter per slot to avoid unnecessary searching, and splitting a slot into sub-slots.

Abstract: Techniques for making machine learning inference calls for database query processing are described. In some embodiments, a method of making machine learning inference calls for database query processing may include generating a first batch of machine learning requests based at least on a query to be performed on data stored in a database service, wherein the query identifies a machine learning service, sending the first batch of machine learning requests to an input buffer of an asynchronous request handler, the asynchronous request handler to generate a second batch of machine learning requests based on the first batch of machine learning requests, and obtaining a plurality of machine learning responses from an output buffer of the asynchronous request handler, the machine learning responses generated by the machine learning service using a machine learning model in response to receiving the second batch of machine learning requests.

Abstract: Techniques for generation of synthetic queries from customer data for training of document querying machine learning (ML) models as a service are described. A service may receive one or more documents from a user, generate a set of question and answer pairs from the one or more documents from the user using a machine learning model trained to predict a question from an answer, and store the set of question and answer pairs generated from the one or more documents from the user. The question and answer pairs may be used to train another machine learning model, for example, a document ranking model, a passage ranking model, a question/answer model, or a frequently asked question (FAQ) model.

Abstract: Techniques for searching documents are described. An exemplary method includes receiving a document search query; querying at least one index based upon the document search query to identify matching data; fetching the identified matched data; determining one or more of a top ranked passage and top ranked documents from the set of documents based upon one or more invocations of one or more machine learning models based at least on the fetched identified matched data and the document search query; and returning one or more of the top ranked passage and the proper subset of documents.

Abstract: Implementations detailed herein include description of a computer-implemented method to migrate a machine learning model from one accelerator portion (such as a portion of a graphical processor unit (GPU)) to a different accelerator portion. In some instances, a state of the first accelerator portion is persisted, the second accelerator portion is configured, the first accelerator portion is then detached from a client application instance, and at least a portion of an inference request is performed using the loaded at least a portion of the machine learning model on the second accelerator portion that had been configured.

Abstract: Described is using flash memory (or other secondary storage), RAM-based data structures and mechanisms to access key-value pairs stored in the flash memory using only a low RAM space footprint. A mapping (e.g. hash) function maps key-value pairs to a slot in a RAM-based index. The slot includes a pointer that points to a bucket of records on flash memory that each had keys that mapped to the slot. The bucket of records is arranged as a linear-chained linked list, e.g., with pointers from the most-recently written record to the earliest written record. Also described are compacting non-contiguous records of a bucket onto a single flash page, and garbage collection. Still further described is load balancing to reduce variation in bucket sizes, using a bloom filter per slot to avoid unnecessary searching, and splitting a slot into sub-slots.

Abstract: Described is using flash memory (or other secondary storage), RAM-based data structures and mechanisms to access key-value pairs stored in the flash memory using only a low RAM space footprint. A mapping (e.g. hash) function maps key-value pairs to a slot in a RAM-based index. The slot includes a pointer that points to a bucket of records on flash memory that each had keys that mapped to the slot. The bucket of records is arranged as a linear-chained linked list, e.g., with pointers from the most-recently written record to the earliest written record. Also described are compacting non-contiguous records of a bucket onto a single flash page, and garbage collection. Still further described is load balancing to reduce variation in bucket sizes, using a bloom filter per slot to avoid unnecessary searching, and splitting a slot into sub-slots.

Abstract: Implementations detailed herein include description of a computer-implemented method.

Abstract: Implementations detailed herein include description of a computer-implemented method.

Abstract: Implementations detailed herein include description of a computer-implemented method. In an implementation, the method at least includes receiving an application instance configuration, an application of the application instance to utilize a portion of an attached accelerator during execution of a machine learning model and the application instance configuration including: an indication of the central processing unit (CPU) capability to be used, an arithmetic precision of the machine learning model to be used, an indication of the accelerator capability to be used, a storage location of the application, and an indication of an amount of random access memory to use.