Abstract: A computing resource service provider deploys resources to process input data sets on an ongoing basis and provide requestors with queryable data structures generated from the input data sets over determined, rolling periods of time. In one embodiment, the input data sets are processed using one or more nearest neighbor search algorithms, and the outputs therefrom are represented in data structures which are rotated as newer data structures are subsequently generated. The disclosed systems and techniques improve resource utilization, processing efficiency, query latency, and result consistency relative to known controls for large and/or complex data processing tasks, such as those employed in machine learning techniques.

Abstract: A replica database may receive a replication instruction from a master. The replica may maintain a version of a collection of data in database pages stored on a storage device. A replication thread may process the instruction. The replication thread may determine that an item affected by the instruction is located in a page that is both loaded into a memory buffer and subject to contention. The page may be modified with information indicating that the instruction has not been processed. A subsequent reader thread may, while processing a request to access the page, apply the instruction and complete processing of the request.

Abstract: A determination is made as to whether a value of a first parameter of a first application is to be obtained using a natural language interaction. Based on received input, a first service of a plurality of services is identified. The first service is to be used to perform a first task associated with the first parameter. Portions of the first application to determine the value of the first parameter and to invoke the first service are generated.

Abstract: Updates performed as part of transaction requests to a data store may be included in a stream of updates. Updates to items in the data store that are included in transactions determined not to be committed to the data store may be excluded from the stream of updates. Records in the stream of updates may include an identifier for the transaction that included the update described by the record. The identifier for the transaction may be used to identify updates to other items in the data store that are included in the same transaction.

Abstract: A system that implements a scalable data storage service may maintain tables in a non-relational data store on behalf of clients. The system may provide a Web services interface through which service requests are received, and an API usable to request that a table be created, deleted, or described; that an item be stored, retrieved, deleted, or its attributes modified; or that a table be queried (or scanned) with filtered items and/or their attributes returned. An asynchronous workflow may be invoked to create or delete a table. Items stored in tables may be partitioned and indexed using a simple or composite primary key. The system may not impose pre-defined limits on table size, and may employ a flexible schema. The service may provide a best-effort or committed throughput model. The system may automatically scale and/or re-partition tables in response to detecting workload changes, node failures, or other conditions or anomalies.

Abstract: Methods, systems, and computer-readable media for automated problem detection for machine learning models are disclosed. A machine learning analysis system receives data associated with use of a machine learning model. The data was collected by a machine learning inference system and comprises input to the model or a plurality of inferences representing output of the machine learning model. The machine learning analysis system performs analysis of the data associated with the use of the machine learning model. The machine learning analysis system detects one or more problems associated with the use of the machine learning model based at least in part on the analysis. The machine learning analysis system initiates one or more remedial actions associated with the one or more problems associated with the use of the machine learning model.

Abstract: Techniques for automated machine learning (ML) pipeline exploration and deployment are described. An automated ML pipeline generation system allows users to easily construct optimized ML pipelines by providing a dataset, identifying a target column in the dataset, and providing an exploration budget. Multiple candidate ML pipelines can be identified and evaluated through an exploration process, and a best ML pipeline can be provided to the requesting user or deployed for production inference. Users can configure, monitor, and adapt the exploration at multiple points in time throughout.

Abstract: Methods, systems, and computer-readable media for debugging and profiling of machine learning model training are disclosed. A machine learning analysis system receives data associated with training of a machine learning model. The data was collected by a machine learning training cluster. The machine learning analysis system performs analysis of the data associated with the training of the machine learning model. The machine learning analysis system detects one or more conditions associated with the training of the machine learning model based at least in part on the analysis. The machine learning analysis system generates one or more alarms describing the one or more conditions associated with the training of the machine learning model.

Abstract: Systems and methods are provided to implement a database system configured to return the state of a data item upon failure of a condition check on the data item during a conditional write. In embodiments, a write request may specify an indicator to return the item state upon condition failure. The request may specify multiple database operations to execute as a transaction, where failure of a single condition check will cause the entire transaction to fail and an item state causing the failure to be returned. The returned state of the data item may include a selection of the item's attributes specified by the request. Advantageously, these features allow a client to easily obtain the precise cause of a write's failure. Moreover, because the item state is returned only when a conditional write fails and only when requested, the response size of most write requests remains unchanged.

Abstract: Techniques for providing an on-demand development environment are described. A service of a provider network receives a request to launch a development environment, such as a notebook, from an electronic device. The service obtains an identification of a computing cluster hosted by the provider network. The service obtains an identification of a compute instance hosted within the provider network, the compute instance executing a software environment to host one or more development environments. The service causes the compute instance to launch a development environment. The service sends a message to configure the launched development environment to execute a computer program written in the development environment using the computing cluster. The service generates a token to secure communications between the electronic device and the development environment and sends the token to an originator of the request.

Abstract: A computing resource service provider deploys resources to process input data sets on an ongoing basis and provide requestors with queryable data structures generated from the input data sets over determined, rolling periods of time. In one embodiment, the input data sets are processed using one or more nearest neighbor search algorithms, and the outputs therefrom are represented in data structures which are rotated as newer data structures are subsequently generated. The disclosed systems and techniques improve resource utilization, processing efficiency, query latency, and result consistency relative to known controls for large and/or complex data processing tasks, such as those employed in machine learning techniques.

Abstract: A data store may implement on-demand data schema modifications. Request to change a schema for a data set in a data store may be received and a description of the change recorded as part of a schema history for the data set. The request to change the schema may then be acknowledged. When access requests directed to the data set are received at the data store, the schema history for the data set may be evaluated. If the schema history indicates that data that is to be accessed in order to service the data store needs to include one or more changes in the schema history, then a version of the schema that includes the one or more changes may be applied to the data.

Abstract: A determination is made as to whether a value of a first parameter of a first application is to be obtained using a natural language interaction. Based on received input, a first service of a plurality of services is identified. The first service is to be used to perform a first task associated with the first parameter. Portions of the first application to determine the value of the first parameter and to invoke the first service are generated.

Abstract: A system that implements a scaleable data storage service may maintain tables in a data store on behalf of storage service clients. The service may maintain data in partitions stored on respective computing nodes in the system. The service may support multiple throughput models, including a committed throughput model and a best effort throughput model. A service request to create a table may specify that requests directed to the table should be serviced under a committed throughput model and may specify the committed throughput level in terms of logical service request units. The service may reserve low-latency storage and other resources sufficient to meet the specified committed throughput level. A client/user may request a modification to the committed throughput level in anticipation of workload changes, such as an increase or decrease in traffic or data volume. In response, the system may increase or decrease the resources reserved for the table.

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: A system that implements a scalable data storage service may maintain tables in a data store on behalf of storage service clients. The service may maintain table data in multiple replicas of partitions that are stored on respective computing nodes in the system. In response to detecting an anomaly in the system, detecting a change in data volume on a partition or service request traffic directed to a partition, or receiving a service request from a client to split a partition, the data storage service may create additional copies of a partition replica using a physical copy mechanism. The data storage service may issue a split command defined in an API for the data store to divide the original and additional replicas into multiple replica groups, and to configure each replica group to maintain a respective portion of the table data that was stored in the partition before the split.

Abstract: Techniques for using scoring algorithms utilizing containers for flexible machine learning inference are described. In some embodiments, a request to host a machine learning (ML) model within a service provider network on behalf of a user is received, the request identifying an endpoint to perform scoring using the ML model. An endpoint is initialized as a container running on a virtual machine based on a container image and used to score data and return a result of said scoring to a user device.

Abstract: Techniques for streaming real-time automated speech recognition (ASR) are described. A user can stream audio data to a frontend service of the ASR service. The frontend service can establish a bi-directional connection to an audio decoder host to perform ASR on the data stream. The audio decoder host may include a streaming ASR engine which can analyze chunks of the audio data stream using an acoustic model to divide the audio data into words, and a language model to identify sentences made of the words spoken in the audio file. The acoustic model can be trained using short audio sentence data (e.g., on the order of 30 seconds to a few minutes), enabling the transcription service to accurately transcribe short chunks of audio data. The results are then punctuated and normalized. The resulting transcript is then streamed back to the user over the bi-directional connection.

Abstract: A system that implements a scalable data storage service may maintain tables in a non-relational data store on behalf of clients. The system may provide a Web services interface through which service requests are received, and an API usable to request that a table be created, deleted, or described; that an item be stored, retrieved, deleted, or its attributes modified; or that a table be queried (or scanned) with filtered items and/or their attributes returned. An asynchronous workflow may be invoked to create or delete a table. Items stored in tables may be partitioned and indexed using a simple or composite primary key. The system may not impose pre-defined limits on table size, and may employ a flexible schema. The service may provide a best-effort or committed throughput model. The system may automatically scale and/or re-partition tables in response to detecting workload changes, node failures, or other conditions or anomalies.

Abstract: A pricing policy to be applied to token population changes at a token bucket used for admission control during burst-mode operations at a work target is determined. Over a time period, changes to the token population of that bucket are recorded. An amount to be charged to a client is determined, based on the recorded changes in token population and an associated pricing amount indicated in the policy.