Abstract: Techniques for resolving a label are described. For example, a method of generating one or more labels corresponding to a substring of a tokenized utterance, each of the labels including one or more slot types, resolving the one or more generated labels into machine-readable values by: generating a tree of the generated one or more labels, the graph including a plurality of leaf nodes, wherein each rule of associated with a leaf node references a resolver, and resolving the generated tree from the leaf nodes to a root of the generated tree, by at each node, calling the referenced resolver and passing a result from each resolver to a subsequent resolver of the tree; and outputting a result of the resolving to an entity as dictated by the request is described.

Abstract: Techniques for dynamic accuracy-based experimentation and deployment of machine learning (ML) models are described. Inference traffic flowing to ML models and the accuracy of the models is analyzed and used to ensure that better performing models are executed more often via model selection. A predictive component can evaluate which model is more likely to be accurate for certain input data elements. Ensemble techniques can combine inference results of multiple ML models to aim to achieve a better overall result than any individual model could on its own.

Abstract: Localized lookups for performing access requests to a database may be implemented. Mapping information for storage nodes of a network-based service storing different data for different databases may be obtained by a routing application co-hosted with a client application of the database at a same container host. Access requests from the client application are handled by the routing application and sent to storage nodes identified using the mapping information. An authorization token may be included along with the requests to verify authorization to perform the access request at the storage nodes.

Abstract: A processing device receives a dataset comprising a plurality of data points, wherein each data point of the plurality of data points comprises a representative vector for the data point and an associated classification for the data point. The processing device determines, for the dataset, a score representative of a degree of clustering of the plurality of data points. The processing device determines a suitability of the dataset for use in machine learning based on the score.

Abstract: Systems, methods, and computer storage media are disclosed. A computer storage medium includes instructions that, when executed, perform actions including maintaining a first view based on a first table of a database. The actions include performing a first transaction within the database. The actions include performing a second transaction within the database, and the second transaction relies on data from the first view. Performing the second transaction includes beginning the second transaction. In response to occurrence of both (i) the first transaction committing a first modification to the first table subsequent to the beginning of the second transaction and (ii) the second transaction causing a modification to the first view based on the first modification, performing the second transaction includes continuing the second transaction based on the data from the first view excluding the modification to the first view and completing the second transaction.

Abstract: A distributed data store may implement passive distribution encryption keys to enable access to encrypted data stored in the distributed data store. Keys to encrypt a data volume stored in the distributed data store may be encrypted according to a distribution key and provided to a client of the distributed data store. Storage nodes that maintain portions of the data volume may receive the encrypted key from a client to enable access to the data volume. The storage nodes may decrypt the key according to the distribution key and enable access to the data volume at the storage nodes. In some embodiments, a key hierarchy may be implemented to encrypt the keys that provide access to the encrypted data. The key hierarchy may include a user key.

Abstract: A search management system and method to perform a search of first set of records maintained in a transactional data store (e.g., a commit log of a relational data store) and a second set of records maintained in a batch-updated data store. The search results corresponding to the transactional data store and the batch-updated data store are merged to generate a search result dataset to provide in response to a search query. The transactional layer or transaction data store is a commit log configured to manage and store records in accordance with recently executed operations (e.g., add and delete record operations) based on communications with one or more customer systems relating to customer data. Records maintained in the commit log are subsequently stored in the batch-updated data store as a result of a batching process.

Abstract: A machine learning model inference routing system in a machine learning service is described herein. The machine learning model inference routing system includes load balancer(s), network traffic router(s), an endpoint registry, and a feedback processing system that collectively allow the machine learning model inference routing system to adjust the routing of inferences based on machine learning model accuracy, demand, and/or the like. In addition, the arrangement of components in the machine learning model inference routing system enables the machine learning service to perform shadow testing, support ensemble machine learning models, and/or improve existing machine learning models using feedback data.

Abstract: A distributed data store may provide continuous backup for data stored in the distributed data store. Updates to data may be replicated amongst storage nodes according to a peer-to-peer replication scheme. A backup node may participate in the peer-to-peer replication scheme to identify additional updates to be applied to a backup version of the data in a separate data store. The backup node may obtain the updates according to the peer-to-peer replication scheme and update the backup version of the data. In some embodiments, configuration changes to the data in the distributed data store may be detected via the peer-to-peer replication scheme such that a backup node can adapt performance of backup operations in conformity with the configuration change.

Abstract: Techniques for auto-scaling hosted machine learning models for production inference are described. A machine learning model can be deployed in a hosted environment such that the infrastructure supporting the machine learning model scales dynamically with demand so that performance is not impacted. The model can be auto-scaled using reactive techniques or predictive techniques.

Abstract: Techniques are described for a nearest neighbor search service that enables users to perform nearest neighbor searches. The nearest neighbor search service includes an interface that enables users to create collections of searchable vectors, add and update vectors to a collection, delete vectors from a collection, and perform searches for nearest neighbors to a given vector. The nearest neighbor search service enables users to add, update, and delete vectors of a collection in real-time while also enabling users to perform searches at the same time.

Abstract: Methods and apparatus for token-based admission control for replicated writes are disclosed. Data objects are divided into partitions, and corresponding to each partition, at least a master replica and a slave replica are stored. A determination as to whether to accept a write request directed to the partition is made based at least in part on one or more of (a) available throughput capacity at the master replica, and (b) an indication, obtained using a token-based protocol, of available throughput capacity at the slave replica. If the write request is accepted, one or more data modification operations are initiated.

Abstract: A storage engine may implement dynamic splitting of contentious data pages. Data pages may store data for a table of a data store as part of an indexing structure for the table. Access to the table may be provided by locating the corresponding data pages via the indexing structure. Access contention for different data pages may be monitored. Data pages may be identified for splitting based on the monitoring. A split operation for an identified data page may be formed to store the data on the identified data page on two different data pages so that subsequent access requests for the data are divided between the two data pages. Monitoring of access contention may also be performed to identify data pages for merging in order to consolidate access requests to a single data page.

Abstract: A data set may be partitioned according to relative differences indicated by a cover tree. A cover tree may be generated for a data set. Items in the data set may be stored at the same or different nodes according to the relative difference between the items indicated in the cover tree. Portions of the cover tree may be assigned to different nodes storing the data set. Access requests for the data set may be performed by sending the access requests to nodes identified according to the assigned portions of the cover tree.

Abstract: A system and method for failover in a distributed system may comprise a computing device that receives client-provided information that groups computing nodes into ordered subsets. The subsets, or nodes in the subsets, may be associated with client-provided instructions for evaluating the health of a node. A node may be selected for failover based on executing the instructions and evaluating associated performance metrics. When a node is selected for failover, a replacement node may be selected based on the ordering of the subsets and the health of candidate nodes as determined based on executing the client-provided instructions.

Abstract: Systems, methods, and computer-readable media are disclosed for an improved database. The systems, methods, and computer-readable media described herein may enhance the response time of databases and improve user experiences. In an example method described herein, a database may receive a search request for data stored in the database. The database may be stored in association with a b-tree to facilitate searching and retrieving data from the database. The method may include allocating a portion of a first memory for read ahead operations and performing the read ahead operations, wherein the read ahead operations comprise loading, into the portion of the first memory, data associated with one or more siblings of a current node, wherein the one or more siblings of the current node are one or more nodes that have a same number of parent nodes as the current node.

Abstract: Techniques are described for a nearest neighbor search service that enables users to perform nearest neighbor searches. The nearest neighbor search service includes an interface that enables users to create collections of searchable vectors, add and update vectors to a collection, delete vectors from a collection, and perform searches for nearest neighbors to a given vector. The nearest neighbor search service enables users to add, update, and delete vectors of a collection in real-time while also enabling users to perform searches at the same time.

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: 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 are described for a nearest neighbor search service that enables users to perform nearest neighbor searches. The nearest neighbor search service includes an interface that enables users to create collections of searchable vectors, add and update vectors to a collection, delete vectors from a collection, and perform searches for nearest neighbors to a given vector. The nearest neighbor search service enables users to add, update, and delete vectors of a collection in real-time while also enabling users to perform searches at the same time.