Abstract: A system that implements a data storage service may store data for a database table in multiple replicated partitions on respective storage nodes. In response to a request to back up a table, the service may back up individual partitions of the table to a remote storage system independently and (in some cases) in parallel, and may update (or create) and store metadata about the table and its partitions on storage nodes of the data storage service and/or in the remote storage system. Backing up each partition may include exporting it from the database in which the table is stored, packaging and compressing the exported partition for upload, and uploading the exported, packaged, and compressed partition to the remote storage system. The remote storage system may be a key-value durable storage system in which each backed-up partition is accessible using its partition identifier as the key.

Abstract: Authorization is divided between a control plane and a data plane for sharing database data. A producer database engine can create a shared database via a data plane interface. A producer can then authorize access to the shared database via a control plane interface to a consumer. A consumer can associate the authorization granted to the consumer with a consumer database engine via the control plane interface.

Abstract: A serverless computing system is configured to provide access to a machine learning model by at least associating an endpoint, comprising code that accesses the machine learning model, with an extension that interfaces between a serverless compute architecture and the endpoint. A request to perform an inference is received by the system and processed by using the serverless compute architecture to execute a compute function. The compute function cases the extension to interface with the endpoint to cause the machine learning model to perform the inference.

Abstract: A database system may associate functions with a database table. A request to associate a function with a table in a database system may be received. An association between the table and the function may be created. The function may include parameters that are determined from values within the table which are then invoked by a request to perform the function. The associated function may cause the collection of the values prior to performance of the function.

Abstract: A system that provides services to clients may receive and service requests, various ones of which may require different amounts of work. The system may determine whether it is operating in an overloaded or underloaded state based on a current work throughput rate, a target work throughput rate, a maximum request rate, or an actual request rate, and may dynamically adjust the maximum request rate in response. For example, if the maximum request rate is being exceeded, the maximum request rate may be raised or lowered, dependent on the current work throughput rate. If the target or committed work throughput rate is being exceeded, but the maximum request rate is not being exceeded, a lower maximum request rate may be proposed. Adjustments to the maximum request rate may be made using multiple incremental adjustments. Service request tokens may be added to a leaky token bucket at the maximum request rate.

Abstract: A serverless computing system is configured to provide access to a machine learning model by at least associating an endpoint, comprising code that accesses the machine learning model, with an extension that interfaces between a serverless compute architecture and the endpoint. A request to perform an inference is received by the system and processed by using the serverless compute architecture to execute a compute function. The compute function cases the extension to interface with the endpoint to cause the machine learning model to perform the inference.

Abstract: A system is configured to provide access to a machine learning model by using a hybrid configuration comprising a dedicate server on which an instance of a model server is installed, and a serverless compute architecture that interfaces with an instance of the model server using an extension. A first portion of requests directed to the model server are processed by the dedicated server, and a second portion of the requests is processed by the serverless compute architecture.

Abstract: A system that provides services to clients may receive and service requests, various ones of which may require different amounts of work. The system may determine whether it is operating in an overloaded or underloaded state based on a current work throughput rate, a target work throughput rate, a maximum request rate, or an actual request rate, and may dynamically adjust the maximum request rate in response. For example, if the maximum request rate is being exceeded, the maximum request rate may be raised or lowered, dependent on the current work throughput rate. If the target or committed work throughput rate is being exceeded, but the maximum request rate is not being exceeded, a lower maximum request rate may be proposed. Adjustments to the maximum request rate may be made using multiple incremental adjustments. Service request tokens may be added to a leaky token bucket at the maximum request rate.

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: A system that implements a data storage service may store data for database tables in multiple replicated partitions on respective storage nodes. In response to a request to back up a table, the service may export individual partitions of the table from the database and package them to be independently uploaded (e.g., in parallel) to a remote storage system (e.g., a key-value durable storage system). Prior to uploading the exported and packaged partitions to the remote storage system, the service may verify that the exported and packaged partitions can be subsequently restored, which may include unpackaging and/or re-inflating the exported and packaged partitions to create additional unpackaged copies of the partitions, re-importing the additional unpackaged copies of the partitions into the database (e.g., as additional replicas), and/or comparing checksums generated for the exported partitions with checksums generated for the additional unpackaged copies of the partitions.

Abstract: A system that provides services to clients may receive and service requests, various ones of which may require different amounts of work. The system may determine whether it is operating in an overloaded or underloaded state based on a current work throughput rate, a target work throughput rate, a maximum request rate, or an actual request rate, and may dynamically adjust the maximum request rate in response. For example, if the maximum request rate is being exceeded, the maximum request rate may be raised or lowered, dependent on the current work throughput rate. If the target or committed work throughput rate is being exceeded, but the maximum request rate is not being exceeded, a lower maximum request rate may be proposed. Adjustments to the maximum request rate may be made using multiple incremental adjustments. Service request tokens may be added to a leaky token bucket at the maximum request rate.

Abstract: A distributed database system may implement dynamic quorum group membership changes. In various embodiments, a quorum set may maintain a replica of a data object among group members according to a protection group policy for the data object. A group member may be identified as to be replaced. In response, a new quorum set may be created from the remaining group members and a new group member. The protection group policy may be updated to include the new group members such that subsequently received updates are maintained at both the previous to quorum set and the new quorum set. Previously received updates may be replicated on the new group member. Upon completion of replicating the previously received updates, the protection group policy for the data object may be revised such that subsequently received updates are maintained at the new quorum set.

Abstract: A system that implements a data storage service may store data for database tables in multiple replicated partitions on respective storage nodes. In response to a request to restore a given table that was backed up in a remote storage system (e.g., key-value durable storage system), the service may create a new table, and may import a copy of each of the partitions of the given table from the remote storage system into the new table. The request to restore the table may specify a modified value for a configuration parameter for the table or for one of its partitions. The service may apply the new configuration parameter value to the table or its partitions during the restore operation. The new configuration parameter value may indicate an increase or decrease in storage capacity or throughput capacity, and its application may automatically trigger a partition split or move operation.

Abstract: A system that provides services to clients may receive and service requests, various ones of which may require different amounts of work. The system may determine whether it is operating in an overloaded or underloaded state based on a current work throughput rate, a target work throughput rate, a maximum request rate, or an actual request rate, and may dynamically adjust the maximum request rate in response. For example, if the maximum request rate is being exceeded, the maximum request rate may be raised or lowered, dependent on the current work throughput rate. If the target or committed work throughput rate is being exceeded, but the maximum request rate is not being exceeded, a lower maximum request rate may be proposed. Adjustments to the maximum request rate may be made using multiple incremental adjustments. Service request tokens may be added to a leaky token bucket at the maximum request rate.

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: A system that implements a data storage service may store data for a database table in multiple replicated partitions on respective storage nodes. In response to a request to back up a table, the service may back up individual partitions of the table to a remote storage system independently and (in some cases) in parallel, and may update (or create) and store metadata about the table and its partitions on storage nodes of the data storage service and/or in the remote storage system. Backing up each partition may include exporting it from the database in which the table is stored, packaging and compressing the exported partition for upload, and uploading the exported, packaged, and compressed partition to the remote storage system. The remote storage system may be a key-value durable storage system in which each backed-up partition is accessible using its partition identifier as the key.

Abstract: A distributed database management system may comprise a plurality of computing nodes. A request to update an item maintained by the system may be acknowledged as durable and committed once an entry corresponding to the request has been written to a log file and quorum among the computing nodes has been achieved. Improved consistency may be achieved by maintaining snapshots of committed item states within queryable in-memory snapshot data structures. Range queries may be performed by merging a secondary index with the snapshots and applying filters. Projections may be completed by retrieving additional data from an item collection maintain on one or more storage devices.

Abstract: Techniques are disclosed for splitting a database partition into two partitions. In embodiments, where the partition is a hash partition, the partition is split at its logical midpoint, so that half of the partition's rows are placed in each split partition. Where the partition is a hash-range partition, a number of rows of the partition are sampled. Where enough samples fall on each side of the logical midpoint, the logical midpoint is again used as the split point. Where not enough samples fall on one side of the logical midpoint, then the median of the samples is used as the split point.

Abstract: A system that provides services to clients may receive and service requests, various ones of which may require different amounts of work. The system may determine whether it is operating in an overloaded or underloaded state based on a current work throughput rate, a target work throughput rate, a maximum request rate, or an actual request rate, and may dynamically adjust the maximum request rate in response. For example, if the maximum request rate is being exceeded, the maximum request rate may be raised or lowered, dependent on the current work throughput rate. If the target or committed work throughput rate is being exceeded, but the maximum request rate is not being exceeded, a lower maximum request rate may be proposed. Adjustments to the maximum request rate may be made using multiple incremental adjustments. Service request tokens may be added to a leaky token bucket at the maximum request rate.

Abstract: A distributed database system may implement dynamic quorum group membership changes. In various embodiments, a quorum set may maintain a replica of a data object among group members according to a protection group policy for the data object. A group member may be identified as to be replaced. In response, a new quorum set may be created from the remaining group members and a new group member. The protection group policy may be updated to include the new group members such that subsequently received updates are maintained at both the previous to quorum set and the new quorum set. Previously received updates may be replicated on the new group member. Upon completion of replicating the previously received updates, the protection group policy for the data object may be revised such that subsequently received updates are maintained at the new quorum set.