Abstract: A computer system and associated methods are disclosed for mitigating side-channel attacks using a shared cache. The computer system includes a host having a main memory and a shared cache. The host executes a virtual machine manager (VMM) that determines respective security keys for a plurality of co-located virtual machines (VMs). A cache controller for the shared cache includes a scrambling function that scrambles addresses of memory accesses performed by threads of the VMs according to the respective security keys. Different cache tiers may implement different scrambling functions optimized to the architecture of each cache tier. Security keys may be periodically updated to further reduce predictability of shared cache to memory address mappings.

Abstract: Methods, systems, and computer-readable media for global column indexing in a graph database are disclosed. A plurality of data elements of a graph database are stored. The triples comprise identifiers, column names, and values. The column names are globally scoped in the graph database and are associated with data types. Indices corresponding to the column names are created. A particular one of the indices comprises one or more of the values associated with the corresponding column name. A query is performed on the graph database using one or more of the indices corresponding to one of more of the column names associated with the query.

Abstract: Voice-controlled devices that include one or more speakers for outputting audio. In some instances, the device includes at least one speaker within a cylindrical housing, with the speaker aimed or pointed away from a microphone coupled to the housing. For instance, if the microphone resides at or near the top of the cylindrical housing, then the speaker may point downwards along the longitudinal axis of the housing and away from the microphone. By pointing the speaker away from the microphone, the microphone will receive less sound from the speaker than if the speaker were pointed toward the microphone). Because the voice-controlled device may perform speech recognition on audio signals generated by the microphone, less sound from the speaker represented in the audio signal may result in more accurate speech recognition, and/or a lesser need to perform acoustic echo cancelation (AEC) on the generated audio signals.

Abstract: Techniques are generally described for machine learning exampled-based annotation of image data. In some examples, a first machine learning model may receive a query image comprising a first depiction of an object-of-interest. In some examples, the first machine learning model may receive a target image representing a scene in which a second depiction of the object-of-interest is visually represented. In various examples, the first machine learning model may generate annotated output image data that identifies a location of the second depiction of the object-of-interest within the target image. In some examples, an object detection model may be trained based at least in part on the annotated output image data.

Abstract: Techniques for using suppression filters for presenting content of a network documents are described. In an example, a computer system receives, from a device associated with an account identifier, a request for an online network document. The computer system determines that a first category identifier of a first item category is associated with a suppression filter. The computer system determines that the account identifier is associated with an online item interaction having an interaction time associated with an item that belongs to the first item category. The computer system determines, based on a comparison of the interaction time and the suppression time, that a presentation of the first content in the online network document is prohibited. The computer system sends, to the device in response to the request, network document data that indicates second content about a third item that belongs to a second item category.

Abstract: Techniques for training a visual neural network are described. In particular, the training of the visual neural network may be performed using one or more contrastive learning loss functions including one or more of a visual-to-visual contrastive learning function using the visual neural network on positive and negative video clips according to a first loss function, a secondary-to-secondary contrastive learning a secondary neural network on secondary positive and negative information generated from the positive and negative video clips, and a secondary-to-visual contrastive learning according to a third loss function using the visual neural network on positive and negative video clips and using the secondary neural network secondary positive and negative information generated from the positive and negative video clips.

Abstract: Systems and methods are described controlling ingestion of data items within a data stream by executions of a serverless function on a serverless compute system. A poller device can act as an intermediary between the data stream and the serverless function, iteratively retrieving data items from the data stream and passing them in invocations of the serverless function. To allow for fine-grained control of ingestion without requiring implementation of complex logic at the poller device, the poller device can enable the serverless function to pass instructions controlling subsequent operation of the poller device. Each execution of the serverless function may determine whether subsequent operation of the poller device should be altered, and if so, instruct the poller device accordingly. The poller device can then modify its operation pursuant to the instructions, enabling highly granular control of streaming data ingestion without inhibiting existing benefits of serverless computing.

Abstract: Various embodiments of systems and methods for providing virtualized (e.g., serverless) broker clusters for a data streaming service are disclosed. A data streaming service uses a front-end proxy layer and a back-end broker layer to provide virtualized broker clusters, for example in a Kafka-based streaming service. Resources included in a virtualized broker cluster are monitored and automatically scaled-up, scaled-down, or re-balanced in a way that is transparent to data producing and/or data consuming clients of the data streaming service.

Abstract: Disclosed are various embodiments for inferring brand similarities using graph neural networks and selection prediction. In one embodiment, a brand-to-brand graph is generated indicating similarities between a set of brands according to at least one of: click-through data or conversion data. Using a first graph neural network (GNN) tower, the brand-to-brand graph is analyzed to determine brand similarities among a first brand identified from a search query and a first set of other brands. Using a second GNN tower, the brand-to-brand graph is analyzed to determine brand similarities among a second brand and a second set of other brands. A level of similarity between the first brand and the second brand is determined based at least in part on an output of the first GNN tower and an output of the second GNN tower.

Abstract: A clock disciplining scheme uses a pulse per second (PPS) signal that is distributed throughout a network to coordinate timing. In determining the time, jitter can occur due to latency between detection of the PPS signal and a software interrupt generated there from. This jitter affects the accuracy of the clock disciplining process. To eliminate the jitter, extra hardware is used to capture when the PPS signal occurred relative to a hardware clock counter associated with the clock disciplining software. In one embodiment, the extra hardware can be a sampling logic, which captures a state of a hardware clock counter upon PPS detection. In another embodiment, the extra hardware can initiate a counter that calculates a delay by the clock disciplining software in reading the hardware clock counter. The disciplining software can then subtract the calculated delay from a hardware clock counter to obtain the original PPS signal.

Abstract: An optical transceiver includes a built-in optical switch to switch between diverse fiber paths in switches in a datacenter or in switches between two datacenters. The built-in optical switch can be used to switch between racks in a datacenter to increase capacity for any rack that requests it. A controller, which receives a signal from a server computer in one of the racks, can be external to the optical transceiver or within the optical transceiver. In either case, the server computer can be provided with additional bandwidth when needed. For connections between datacenters, the built-in optical switch allows for optical line protection, but without the need for a splitter circuit, which incurs a significant power loss and requires a more expensive transceiver. Consequently, the built-in optical switch within an optical transceiver can be used in a variety of contexts to increase efficiency and reduce overall costs for network devices.

Abstract: Systems and methods for sender profile and/or recipient profile disambiguation and/or confirmation are disclosed. In instances where a sender profile is not indicated by a user sending a communication from a communal device, heuristic data may be utilized to infer the sender profile. Similar heuristic data may also be used when selection of the sender profile is associated with a low confidence level. Heuristic data may also be used to infer the recipient profile when the user does not indicate the recipient profile or when selection of the recipient profile is associated with a low confidence. Various confirmations may result from the sender and recipient profile disambiguation.

Abstract: The present disclosure generally relates to a system of a delivery device for combining sensor data from various types of sensors to generate a map that enables the delivery device to navigate from a first location to a second location to deliver an item to the second location. The system obtains data from RGB, LIDAR, and depth sensors and combines this sensor data according to various algorithms to detect objects in an environment of the delivery device, generate point cloud and pose information associated with the detected objects, and generates object boundary data for the detected objects. The system further identifies object states for the detected object and generates the map for the environment based on the detected object, the generated object proposal data, the labeled point cloud data, and the object states. The generated map may be provided to other systems to navigate the delivery device.

Abstract: Systems and methods are provided for implementing a multi-service file system for a hosted computing instance via a locally-addressable secure compute layer. Software within the instance can submit file operations to the secure compute layer, which the secure compute layer can translate into calls to one or more network-accessible storage services. To provide a multi-service file system, the secure compute layer can obtain mapping data mapping file system objects within the virtualized file system to different network-accessible storage services. On receiving a file operation, the secure compute layer can determine one or more network-accessible storage services corresponding to the file operation, and submit appropriate calls to the one or more network-accessible storage services.

Abstract: Techniques for performing machine learning inference calls in database query processing are described. A method for performing machine learning inference calls in database query processing may include generating a query plan to optimize a query for batch processing of data stored in a database service, the query plan including a batch mode operator to execute a function reference and an execution context associated with the batch mode operator, executing the query plan to invoke a function associated with the function reference, wherein the function sends a batch of requests, generated using the execution context, to a remote service and obtains a plurality of responses from the remote service, and generating a query response based on the plurality of responses.

Abstract: A data service implements a configurable data compressor/decompressor using a recipe generated for a particular data set type and using compression operators of a common registry (e.g., pantry) that are referenced by the recipe, wherein the recipe indicates at which nodes of a compression graph respective ones of the compression operators of the registry are to be implemented. The configurable data compressor/decompressor provides a customizable framework for compressing data sets of different types (e.g., belonging to different data domains) using a common compressor/decompressor implemented using a common set of compression operators.

Abstract: Techniques for change data capture (CDC) log augmentation are described. In some examples, a user configures CDC log augmentation by indicating which data should be included in a CDC log, and the database, when generating a CDC log associated with this configuration, can obtain the associated data and augment the CDC log by inserting this data into it. The augmented data can include one or more fields from a record in a separate database table, where the record can be identified based on the changed record represented by the CDC log.