Abstract: Systems and methods are described for allocating requests to implement new workloads within a dynamic set of servers. Existing load balancing techniques can result in “focus firing” on new servers added to the set, since a load balancer may view a new server as underloaded. With sufficient intensity, focus firing can result in overshooting target load for the new server, and the new server in fact becoming overloaded. The present disclosure modifies selection of servers as potential targets for a workload by at least partly biasing against selection of young servers. The bias imposed can be scaled to avoid overloading new servers.

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: Systems and methods are described for detecting and preventing execution of malware on an on-demand code execution system. An on-demand code execution system may execute user-submitted code on virtual machine instances, which may be provisioned with various computing resources (memory, storage, processors, network bandwidth, etc.). These resources may be utilized in varying amounts or at varying rates during execution of the user-submitted code. The user-submitted code may also be unavailable for inspection for security or other reasons. A malware detection system may thus identify user-submitted code that corresponds to malware by monitoring resource utilization during execution of the code and generating a resource utilization signature, which enables comparison between the signature of the user-submitted code and resource utilization signatures of codes previously identified as malware.

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: Systems and methods are described for enabling garbage collection on data storage systems. Traditional garbage collection often attempts to track use of data items on an individual level, deleting each item when it is no longer used. In distributed systems, tracking use on an individual level is difficult, and may require centralized knowledge across the system with respect to individual data items. Provided herein is a “coarse-grained” garbage collection mechanism, which divides objects into logical groups referred to as “roots.” Each root has a life cycle. While active, new data can be stored in a root. While inactive, use of data within a root can cause that date to be copied to a different, active root. When the system detects that data hasn't been used in an inactive root for a threshold period, the root can be considered “dead” and data within the root may be deleted.

Abstract: Systems and methods are described for providing for serverless inferences against a trained machine learning (ML) model. Rather than obtaining one or more dedicated devices to conduct inferences, users are enabled to create a task on a serverless system that, when invoked, passing input data to a trained ML model and provides a result. To satisfy varying user requirements for inference speed, the system includes a variety of hardware configurations. The system can efficiently allocate resources between different tasks by invoking the task on a particular hardware configuration that is selected based on a current availability of the selected hardware configuration to host an execution environment in which the task is implemented and an expected time to invoke the task on the individual hardware configuration. The system can therefore efficiently allocate resources among inferences using a variety of different ML models.

Abstract: In some cases, a localization service may identify candidate strings in the source code of an application. Further, the localization service may determine whether the candidate strings are displayed literals in a first human-perceivable language. In addition, the localization service may replace the identified displayed literals with identification tokens to generate pivot source code. In some examples, an identification token may include a JavaScript function that returns a translation of a displayed literal in a second human-perceivable language or any other desired human-perceivable language. Further, the localization service may verify pivot source code by comparing a localized application corresponding to the pivot source code to the application with the original source code of the application.