Amazon Technologies, Inc.

Vaibhav Aggarwal

Peter Sirota

Richard Jeffrey Cole

Kathryn Marie Shih

Carl Louis Christofferson

Jai Vasanth

Andrew J Hitchcock

James P Bartlett

Eider Brantly Moore

Richard Rex McKnight

Lorenzo Minore

Methods and systems for cost-minimizing job scheduling are disclosed. A definition of a task is received. The definition comprises a need-by time. The need-by time comprises a deadline for completion of execution of the task. An estimated duration to complete the execution of the task is determined for each of a plurality of computing resources. One or more of the computing resources are selected based on an estimated cost of completing the execution using the computing resources. The execution of the task is initiated at a scheduled time using the selected one or more computing resources. The scheduled time is earlier than the need-by time by at least the estimated duration.

Cost-minimizing task scheduler

Methods and systems for optimization of task execution are disclosed. A definition of a task is received. A plurality of parameter values for execution of the task are selected based on an execution history for a plurality of prior tasks performed for a plurality of clients. The plurality of parameter values are selected to optimize one or more execution constraints for the execution of the task. The execution of the task is initiated using one or more computing resources configured with the selected parameter values.

Parameter selection for optimization of task execution based on execution history for prior tasks

Methods and systems for using a scheduler in a data pipeline are disclosed. A plurality of objects in a first layer are created, each representing a respective regularly scheduled task. A plurality of objects in a second layer are created, each representing a respective scheduled instance of a regularly scheduled task. It is determined whether each object in the second layer is ready to execute. For at least one object in the second layer, it is determined if the object has received notifications from any objects on which it depends. For each object that is ready to execute, the regularly scheduled task associated with the object is performed. For each object that is not ready to execute, the object is put to sleep.

Scheduler for data pipeline

Methods and systems for task timeouts as a function of input data size are disclosed. A definition of a task is received. The definition of the task indicates a set of input data for the task. A timeout duration for the task is determined based on the set of input data. The timeout duration varies with one or more characteristics of the set of input data. The execution of the task is initiated. The execution of the task is stopped if the execution of the task exceeds the timeout duration.

Task timeouts based on input data characteristics

Techniques are described for managing distributed execution of programs, including by dynamically scaling a cluster of multiple computing nodes performing ongoing distributed execution of a program, such as to increase and/or decrease computing node quantity. An architecture may be used that has core nodes that each participate in a distributed storage system for the distributed program execution, and that has one or more other auxiliary nodes that do not participate in the distributed storage system. Furthermore, as part of performing the dynamic scaling of a cluster, computing nodes that are only temporarily available may be selected and used, such as computing nodes that might be removed from the cluster during the ongoing program execution to be put to other uses and that may also be available for a different fee (e.g., a lower fee) than other computing nodes that are available throughout the ongoing use of the cluster.

Dynamic scaling of a cluster of computing nodes

Use of temporarily available computing nodes for dynamic scaling of a cluster

Methods and systems for a connector interface in a data pipeline are disclosed. A pipeline comprising two data source nodes and an activity node is configured. Each data source node represents data from a different data source, and the activity node represents a workflow activity that uses the data as input. Two connectors which implement the same connector interface are triggered. In response, data is acquired at each connector from the corresponding data source through the connector interface. The data is sent from the connectors to the activity node through the connector interface. The workflow activity is performed using the acquired data.

Connector interface for data pipeline

Techniques are described for managing distributed execution of programs, including by dynamically scaling a cluster of multiple computing nodes used to perform ongoing distributed execution of a program, such as to increase and/or decrease the quantity of computing nodes in the cluster at various times and for various reasons. An architecture may be used that facilitates the dynamic scaling of a cluster, including by having at least some of the computing nodes act as core nodes that each participate in a distributed storage system for the distributed program execution, and having one or more other computing nodes that act as auxiliary nodes that do not participate in the distributed storage system. If computing nodes are selected to be removed from the cluster during ongoing distributed execution of a program, one or more nodes of the auxiliary computing node type may be selected for the removal.

Growing community of inventors

Bellevue, WA, United States of America

Vaibhav Aggarwal