MACHINE LEARNING SYSTEM FLOW PROCESSING

• US 20160358103A1
• Filed: 06/05/2015
• Published: 12/08/2016
• Est. Priority Date: 06/05/2015
• Status: Active Grant

• Current Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Original Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Inventors: Farnham, Rodrigo Bouchardet, Paton, James Robert, Sidorov, Aleksandr, Vagata, Pamela Shen, Xie, Xiaowen, Bowers, Stuart Michael, Mehanna, Hussein Mohamed Hassan, Azzolini, Alisson Gusatti, Dunn, Jeffrey Scott
• Granted Patent: US 10,395,181 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 9/4881 Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/12 using genetic models

Abstract

Some embodiments include a method of machine learner workflow processing. For example, a workflow execution engine can receive an interdependency graph of operator instances for a workflow run. The operator instances can be associated with one or more operator types. The workflow execution engine can assign one or more computing environments from a candidate pool to execute the operator instances based on the interdependency graph. The workflow execution engine can generate a schedule plan of one or more execution requests associated with the operator instances. The workflow execution engine can distribute code packages associated the operator instances to the assigned computing environments. The workflow execution engine can maintain a memoization repository to cache one or more outputs of the operator instances upon completion of the execution requests.

Claims

1. A computer-implemented method, comprising:

initializing a workflow run in a machine learning system by identifying a text string defining a workflow;

traversing syntax of the text string to determine an interdependency graph of one or more data processing operator instances of the workflow;

generating an execution schedule of the workflow run based on the interdependency graph; and

managing execution of the workflow run in multiple computing environments according to the execution schedule; and

indexing an output of a data processing operator instance from among the data processing operator instances in a memoization repository, wherein the output is indexed as a result of processing an identifiable input through a data processing operator type associated with the data processing operator instance.

2. The computer-implemented method of claim 1, further comprising labeling the data processing operator instances in the interdependency graph as having one or more input types, wherein the input types include promised objects, constant objects, externally available objects, run-time initiated objects, or any combination thereof.

3. The computer-implemented method of claim 1, further comprising serializing the data processing operators into an operator database for access by at least one of the computing environments.

4. The computer-implemented method of claim 1, further comprising:

identifying an execution request to run the identifiable input through another instance of the data processing operator type; and

sending the indexed output from the memoization repository in response to the execution request.

5. The computer-implemented method of claim 4, wherein the execution request and the other instance are associated with another workflow run different from the initialized workflow run.

6. The computer-implemented method of claim 4, wherein the execution request and the other instance are associated with another workflow different from the workflow defined by the text string.

7. The computer-implemented method of claim 4, wherein indexing the output includes tracking a version number of the data processing operator type.

8. The computer-implemented method of claim 1, wherein said generating the execution schedule includes identifying one or more independent data processing operator instances in the interdependency graph to start the execution of the workflow run.

9. The computer-implemented method of claim 1, wherein generating the execution schedule includes scheduling execution of a dependent data processing operator instance in the interdependency graph to be triggered by completion of another data processing operator instance.

10. The computer-implemented method of claim 9, wherein the other data processing operator instance is an independent data processing operator instance in the interdependency graph or another dependent data processing operator instance.

11. The computer-implemented method of claim 9, further comprising:

receiving a resulting output from the other data processing operator instance; and

caching the resulting output in the memoization repository for access by a computing environment executing the dependent data processing operator instance.

12. A computer readable data memory storing computer-executable instructions that, when executed by a computer system, cause the computer system to perform a computer-implemented method, the instructions comprising:

instructions for accessing an interdependency graph of data processing operator instances for a workflow run, the data processing operator instances associated with one or more data processing operator types;

instructions for assigning one or more computing environments to execute the data processing operator instances based on the interdependency graph and resource constraints associated with the data processing operator instances;

instructions for generating a schedule plan of one or more execution requests associated with the data processing operator instances to the assigned computing environments; and

instructions for facilitating passing of an output from at least one of the data processing operator instances by caching the output in a memoization repository for access by another data processing operator instance.

13. The computer readable data memory of claim 12, wherein the resource constraints are defined as part of the workflow run or as respective attributes of the data processing operator types.

14. The computer readable data memory of claim 12, further comprising instructions for distributing one or more code packages associated with the data processing operator types to the assigned computing environments.

15. The computer readable data memory of claim 14, further comprising instructions for notifying at least one of the computing environments of an execution request associated with a target data processing operator instance in the interdependency graph.

16. The computer readable data memory of claim 15, further comprising:

instructions for receiving an execution schedule query from the at least one of the computing environments; and

instructions for identifying, in response to receiving the execution schedule query, a task-related object associated with the execution request for the at least one of the computing environment.

17. The computer readable data memory of claim 16, wherein the task related object is a serialized data processing operator type definition associated with the target data processing operator instance, an input dataset for the target data processing operator instance, an input parameter for the target data processing operator instance, or any combination thereof.

18. The computer readable data memory of claim 16, wherein said notifying includes identifying, to the at least one of the computing environments, where to retrieve an input dataset or an input parameter for the target data processing operator instance, where to store a resulting output of the target data processing operator instance, one or more identifiers of code or executable libraries required to execute the target processing operator instance, one or more identifiers of other data processing operator instances adjacent to the target data processing operator instance in the interdependency graph, or any combination thereof.

19. A machine learning system, comprising:

an operator repository configured to store one or more operator definitions;

a workflow repository configured to store a workflow definition; and

an execution scheduler engine configured to generate an interdependency graph defining a workflow run by matching input and output types of the operator definitions in the workflow definition, wherein the interdependency graph identifies one or more independent operator instances that are capable of parallel execution;

wherein the execution scheduler engine is configured to generate an execution schedule of the workflow run, wherein the workflow run is scheduled for execution on multiple computing environments, wherein the independent operator instances are scheduled to be executed in parallel; and

wherein the execution scheduler engine is configured to generate the execution schedule by maximizing parallel processing and memoization utilization.

20. The machine learning system of claim 19, wherein the execution scheduler engine is configured to generate the execution schedule by minimizing resource commitment while taking into account another workflow run scheduled to be executed by the machine learning system, the other workflow run sharing a computing environment candidate pool with the workflow run.