Mapreduce implementation in an on-demand network code execution system and stream data processing system

Abstract

Systems and methods are described for providing an implementation of the MapReduce programming model utilizing tasks executing on an on-demand code execution system, utilizing a stream data processing system as an intermediary between map and reduce function. A map task implementing a map function can process portions of a data set, to generate outputs associated with different values for a measured attribute of the data set. Executions of the map task can publish outputs to a data stream on the stream data processing system, which stream is configured to utilize the measured attribute as a partition key for the stream. Based on the partition key, the stream data processing system can divide the stream into sub-streams, each containing a relevant subset of the outputs. The on-demand code execution system can execute a reduce task to apply the reduce function to the outputs of each sub-stream, thereby completing the MapReduce process.

Description

BACKGROUND

Computing devices can utilize communication networks to exchange data. Companies and organizations operate computer networks that interconnect a number of computing devices to support operations or to provide services to third parties. The computing systems can be located in a single geographic location or located in multiple, distinct geographic locations (e.g., interconnected via private or public communication networks). Specifically, data centers or data processing centers, herein generally referred to as a “data center,” may include a number of interconnected computing systems to provide computing resources to users of the data center. The data centers may be private data centers operated on behalf of an organization or public data centers operated on behalf, or for the benefit of, the general public.

To facilitate increased utilization of data center resources, virtualization technologies allow a single physical computing device to host one or more instances of virtual machines that appear and operate as independent computing devices to users of a data center. With virtualization, the single physical computing device can create, maintain, delete, or otherwise manage virtual machines in a dynamic manner. In turn, users can request computer resources from a data center, including single computing devices or a configuration of networked computing devices, and be provided with varying numbers of virtual machine resources.

In some scenarios, virtual machine instances may be configured according to a number of virtual machine instance types to provide specific functionality. For example, various computing devices may be associated with different combinations of operating systems or operating system configurations, virtualized hardware resources and software applications to enable a computing device to provide different desired functionalities, or to provide similar functionalities more efficiently. These virtual machine instance type configurations are often contained within a device image, which includes static data containing the software (e.g., the OS and applications together with their configuration and data files, etc.) that the virtual machine will run once started. The device image is typically stored on the disk used to create or initialize the instance. Thus, a computing device may process the device image in order to implement the desired software configuration.

One example use of data centers is to process or analyze large data sets, which may be impractical to analyze using a single computing device. Various techniques have been developed to allow for multiple computing devices (or in some instances multiple processors within a single computing device) to process data concurrently. This concurrent data processing is sometimes referred to as “parallelization.” One technique for allowing parallelization in processing data sets is the “MapReduce” programming model. This programming model generally requires a centralized “infrastructure” or “framework,” which controls execution of two functions by individual computing devices within a set of devices. Execution of the first function, a “map” function, causes multiple devices to process portions (or “chunks”) of a full set of raw data to generate a set of intermediate results, such as counts of individual words within a corpus of text. Execution of the second function, a “reduce” function, causes one or more devices to combine multiple sets of intermediate results (from multiple map functions) to produce a set of aggregated results. The reduce functions may be executed multiple times, with each execution further reducing the number of aggregated results, until a single aggregate result record is created. In traditional implementations of the MapReduce programming model, an infrastructure or framework must typically execute continuously to coordinate execution of map and reduce functions until a result is provided.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a block diagram depicting an illustrative environment in which an on-demand code execution environment can operate to process data sets according to the MapReduce programming model, utilizing one or more tasks executing on the on-demand code execution environment and a stream data processing system as an intermediary data store between task executions;

FIG. 2 depicts a general architecture of a computing device providing a worker manager 140 of FIG. 1;

FIG. 3 is a flow diagram depicting illustrative interactions for initiating an implementation of the MapReduce programming model on the on-demand code execution environment of FIG. 1, including initiating multiple executions a map task;

FIG. 4A is a flow diagram depicting illustrative interactions of map and reduce task executions on the on-demand code execution environment of FIG. 1 to process data according to the MapReduce programming model;

FIG. 4B is an alternative view of the interactions of FIG. 4A, depicting a logical data flow during those interactions;

FIG. 5 is an illustrative routine for implementing a map function on an the on-demand code execution system of FIG. 1;

FIG. 6 is an illustrative routine for utilizing a stream data processing system of FIG. 1 to divide outputs of the map functions into sub-streams according to an attribute of the outputs relevant to a reduce function; and

FIG. 7 is an illustrative routine for reducing the messages in each sub-stream via a reduce function implemented on the on-demand code execution system of FIG. 1.

DETAILED DESCRIPTION

Generally described, aspects of the present disclosure relate to an on-demand code execution environment that enables clients to request execution of user-defined code in order to process data according to the MapReduce programming model. More specifically, aspects of the present disclosure enable an implementation of the MapReduce programming model on the on-demand code execution environment without the need for a dedicated framework or infrastructure to manage execution of map and reduce functions within the implementation. Instead, embodiments of the present disclosure may utilize a stream data processing system as an intermediary between executions of map and reduce functions, reducing or eliminating the need for other coordinating devices and minimizing redundant data processing or transmission during implementation of the MapReduce model.

The MapReduce programming model is a logical data processing model used to enable a distributed computing system (e.g., including multiple distinct devices in communication with one another) to process data sets, and particularly large data sets (e.g., on the order of gigabytes, terabytes, petabytes, etc.). Generally, the MapReduce programming model includes at least two data processing stages: a map stage and a reduce stage. During the map stage, individual devices within the distributed computing system collect a portion of the data set (e.g., chunks of n gigabytes) and “map” the content of the portion into a set of outputs. Each set of outputs is then processed by a “reduce” function to aggregate, combine, or otherwise determine a result of processing the content within the output set. A common example of the MapReduce paradigm is that of counting the occurrence of words within a corpus. For example, suppose that a user obtains a multi-terabyte textual data set, and would like to determine what words exist within the data set and how often each word occurs. Utilizing a single device, processing the data set may take excessive amounts of time. However, under the MapReduce model, each device within a distributed system may process a limited portion (or “chunk”) of the data set according to a map function. In this example, the map function may correspond to counting the occurrence of each word within the portion of the data set. By increasing the number of devices within the system, the speed of processing can be increased. The limitation of such processing is that each device would process only a limited portion of the data set, and thus be aware only of a count of words within that portion. To aggregate the counts across devices, a reduce function is used. While it may be possible to use a single device to aggregate the word counts generated at each map function, this configuration once again creates a bottleneck to processing. To address this, the MapReduce model enables multiple reduce functions to be implemented across the distributed system, each function processing a subset of the outputs of the map functions, according to an attribute of the outputs. For example, in the example of a word count, 26 reduce functions may be implemented, each creating a count of words starting with a particular character in the English alphabet. Illustratively, a first reduce function may obtain, from the map functions, all counts of words starting with the letter ‘a,’ a second reduce function may obtain all counts of words starting with ‘b,’ etc. Each reduce function can therefore gain an overall word count for a certain portion of the words in the corpus. The outputs of the reduce function can be combined to obtain the total word count. Often, combining the outputs of a reduce function is a relatively low-cost operation, as those outputs are distinct and non-overlapping. Thus, the MapReduce programming model can be utilized to facilitate rapid, distributed processing of data sets.

Traditionally, the MapReduce model has been implemented on dedicated distributed systems, configured with framework software that handles the generation and configuration of worker computer nodes within the system. Worker nodes are commonly dedicated solely to implementation of the MapReduce model, limiting their ability to undertake other computing tasks. Establishing such a distributed system can therefore be time consuming and inefficient, particularly where a user does not require a continual, dedicated system for implementation of the MapReduce model.

Embodiments of the present disclosure address the above-noted issues by enabling implementation of the MapReduce programming model on an on-demand code execution system (sometimes referred to as a “serverless” computing system) in conjunction with a stream data processing system. As will be described in more detail below, implementation of the MapReduce programming model on an on-demand code execution environment can provide many benefits over traditional MapReduce implementations, such as reducing the need to manage configuration and deployment of worker computing nodes (the management of which may be provided by the on-demand code execution environment), as well as reducing the need for a dedicated “infrastructure” or “framework” to coordinate working computing nodes. Illustratively, rather than utilizing a dedicated infrastructure or framework, an implementation of the MapReduce programming model as described herein may utilize a message stream on a stream data processing system, which may act as an intermediary between execution of “map” tasks on the on-demand code execution system and “reduce” tasks on the system. Specifically, outputs of map tasks may be placed onto the message stream, which outputs can cause execution of reduce tasks on the on-demand code execution system. The stream data processing system can be configured to automatically scale based on the volume of messages within the message stream, and to partition the stream in a manner similar to partitioning of outputs in a traditional MapReduce implementation (e.g., according to a values of a particular attribute in the outputs). Each partition of the stream can be passed to an individual reduce function, thus requiring little external input or control to provide parallelization of a reduce function. As such, embodiments of the present disclosure can greatly simply implementation of the MapReduce programming model.

Prior attempts have been made to implement the MapReduce programming model. For example, in U.S. Pat. No. 11,119,813 to Kasaragod, entitled “MAPREDUCE IMPLEMENTATION USING AN ON-DEMAND NETWORK CODE EXECUTION SYSTEM” (the “'813 patent”), Applicant disclosed systems and methods for implementing the MapReduce programming model on an on-demand code execution system by utilization of a coordinator function. The present disclosure extends on that prior disclosure by providing an implementation of the MapReduce programming model on an on-demand code execution system even absent a coordinator function. As such, the present disclosure may provide improved operation over the prior disclosure by, for example, reducing network traffic over the on-demand code execution system (e.g., to call the coordinator function) and reducing computing resources that would otherwise be used to provide such a function. The entirety of the '813 patent is hereby incorporated by reference.

As described in detail herein, the on-demand code execution system may provide a network-accessible service enabling users to submit or designate computer-executable code to be executed by virtual machine instances on the on-demand code execution system. Each set of code on the on-demand code execution system may define a “task,” and implement specific functionality corresponding to that task when executed on a virtual machine instance of the on-demand code execution system. Individual implementations of the task on the on-demand code execution system may be referred to as an “execution” of the task (or a “task execution”). The on-demand code execution system can further enable users to trigger execution of a task based on a variety of potential events, such as detecting new data at a network-based storage system, transmission of an application programming interface (“API”) call to the on-demand code execution system, or transmission of a specially formatted hypertext transport protocol (“HTTP”) packet to the on-demand code execution system. Thus, users may utilize the on-demand code execution system to execute any specified executable code “on-demand,” without requiring configuration or maintenance of the underlying hardware or infrastructure on which the code is executed. Further, the on-demand code execution system may be configured to execute tasks in a rapid manner (e.g., in under 100 milliseconds [ms]), thus enabling execution of tasks in “real-time” (e.g., with little or no perceptible delay to an end user).

Because the on-demand code execution system can provide the ability to execute a task on-demand, without configuration of an underlying device on which to execute the code, the on-demand code execution system can provide an excellent platform on which to implement a map function. For example, a user may submit a map task to the system, corresponding to code that, when executed, processes a portion of an underlying data set and provides an output. The portion of the data set may be specified within the call to execute the map task, or may be determined at runtime during execution of the map task. For example, an execution of the map task may inspect a location of a data set for a next unprocessed portion of a given size (e.g., a maximum size for an individual execution of the on-demand code execution system), mark the portion as being processed, retrieve and process the portion, and then mark the portion as processed. The execution may then continue to process portions, or to call for an additional execution of the map task, until all portions of the data set have been processed. Rather than requiring centralized coordination, an end user may simply invoke as many executions of the map task as they desire, thus parallelizing the map task across those executions.

In the MapReduce programming model, the outputs of each map function are distributed to reduce functions, according to an attribute of those outputs. This is sometimes referred to as “shuffling” the data, because the map function often intakes data naively (e.g., as arbitrary portions of the data set) and outputs data “intelligently” according to the content of that data. As discussed above, it is often beneficial for the outputs to be distributed across multiple reduce functions, enabling parallelization as the outputs of the map functions are aggregated to provide a final result.

To provide this distribution of outputs without requiring a framework or coordinator, each map task may write their outputs into a message stream on a stream data processing system. Generally, a stream data processing system provides the ability for upstream devices to place data onto a message stream, such as by publishing “messages” onto the stream. Downstream devices may then obtain those messages, often in a “first-in-first-out” (“FIFO”) or nearly FIFO order. In some instances, the stream data processing system “pushes” messages to downstream devices. In other instances, downstream devices “pull” messages from the message stream on request. Generally, the stream data processing system is configured to provide resiliency, such that data successfully published to the stream is unlikely to be lost due to failures of devices of the stream data processing system. Moreover, the stream data processing system can be configured to provide parallelization of the devices that maintain the message stream. For example, a user configuring a message stream may designate a partition key for the stream, used to divide the stream into sub-streams, each sub-stream handled by one or more parallelized devices. The stream data processing system may modify the number of sub-streams based on the volume of data on the stream, to maintain the ability to robustly receive and transmit messages. For example, a low-volume stream may be handled entirely by a single device, while a high-volume stream may by split into numerous sub-streams, each handled by one or more devices. The partition key may correspond to any attribute of the data published to the stream. Often, it is desirable that the partition key represent an attribute of the data for which values are relatively evenly distributed within the data published to the stream. For example, where messages on the stream pertain to actions of user computing devices, the partition key may be selected as a unique identifier of the computing devices. The stream data processing system may then partition the messages according to a distribution of the unique identifiers, such as by modulus dividing the identifier by the current number of sub-streams or applying a hashing to the identifier (e.g., according to a consistent hashing scheme, examples of which are known in the art). Examples of stream data processing systems known in the art include the AMAZON™ KINESIS™ network service and the APACHE™ KAFKA™ system.

In the context of the present disclosure, a stream data processing system may be utilized as an intermediary between executions of map tasks and reduce tasks implementing the MapReduce model. In accordance with one embodiment, the output of each map task execution may be published onto a message stream whose partition key is selected as an attribute of the underlying data set according to which a final result is desired. For example, in the instance of a word count application, the partition key of the message stream may be selected as the word being counted. Because the stream data processing system can function to automatically divide the message stream into sub-streams based on the volume of the stream, the message stream can be utilized by embodiments of the present disclosure to automatically manage parallelization of a reduce task (implementing a reduce function in the MapReduce model). For example, the partition key of a message stream may be used to generate sub-streams (sometimes referred to as “shards”), each including a subset of messages selected according to the partition key (e.g., according to a modulo division or hashing operation). The number of sub-streams may be automatically scaled based on the volume of messages in the stream, which in turn may be based on the parallelization of a map task and the output of the map task executions. In one embodiment, each sub-stream of a messages stream forms an input to a corresponding reduce task execution, thus enabling the parallelization of the reduce task to scale according to the volume of messages produced by the map tasks. Because the partition key of the message stream is selected according to the attribute being mapped (e.g., the word being counted), the portioning of the message stream accomplishes the “shuffling” of data between the map task executions and the reduce task executions. Each reduce task execution can therefore obtain messages from a corresponding sub-stream, apply a reduce function to those messages, and provide an output. The aggregate outputs of the reduce tasks executions can represent a final output of the MapReduce implementation.

As will be appreciated by one of skill in the art in light of the present disclosure, the embodiments disclosed herein improves the ability of computing systems, such as on-demand code execution environments, to process and analyze data sets in a parellizable manner. Moreover, the embodiments disclosed herein represent an improvement to the MapReduce programming model, by reducing or eliminating the need for a persistent framework or infrastructure and by reducing the need for an implementation of the MapReduce programming model to generate or manage worker computing devices that execute map or reduce functions within the implementation. Moreover, the presently disclosed embodiments address technical problems inherent within computing systems; specifically, the limited nature of computing resources in processing large data sets and the inherent complexities in providing rapid, parallelized executions. These technical problems are addressed by the various technical solutions described herein, including the utilization of a stream data processing system as an intermediary between map function executions and reduce function executions. Thus, the present disclosure represents an improvement on existing data processing systems and computing systems in general.

The general execution of tasks on the on-demand code execution environment will now be discussed. Specifically, to execute tasks, the on-demand code execution environment described herein may maintain a pool of pre-initialized virtual machine instances that are ready for use as soon as a user request is received. Due to the pre-initialized nature of these virtual machines, delay (sometimes referred to as latency) associated with executing the user code (e.g., instance and language runtime startup time) can be significantly reduced, often to sub-100 millisecond levels. Illustratively, the on-demand code execution environment may maintain a pool of virtual machine instances on one or more physical computing devices, where each virtual machine instance has one or more software components (e.g., operating systems, language runtimes, libraries, etc.) loaded thereon. When the on-demand code execution environment receives a request to execute the program code of a user (a “task”), which specifies one or more computing constraints for executing the program code of the user, the on-demand code execution environment may select a virtual machine instance for executing the program code of the user based on the one or more computing constraints specified by the request and cause the program code of the user to be executed on the selected virtual machine instance. The program codes can be executed in isolated containers that are created on the virtual machine instances. Since the virtual machine instances in the pool have already been booted and loaded with particular operating systems and language runtimes by the time the requests are received, the delay associated with finding compute capacity that can handle the requests (e.g., by executing the user code in one or more containers created on the virtual machine instances) is significantly reduced.

The on-demand code execution environment may include a virtual machine instance manager configured to receive user code (threads, programs, etc., composed in any of a variety of programming languages) and execute the code in a highly scalable, low latency manner, without requiring user configuration of a virtual machine instance. Specifically, the virtual machine instance manager can, prior to receiving the user code and prior to receiving any information from a user regarding any particular virtual machine instance configuration, create and configure virtual machine instances according to a predetermined set of configurations, each corresponding to any one or more of a variety of run-time environments. Thereafter, the virtual machine instance manager receives user-initiated requests to execute code, and identifies a pre-configured virtual machine instance to execute the code based on configuration information associated with the request. The virtual machine instance manager can further allocate the identified virtual machine instance to execute the user's code at least partly by creating and configuring containers inside the allocated virtual machine instance. Various embodiments for implementing a virtual machine instance manager and executing user code on virtual machine instances is described in more detail in U.S. Pat. No. 9,323,556, entitled “PROGRAMMATIC EVENT DETECTION AND MESSAGE GENERATION FOR REQUESTS TO EXECUTE PROGRAM CODE” and filed Sep. 30, 2014 (“the '556 patent”), the entirety of which is hereby incorporated by reference.

As used herein, the term “virtual machine instance” is intended to refer to an execution of software or other executable code that emulates hardware to provide an environment or platform on which software may execute (an “execution environment”). Virtual machine instances are generally executed by hardware devices, which may differ from the physical hardware emulated by the virtual machine instance. For example, a virtual machine may emulate a first type of processor and memory while being executed on a second type of processor and memory. Thus, virtual machines can be utilized to execute software intended for a first execution environment (e.g., a first operating system) on a physical device that is executing a second execution environment (e.g., a second operating system). In some instances, hardware emulated by a virtual machine instance may be the same or similar to hardware of an underlying device. For example, a device with a first type of processor may implement a plurality of virtual machine instances, each emulating an instance of that first type of processor. Thus, virtual machine instances can be used to divide a device into a number of logical sub-devices (each referred to as a “virtual machine instance”). While virtual machine instances can generally provide a level of abstraction away from the hardware of an underlying physical device, this abstraction is not required. For example, assume a device implements a plurality of virtual machine instances, each of which emulate hardware identical to that provided by the device. Under such a scenario, each virtual machine instance may allow a software application to execute code on the underlying hardware without translation, while maintaining a logical separation between software applications running on other virtual machine instances. This process, which is generally referred to as “native execution,” may be utilized to increase the speed or performance of virtual machine instances. Other techniques that allow direct utilization of underlying hardware, such as hardware pass-through techniques, may be used, as well.

While a virtual machine executing an operating system is described herein as one example of an execution environment, other execution environments are also possible. For example, tasks or other processes may be executed within a software “container,” which provides a runtime environment without itself providing virtualization of hardware. Containers may be implemented within virtual machines to provide additional security, or may be run outside of a virtual machine instance.

The foregoing aspects and many of the attendant advantages of this disclosure will become more readily appreciated as the same become better understood by reference to the following description, when taken in conjunction with the accompanying drawings.

FIG. 1 is a block diagram of an illustrative operating environment 100 in which an on-demand code execution system 110 may operate based on communication with user computing devices 102, auxiliary services 106, and network-based data storage services 108. By way of illustration, various example user computing devices 102 are shown in communication with the on-demand code execution system 110, including a desktop computer, laptop, and a mobile phone. In general, the user computing devices 102 can be any computing device such as a desktop, laptop or tablet computer, personal computer, wearable computer, server, personal digital assistant (PDA), hybrid PDA/mobile phone, mobile phone, electronic book reader, set-top box, voice command device, camera, digital media player, and the like. The on-demand code execution system 110 may provide the user computing devices 102 with one or more user interfaces, command-line interfaces (CLI), application programing interfaces (API), and/or other programmatic interfaces for generating and uploading user-executable code, invoking the user-provided code (e.g., submitting a request to execute the user codes on the on-demand code execution system 110), scheduling event-based jobs or timed jobs, tracking the user-provided code, and/or viewing other logging or monitoring information related to their requests and/or user codes. Although one or more embodiments may be described herein as using a user interface, it should be appreciated that such embodiments may, additionally or alternatively, use any CLIs, APIs, or other programmatic interfaces.

The illustrative environment 100 further includes one or more auxiliary services 106, which can interact with the on-demand code execution system 110 to implement desired functionality on behalf of a user. Auxiliary services 106 can correspond to network-connected computing devices, such as servers, which generate data accessible to the on-demand code execution system 110 or otherwise communicate to the on-demand code execution system 110. For example, the auxiliary services 106 can include web services (e.g., associated with the user computing devices 102, with the on-demand code execution system 110, or with third parties), databases, really simple syndication (“RSS”) readers, social networking sites, or any other source of network-accessible service or data source. In some instances, auxiliary services 106 may be associated with the on-demand code execution system 110, e.g., to provide billing or logging services to the on-demand code execution system 110. In some instances, auxiliary services 106 actively transmit information, such as API calls or other task-triggering information, to the on-demand code execution system 110. In other instances, auxiliary services 106 may be passive, such that data is made available for access by the on-demand code execution system 110. As described below, components of the on-demand code execution system 110 may periodically poll such passive data sources, and trigger execution of tasks within the on-demand code execution system 110 based on the data provided. While depicted in FIG. 1 as distinct from the user computing devices 102 and the on-demand code execution system 110, in some embodiments, various auxiliary services 106 may be implemented by either the user computing devices 102 or the on-demand code execution system 110.

The illustrative environment 100 further includes one or more network-based data storage services 108, configured to enable the on-demand code execution system 110 to store and retrieve data from one or more persistent or substantially persistent data sources. Illustratively, the network-based data storage services 108 may enable the on-demand code execution system 110 to retrieve a set of data to be analyzed, and store information (e.g., results) regarding that analysis. The network-based data storage services 108 may represent, for example, a relational or non-relational database. In another example, the network-based data storage services 108 may represent a network-attached storage (NAS), configured to provide access to data arranged as a file system. Various other functionalities may be included within network-based data storage services 108 usable within embodiments of the present disclosure. The network-based data storage services 108 may further enable the on-demand code execution system 110 to query for and retrieve information regarding data stored within the on-demand code execution system 110, such as by querying for a number of relevant files or records, sizes of those files or records, file or record names, file or record creation times, etc. In some instances, the network-based data storage services 108 may provide additional functionality, such as the ability to separate data into logical groups (e.g., groups associated with individual accounts, etc.). While shown as distinct from the auxiliary services 106, the network-based data storage services 108 may in some instances also represent a type of auxiliary service 106.

The illustrative environment 100 further includes a stream data processing system 170. As discussed above, the stream data processing system can provides the ability for upstream devices to place data onto a message stream 172, such as by publishing “messages” onto the stream 172, which may be designated based on a specific “topic.” While a single stream 172 is shown in FIG. 1, the system 170 may provide multiple streams on behalf of multiple parties. The system 170 can make messages within the stream 172 available to downstream devices, often in a “first-in-first-out” (“FIFO”) or nearly FIFO order. In some instances, the stream data processing system 170 “pushes” messages to downstream devices. In other instances, downstream devices “pull” messages from the message stream 172 on request. Generally, the stream data processing system 170 is configured to provide resiliency, such that data successfully published to the stream is unlikely to be lost due to failures of devices of the stream data processing system 170. For example, the system 170 may duplicate messages placed onto the stream 172 onto multiple computing devices used to implement the stream (e.g., physical computing devices or virtual devices implemented on physical hosts). Moreover, the stream data processing system 170 can be configured to provide parallelization of the devices that maintain the message stream 172. For example, a user configuring a message stream may designate a partition key for the stream, used to divide the stream into sub-streams, each sub-stream handled by one or more parallelized devices. The sub-streams are shown in FIG. 1 as message shards 174A-N. Each message shard 174 can generally represent one or more computing devices configured to obtain and make available a subset of messages on the message stream, selected by the system 170 according to the partition key and a volume of messages on the stream 170 (e.g., such that additional shards are created, or excess shards are destroyed, based on a capacity of the shards 174 to service messages on the stream 172). Examples of stream data processing systems known in the art include the AMAZON™ KINESIS™ network service and the APACHE™ KAFKA™ system.

The user computing devices 102, auxiliary services 106, network-based data storage services 108, and stream data processing system 170 may communicate with the on-demand code execution system 110 via network 104, which may include any wired network, wireless network, or combination thereof. For example, the network 104 may be a personal area network, local area network, wide area network, over-the-air broadcast network (e.g., for radio or television), cable network, satellite network, cellular telephone network, or combination thereof. As a further example, the network 104 may be a publicly accessible network of linked networks, possibly operated by various distinct parties, such as the Internet. In some embodiments, the network 104 may be a private or semi-private network, such as a corporate or university intranet. The network 104 may include one or more wireless networks, such as a Global System for Mobile Communications (GSM) network, a Code Division Multiple Access (CDMA) network, a Long Term Evolution (LTE) network, or any other type of wireless network. The network 104 can use protocols and components for communicating via the Internet or any of the other aforementioned types of networks. For example, the protocols used by the network 104 may include Hypertext Transfer Protocol (HTTP), HTTP Secure (HTTPS), Message Queue Telemetry Transport (MQTT), Constrained Application Protocol (CoAP), and the like. Protocols and components for communicating via the Internet or any of the other aforementioned types of communication networks are well known to those skilled in the art and, thus, are not described in more detail herein.

The on-demand code execution system 110 and stream data processing system 170 are depicted in FIG. 1 as operating in a distributed computing environment including several computer systems that are interconnected using one or more computer networks (not shown in FIG. 1). Either or both of the on-demand code execution system 110 and stream data processing system 170 could also operate within a computing environment having a fewer or greater number of devices than are illustrated in FIG. 1. Thus, the depiction of the on-demand code execution system 110 and stream data processing system 170 in FIG. 1 should be taken as illustrative and not limiting to the present disclosure. For example, the on-demand code execution system 110 and stream data processing system 170 or various constituents thereof could implement various Web services components, hosted or “cloud” computing environments, and/or peer to peer network configurations to implement at least a portion of the processes described herein.

Further, the on-demand code execution system 110 and stream data processing system 170 may be implemented directly in hardware or software executed by hardware devices and may, for instance, include one or more physical or virtual servers implemented on physical computer hardware configured to execute computer executable instructions for performing various features that will be described herein. The one or more servers may be geographically dispersed or geographically co-located, for instance, in one or more data centers. In some instances, the one or more servers may operate as part of a system of rapidly provisioned and released computing resources, often referred to as a “cloud computing environment.”

In the example of FIG. 1, the on-demand code execution system 110 and stream data processing system 170 are illustrated as connected to the network 104. In some embodiments, any of the components within the on-demand code execution system 110 and stream data processing system 170 can communicate with other components of the on-demand code execution system 110 and stream data processing system 170 via the network 104. In other embodiments, another network (such as a private network not shown in FIG. 1) may enable communication between components within each of the on-demand code execution system 110 and stream data processing system 170 or between those systems.

In FIG. 1, users, by way of user computing devices 102, may interact with the on-demand code execution system 110 to provide executable code, and establish rules or logic defining when and how such code should be executed on the on-demand code execution system 110, thus establishing a “task.” For example, a user may wish to run a piece of code in connection with a web or mobile application that the user has developed. One way of running the code would be to acquire virtual machine instances from service providers who provide infrastructure as a service, configure the virtual machine instances to suit the user's needs, and use the configured virtual machine instances to run the code. In order to avoid the complexity of this process, the user may alternatively provide the code to the on-demand code execution system 110, and request that the on-demand code execution system 110 execute the code using one or more pre-established virtual machine instances. The on-demand code execution system 110 can handle the acquisition and configuration of compute capacity (e.g., containers, instances, etc., which are described in greater detail below) based on the code execution request, and execute the code using the compute capacity. The on-demand code execution system 110 may automatically scale up and down based on the volume, thereby relieving the user from the burden of having to worry about over-utilization (e.g., acquiring too little computing resources and suffering performance issues) or under-utilization (e.g., acquiring more computing resources than necessary to run the codes, and thus overpaying). In accordance with embodiments of the present disclosure, the tasks established by a user may correspond to code executable to implement “map” and “reduce” functions with respect to a data set.

To enable interaction with the on-demand code execution system 110, the environment 110 includes one or more frontends 120, which enable interaction with the on-demand code execution system 110. In an illustrative embodiment, the frontends 120 serve as a “front door” to the other services provided by the on-demand code execution system 110, enabling users (via user computing devices 102) to provide, request execution of, and view results of computer executable code. The frontends 120 include a variety of components to enable interaction between the on-demand code execution system 110 and other computing devices. For example, each frontend 120 may include a request interface providing user computing devices 102 with the ability to upload or otherwise communication user-specified code to the on-demand code execution system 110 and to thereafter request execution of that code. In one embodiment, the request interface communicates with external computing devices (e.g., user computing devices 102, auxiliary services 106, etc.) via a graphical user interface (GUI), CLI, or API. The frontends 120 process the requests and makes sure that the requests are properly authorized. For example, the frontends 120 may determine whether the user associated with the request is authorized to access the user code specified in the request.

References to user code as used herein may refer to any program code (e.g., a program, routine, subroutine, thread, etc.) written in a specific program language. In the present disclosure, the terms “code,” “user code,” and “program code,” may be used interchangeably. Such user code may be executed to achieve a specific function, for example, in connection with a particular web application or mobile application developed by the user. As noted above, individual collections of user code (e.g., to achieve a specific function) are referred to herein as “tasks,” while specific executions of that code are referred to as “task executions” or simply “executions.” Tasks may be written, by way of non-limiting example, in JavaScript (e.g., node.js), Java, Python, and/or Ruby (and/or another programming language). Tasks may be “triggered” for execution on the on-demand code execution system 110 in a variety of manners. In one embodiment, a user or other computing device may transmit a request to execute a task may, which can generally be referred to as “call” to execute of the task. Such calls may include the user code (or the location thereof) to be executed and one or more arguments to be used for executing the user code. For example, a call may provide the user code of a task along with the request to execute the task. In another example, a call may identify a previously uploaded task by its name or an identifier. In yet another example, code corresponding to a task may be included in a call for the task, as well as being uploaded in a separate location (e.g., storage of an auxiliary service 106 or a storage system internal to the on-demand code execution system 110) prior to the request being received by the on-demand code execution system 110. The on-demand code execution system 110 may vary its execution strategy for a task based on where the code of the task is available at the time a call for the task is processed. A request interface of the frontend 120 may receive calls to execute tasks as Hypertext Transfer Protocol Secure (HTTPS) requests from a user. Also, any information (e.g., headers and parameters) included in the HTTPS request may also be processed and utilized when executing a task. As discussed above, any other protocols, including, for example, HTTP, MQTT, and CoAP, may be used to transfer the message containing a task call to the request interface.

A call to execute a task may specify one or more third-party libraries (including native libraries) to be used along with the user code corresponding to the task. In one embodiment, the call may provide to the on-demand code execution system 110 a ZIP file containing the user code and any libraries (and/or identifications of storage locations thereof) corresponding to the task requested for execution. In some embodiments, the call includes metadata that indicates the program code of the task to be executed, the language in which the program code is written, the user associated with the call, and/or the computing resources (e.g., memory, etc.) to be reserved for executing the program code. For example, the program code of a task may be provided with the call, previously uploaded by the user, provided by the on-demand code execution system 110 (e.g., standard routines), and/or provided by third parties. In some embodiments, such resource-level constraints (e.g., how much memory is to be allocated for executing a particular user code) are specified for the particular task, and may not vary over each execution of the task. In such cases, the on-demand code execution system 110 may have access to such resource-level constraints before each individual call is received, and the individual call may not specify such resource-level constraints. In some embodiments, the call may specify other constraints such as permission data that indicates what kind of permissions or authorities that the call invokes to execute the task. Such permission data may be used by the on-demand code execution system 110 to access private resources (e.g., on a private network).

In some embodiments, a call may specify the behavior that should be adopted for handling the call. In such embodiments, the call may include an indicator for enabling one or more execution modes in which to execute the task referenced in the call. For example, the call may include a flag or a header for indicating whether the task should be executed in a debug mode in which the debugging and/or logging output that may be generated in connection with the execution of the task is provided back to the user (e.g., via a console user interface). In such an example, the on-demand code execution system 110 may inspect the call and look for the flag or the header, and if it is present, the on-demand code execution system 110 may modify the behavior (e.g., logging facilities) of the container in which the task is executed, and cause the output data to be provided back to the user. In some embodiments, the behavior/mode indicators are added to the call by the user interface provided to the user by the on-demand code execution system 110. Other features such as source code profiling, remote debugging, etc. may also be enabled or disabled based on the indication provided in a call.

In accordance with embodiments of the present disclosure, user-submitted code may correspond to a map function and a reduce function for a set of data. In one embodiment, each of the map function and the reduce function may correspond to a distinct set of code, corresponding to a distinct task on the on-demand code execution system 110. In another embodiment, both the map and reduce function may correspond to a single task on the on-demand code execution system 110. In accordance with the MapReduce programming model, the specific functionalities of the map and reduce functions may vary according to the data to be processed. However, in general terms, a map function may correspond to code that processes a portion of an initial data set (e.g., “raw” data) in order to generate intermediate results, while a reduce function corresponds to code to “reduce” or aggregate multiple intermediate results to an aggregate result. For example, a map function may correspond to code to process a corpus of words (e.g., a book or a collection of books) and generate a count of individual words (or, in some instances, specific classes of words, such as surnames) within the corpus. A corresponding reduce function can correspond to code that aggregates the counts of individual words, as produced by individual executions of the map function. Illustratively, a reduce function may take one thousand intermediate count data files, as produced by one thousand instances of the map function, and generate a single data file aggregating the counts of each word within the one thousand intermediate count data files. Where only a single reduce function is executed, the single data file can represent the result of the analysis. Where multiple reduce functions are executed, resulting in multiple outputs (each aggregating counts of words within a set of inputs), an additional reduce function can process those multiple outputs to further aggregate word counts. This process can continue until single reduce function is called, resulting in a single output file as the result of the analysis. Alternatively, each of multiple reduce functions may write output to a common location (e.g., a database) as a final output of the analysis.

Because map and reduce functions are often specific to the data to be processed, code corresponding to these functions may be provided by an end user requesting analysis of a corresponding data set. In some instances, the on-demand code execution system 110 may also provide one or more sets of corresponding map and reduce functions (e.g., corresponding to commonly desired analysis types and configured to process data within an expected format).

To enable storage of tasks, the on-demand code execution system 110 may include a task data store 160, which may correspond to a persistent or substantially persistent data store, such as a hard drive (HDD), a solid state drive (SDD), network attached storage (NAS), a tape drive, or any combination thereof. In accordance with embodiments of the present disclosure, the task data store may include one or more map tasks 164 and one or more reduce tasks 166.

To manage requests for code execution, the frontend 120 can include an execution queue (not shown in FIG. 1), which can maintain a record of requested task executions. Illustratively, the number of simultaneous task executions by the on-demand code execution system 110 is limited, and as such, new task executions initiated at the on-demand code execution system 110 (e.g., via an API call, via a call from an executed or executing task, etc.) may be placed on the execution queue 124 and processed, e.g., in a first-in-first-out order. In some embodiments, the on-demand code execution system 110 may include multiple execution queues, such as individual execution queues for each user account. For example, users of the on-demand code execution system 110 may desire to limit the rate of task executions on the on-demand code execution system 110 (e.g., for cost reasons). Thus, the on-demand code execution system 110 may utilize an account-specific execution queue to throttle the rate of simultaneous task executions by a specific user account. In some instances, the on-demand code execution system 110 may prioritize task executions, such that task executions of specific accounts or of specified priorities bypass or are prioritized within the execution queue. In other instances, the on-demand code execution system 110 may execute tasks immediately or substantially immediately after receiving a call for that task, and thus, the execution queue may be omitted.

In addition to tasks executed based on explicit user calls and data from auxiliary services 106, the on-demand code execution system 110 may in some instances operate to trigger execution of tasks independently. For example, the on-demand code execution system 110 may operate (based on instructions from a user) to trigger execution of a task at each of a number of specified time intervals (e.g., every 10 minutes).

The frontend 120 can further includes an output interface (not shown in FIG. 1) configured to output information regarding the execution of tasks on the on-demand code execution system 110. Illustratively, the output interface may transmit data regarding task executions (e.g., results of a task, errors related to the task execution, or details of the task execution, such as total time required to complete the execution, total data processed via the execution, etc.) to the user computing devices 102 or to auxiliary services 106, which may include, for example, billing or logging services. The output interface may further enable transmission of data, such as service calls, to auxiliary services 106. For example, the output interface may be utilized during execution of a task to transmit an API request to an auxillary service 106 (e.g., to store data generated during execution of the task).

In some embodiments, the on-demand code execution system 110 may include multiple frontends 120. In such embodiments, a load balancer (not shown in FIG. 1) may be provided to distribute the incoming calls to the multiple frontends 120, for example, in a round-robin fashion. In some embodiments, the manner in which the load balancer distributes incoming calls to the multiple frontends 120 may be based on the location or state of other components of the on-demand code execution system 110. For example, a load balancer may distribute calls to a geographically nearby frontend 120, or to a frontend with capacity to service the call. In instances where each frontend 120 corresponds to an individual instance of another component of the on-demand code execution environment, such as the active pools 140A described below, the load balancer may distribute calls according to the capacities or loads on those other components. As will be described in more detail below, calls may in some instances be distributed between frontends 120 deterministically, such that a given call to execute a task will always (or almost always) be routed to the same frontend 120. This may, for example, assist in maintaining an accurate execution record for a task, to ensure that the task executes only a desired number of times. While distribution of calls via a load balancer is illustratively described, other distribution techniques, such as anycast routing, will be apparent to those of skill in the art.

The on-demand code execution system 110 further includes one or more worker managers 140 that manage the execution environments, such as virtual machine instances 150 (shown as VM instance 150A and 150B, generally referred to as a “VM”), used for servicing incoming calls to execute tasks, and that manage the memory states of execution environments. While the following will be described with reference to virtual machine instances 150 as examples of such environments, embodiments of the present disclosure may utilize other environments, such as software containers. In the example illustrated in FIG. 1, each worker manager 140 manages an active pool 140A, which is a group (sometimes referred to as a pool) of virtual machine instances 150 executing on one or more physical host computing devices that are initialized to execute a given task (e.g., by having the code of the task and any dependency data objects loaded into the instance). The active pool 140A illustratively is implemented using primary memory (e.g., RAM) of host devices implementing or under control of the worker manager 140.

Although the virtual machine instances 150 are described here as being assigned to a particular task, in some embodiments, the instances may be assigned to a group of tasks, such that the instance is tied to the group of tasks and any tasks of the group can be executed within the instance. For example, the tasks in the same group may belong to the same security group (e.g., based on their security credentials) such that executing one task in a container on a particular instance 150 after another task has been executed in another container on the same instance does not pose security risks. As another example, the tasks of the group may share common dependencies, such that an environment used to execute one task of the group can be rapidly modified to support execution of another task within the group.

Once a triggering event to execute a task has been successfully processed by a frontend 120, the frontend 120 passes a request to a worker manager 140 to execute the task. In one embodiment, each frontend 120 may be associated with a corresponding worker manager 140 (e.g., a worker manager 140 co-located or geographically nearby to the frontend 120) and thus, the frontend 120 may pass most or all requests to that worker manager 140. In another embodiment, a frontend 120 may include a location selector configured to determine a worker manager 140 to which to pass the execution request. In one embodiment, the location selector may determine the worker manager 140 to receive a call based on hashing the call, and distributing the call to a worker manager 140 selected based on the hashed value (e.g., via a hash ring). Various other mechanisms for distributing calls between worker managers 140 will be apparent to one of skill in the art.

Thereafter, the worker manager 140 may modify a virtual machine instance 150 (if necessary) and execute the code of the task within the instance 150. As shown in FIG. 1, respective instances 150 may have operating systems (OS) 152 (shown as OS 152A and 152B), language runtimes 154 (shown as runtime 154A and 154B), and user code 156 (shown as user code 156A and 156B). The OS 152, runtime 154, and user code 156 may collectively enable execution of the user code to implement the task. In some instances, each VM 150 may be associated with additional information, such as state information, maintained across individual executions of a task. For example, when initially created, a VM 150 may initialize the OS 152, and each time the user code 156 is executed in the VM 150, a state of the VM 150 may change. State of a VM 150 may be maintained, for example, within registers of a virtual CPU of the VM 150, within RAM of the VM 150, within a virtual disk drive of the VM 150, or the like.

As noted above, tasks may be triggered for execution at the on-demand code execution system 110 based on explicit calls from user computing devices 102 (e.g., as received at the request interface 122). Alternatively or additionally, tasks may be triggered for execution at the on-demand code execution system 110 based on data retrieved from one or more auxiliary services 106, network-based data storage services 108, or the stream data processing system 170. To facilitate interaction with auxiliary services 106, the system 110 including a polling fleet 130, which operates to poll auxiliary services 106, data storage services 108, or the stream data processing system 170 for data. Illustratively, the polling fleet 130 may include one or more computing devices (shown in FIG. 1 as poller devices 132A-N) configured to periodically transmit a request to the auxiliary services 106, data storage services 108, or stream data processing system 170 to retrieve any newly available data (e.g., social network “posts,” news articles, files, records, etc.), and to determine whether that data corresponds to a user-established criteria triggering execution a task on the on-demand code execution system 110. Illustratively, criteria for execution of a task may include, but is not limited to, whether new data is available at the auxiliary services 106, data storage services 108, or the stream data processing system 170, the type or content of the data, or timing information corresponding to the data. In some instances, the auxiliary services 106, data storage services 108, or stream data processing system 170 may function to notify the frontend 120 of the availability of new data, and thus the polling fleet 130 may be unnecessary with respect to such services.

In accordance with embodiments of the present disclosure, the poller fleet 130 can be configured to include a dynamic number of poller devices 132A-N (e.g., implemented as virtual machine instances on an underlying computing system), based on the number of message shards 174 within a message stream 172 onto which outputs of map task executions are placed. For example, as shown by the dotted lines of FIG. 1, message shard 174A may correspond to poller device 132A, message shard 174B may correspond to poller device 132B, etc. Thus, as the number of message shards 174 changes (e.g., due to volume of the message stream, which can in turn be due to operation of map task executions), the number of poller devices 132 may also change. As such, the poller fleet 130 may be in communication with stream data processing system 170, and the system 170 may notify the poller fleet 130 of changes to the message shards 174. In such a configuration, each poller device 132A can be configured to poll a message shard 174 to retrieve messages in the sub-stream corresponding to the message shard. The messages may be retrieved individually or in batches (e.g., batches of 10 messages, 50 messages, 100 messages, 500 messages, etc.). Thereafter, the poller device 132 may invoke a call to a reduce task to process each message. In one embodiment, a call to the reduce task may be made for each individual message. In another embodiment, a call to the reduce task may be made for each batch of messages, with a batch size matching or differing from the batch size for retrieval of messages from the message shard. In some instances, the call from each poller device 132 to a reduce task execution may be made synchronously, such that the poller device 132 waits for confirmation that the execution was successful prior to making a next call. Use of synchronous calls may beneficially rate-limit calls to the reduce task, such that a generally one-to-one correspondence is maintained between the number of poller devices 132 and reduce task executions.

While some functionalities are generally described herein with reference to an individual component of the on-demand code execution system 110 or the stream data processing system 170, other components or a combination of components may additionally or alternatively implement such functionalities. For example, while a poller device 132A may operate to poll a message shard 174 for messages, the message shards 174 may additionally or alternatively be configured to notify the on-demand code execution system 110 (e.g., the frontend) of new messages on the shard 174.

FIG. 2 depicts a general architecture of a computing system (referenced as worker manager 140) that manages the virtual machine instances in the on-demand code execution system 110. The general architecture of the worker manager 140 depicted in FIG. 2 includes an arrangement of computer hardware and software modules that may be used to implement aspects of the present disclosure. The hardware modules may be implemented with physical electronic devices, as discussed in greater detail below. The worker manager 140 may include many more (or fewer) elements than those shown in FIG. 2. It is not necessary, however, that all of these generally conventional elements be shown in order to provide an enabling disclosure. Additionally, the general architecture illustrated in FIG. 2 may be used to implement one or more of the other components illustrated in FIG. 1. As illustrated, the worker manager 140 includes a processing unit 190, a network interface 192, a computer readable medium drive 194, and an input/output device interface 196, all of which may communicate with one another by way of a communication bus. The network interface 192 may provide connectivity to one or more networks or computing systems. The processing unit 190 may thus receive information and instructions from other computing systems or services via the network 104. The processing unit 190 may also communicate to and from memory 180 and further provide output information for an optional display (not shown) via the input/output device interface 196. The input/output device interface 196 may also accept input from an optional input device (not shown).

The memory 180 may contain computer program instructions (grouped as modules in some embodiments) that the processing unit 190 executes in order to implement one or more aspects of the present disclosure. The memory 180 generally includes random access memory (RAM), read only memory (ROM) and/or other persistent, auxiliary or non-transitory computer readable media. The memory 180 may store an operating system 184 that provides computer program instructions for use by the processing unit 190 in the general administration and operation of the worker manager 140. The memory 180 may further include computer program instructions and other information for implementing aspects of the present disclosure. For example, in one embodiment, the memory 180 includes a user interface unit 182 that generates user interfaces (and/or instructions therefor) for display upon a computing device, e.g., via a navigation and/or browsing interface such as a browser or application installed on the computing device. In addition, the memory 180 may include and/or communicate with one or more data repositories (not shown), for example, to access user program codes and/or libraries.

In addition to and/or in combination with the user interface unit 182, the memory 180 may include an instance allocation unit 186 and a user code execution unit 188 that may be executed by the processing unit 190. In one embodiment, the user interface unit 182, instance allocation unit 186, and user code execution unit 188 individually or collectively implement various aspects of the present disclosure, e.g., finding compute capacity (e.g., a container) to be used for executing user code, causing the user code to be loaded and executed on the container, etc. as described further below.

The instance allocation unit 186 finds the compute capacity to be used for servicing a request to execute user code. For example, the instance allocation unit 186 identifies a virtual machine instance and/or a container that satisfies any constraints specified by the request and assigns the identified virtual machine instance and/or container to the user or the request itself. The instance allocation unit 186 may perform such identification based on the programming language in which the user code is written. For example, if the user code is written in Python, and the instance allocation unit 186 may find n virtual machine instance (e.g., in the active pool 140A of FIG. 1) having the Python runtime pre-loaded thereon and assign the virtual machine instance to the user. In another example, if the program code specified in the request of the user is already loaded on an existing container or on another virtual machine instance assigned to the user (e.g., in the active pool 140A of FIG. 1), the instance allocation unit 186 may cause the request to be processed in the container or in a new container on the virtual machine instance. In some embodiments, if the virtual machine instance has multiple language runtimes loaded thereon, the instance allocation unit 186 may create a new container on the virtual machine instance and load the appropriate language runtime on the container based on the computing constraints specified in the request.

The user code execution unit 188 manages the execution of the program code specified by the request of the user once a particular virtual machine instance has been assigned to the user associated with the request and a container on the particular virtual machine instance has been assigned to the request. If the code is pre-loaded in a container on the virtual machine instance assigned to the user, the code is simply executed in the container. If the code is available via a network storage (e.g., storage service 108 of FIG. 1), the user code execution unit 188 downloads the code into a container on the virtual machine instance and causes the code to be executed (e.g., by communicating with the frontend 120 of FIG. 1) once it has been downloaded.

While the instance allocation unit 186 and the user code execution unit 188 are shown in FIG. 2 as part of the worker manager 140, in other embodiments, all or a portion of the instance allocation unit 186 and the user code execution unit 188 may be implemented by other components of the on-demand code execution system 110 and/or another computing device. For example, in certain embodiments of the present disclosure, another computing device in communication with the on-demand code execution system 110 may include several modules or components that operate similarly to the modules and components illustrated as part of the worker manager 140.

In some embodiments, the worker manager 140 may further include components other than those illustrated in FIG. 2. For example, the memory 180 may further include a container manager for managing creation, preparation, and configuration of containers within virtual machine instances.

With reference to FIG. 3, illustrative interactions are depicted for initiating an implementation of the MapReduce programming model on the on-demand code execution environment 110 of FIG. 1, including initiating multiple executions a map task. Specifically, the interactions of FIG. 3 are illustrative of those that may be undertaken by the system 110 to receive and respond to a user request to process a data set on a network based data storage service 108, according to map and reduce functions provided by the user.

The interactions of FIG. 3 begin at (1), where a user device 102 configures the on-demand code execution system 110 to implement a MapReduce processing model by configuring map and reduce tasks on the system 110. In the illustrative interactions of FIG. 3, the map function and reduce function are designated as tasks within the on-demand code execution system 110, which may have been previously created by the user device 102 or which may be created by the on-demand code execution system 110 as part of configuring the map and reduce tasks (e.g., the user may submit code for the map and reduce tasks as part of configuring those tasks to implement the MapReduce model). In other embodiments, a user may designate other map and reduce functions, such as those carried out by tasks made available by the on-demand code execution system 110 or other users. In addition to designation of map and reduce functions, the configuration generally includes linking the map and reduce tasks, such that outputs of executions of the map task are processed via corresponding executions of the reduce task. Illustratively, linking the map and reduce tasks may include designating an output location of the map task, such as a stream data processing system 170, and further designating the inclusion of new data at that output location as a “trigger” on which to execute the reduce task. In some embodiments, linking of map and reduce tasks may be accomplished at least partially by modification of the code of the respective tasks. For example, the stream data processing system 170 may be designated as an output location for the map task by alteration of the code for the map task. In other embodiments, linking of the map and reduce tasks may be accomplished without requiring modification of the code of the respective tasks. For example, the code of the map task may specify an “output location” parameter, and configuration of the map and reduce tasks may include noting to the system 110 that the output location parameter should, on execution of the map task, be designated as the stream data processing system 170. In some instances, the system 110 may be configured to automatically link map and reduce tasks, without the need for a user device 102 to specify use of the stream data processing system 170. For example, a user device 102 may designate map and reduce tasks to the system 110 along with a request to implement a MapReduce model measuring an attribute of a target data set, and the system 110 may interact with the stream data processing system 170 to generate a message stream partitioned according to the attribute, set an output parameter of the map task to the message stream, and configure publication of data on the message stream as a trigger on which to execute the reduce task.

In some embodiments, configuration of the map and reduce tasks may include designation of a set of data to be processed according to the MapReduce model. In the illustrative interactions of FIG. 3, the set of data resides on a data storage service 108A in communication with the on-demand code execution system 110 (which data storage service 108A and on-demand code execution system 110 may, for example, be implemented by a common entity). Accordingly, the configuration information may include an identifier of the set of data, such as a uniform resource identifier (URI) or other identifier of the data on the data storage service 108A.

The illustrative configuration information further includes parallelization parameters desired by the user device 102 for at least the executions of the map task. As discussed below, these parallelization parameters may be utilized to determine a number of executions of the map task to use when processing the data set. In some instances, these parallelization parameters may be pre-specified (e.g., by the user), and thus need not by specified within the configuration information. For example, the user may specify parallelization parameters by specification of a maximum number of concurrent task executions which the system 110 should allow for the user. In another example, the parallelization parameters may be set to default values by the system 110, and thus similarly may not be required to be specified. In some instances, the parallelization parameters may further include parameters for parallelization of executions of the reduce task. Illustratively, parallelization parameters for parallelization of executions of the reduce task may control a number (e.g., a minimum or maximum number) of shards 174 within the message stream 172.

Thereafter, the frontend 120, at (2), transmits the provided map and reduce tasks (if required) to the task data store 160, for later retrieval and execution. The frontend 120 may further store configuration information by the map and reduce tasks implement the MapReduce model.

At (3), the user device 120 submits a call to the on-demand code execution system 110 to initiate the processing of the data set according to the MapReduce model. In one embodiment, the call may be submitted based on call information returned by the system 110 to the user during configuration of the map and reduce tasks, such as an API endpoint to which calls may be submitted to invoke MapReduce processing. In another embodiment, the call may include a call to execute one or more instances of the map task, similarly to how other tasks may be called on the system 110.

While configuration of a MapReduce process and a call to initiate that processing are described above separately, they may in some instances occur concurrently. For example, a user may configure and initiate a MapReduce model of processing via a single call to the system 110. Moreover, while configuration information is discussed above as being sent prior to initiation of MapReduce processing, some or all configuration information may be passed within a call to initiate MapReduce processing.

Thereafter, at (4), the frontend 120 distributes to the worker manager 140 instructions to execute the map task according to the parallelization parameters. For example, where the parallelization parameters specify that a parallelization level of 100, the frontend 120 may distribute to the worker manager 140 instructions to initiate 100 map task executions. In the illustrative embodiments of FIG. 1, the map task executions may be configured to process the data set in a peer-to-peer manner, such that the map task executions coordinate to process the entire data set. For example, each map task executions may operate to “check out” a portion of the data set during processing, and to mark the portion when processing is completed. This process may illustratively continue until no data remains to be processed. In this manner, the entirety of the data set may be processed without requiring centralized coordination of the map task executions.

At (5), the worker manager retrieves code corresponding to the map task from the task data store 160, if required (e.g., if the code is not already provisioned within an execution environment managed by the worker manager 140). At (6), the worker manager initiates a set of executions of the map task within an execution environment. Further details of how such an execution environment may be selected, provisioned with requisite data, and managed, are provided in the '556 patent, incorporated by reference above.

With reference to FIG. 4A, illustrative interactions are shown of map and reduce task executions on the on-demand code execution environment of FIG. 1 to process data according to the MapReduce programming model. For example, the interactions of FIG. 4A may occur in response the interactions of FIG. 3, described above. For ease of description, the interactions of FIG. 4A are discussed with reference to the aggregate functionality of multiple map and reduce task executions. An alternative view of these interactions as they pertain to individual executions will be discussed with reference to FIG. 4B, below. Moreover, the interactions of FIG. 4A depict a data flow between map and reduce task executions, and are discussed sequentially. However, in practice, multiple instances of such a data flow may occur concurrently due to parallelization of functionalities within the system 110.

The interactions of FIG. 4A begin at (1), where the map task executions 402 retrieve base (or “raw”) data from the data storage services 108. Illustratively, each map task execution 402 may retrieve a separate portion (or “chunk”) of the data set to be processed. At (2), the map task executions execute map functionality, as reflected in the code of the map task, to analyze the portion of the data set and to determine a set of intermediate results. For example, each map task execution may generate a count of words with the portion of the data set. Thereafter, at (3), the map task execution writes an output of the task execution to the message stream 172, by publishing the output as a message on the message stream. In one embodiment, each message published to the message stream represents the output of the map task execution with respect to a complete portion or chunk of the data set. In other embodiments, a map task execution may periodically publish messages to the message stream 172 during processing of a portion, and thus each message may reflect a fraction of the output of a map task execution with respect to a portion. In one embodiment, each message published to the message stream 172 may include identifying information as to the portion of the data set represented in the message, to assist in error checking or deduplication of messages (e.g., in the case of a failed map task execution).

At (4), poller devices 132 retrieve messages from the message stream 172, including the outputs of the map task executions. The poller devices 132, at (5), call for executions of the reduce task to process those outputs. Illustratively, poller devices 132 may obtain messages from the stream 172 in batches, and operate to call for executions of the reduce function for each message or for collections of messages, to process the content of the message. The poller devices 132 may continue to retrieve messages from the message stream until no further outputs are provided by the map task executions to the stream 172. In some instances, a final execution of the map task (e.g., an execution which determines that it has processed a final portion of the data set) may publish a “complete” message to the data stream, which the poller fleet 130 may utilize to determine that no further executions of the reduce task are necessary.

On receiving calls to the reduce task from the poller fleet 130, the frontend 120 distributes to the worker manager 140 the calls for executions of the reduce task, at (6). The worker manager 140 then initiates executions of the reduce task 404 to reduce the outputs of the map task executions, as included within the messages passed to the reduce task executions 404. Each reduce task may aggregate or otherwise reduce the outputs of the map task executions to produce a portion of a result of the MapReduce implementation. For example, where the map task executions produce counts of words within each portion of the data set processed by the map task executions, the reduce tasks may aggregate those counts to generate a count of each word across the data set. The reduced results produced by each reduce task execution can then be written to the data storage service 108 (or otherwise output to a receiving device).

While the illustration of FIG. 4A depicts retrieval of base data from the data storage service 108, in some embodiments the map task executions 402 may retrieve data from other sources. For example, the map task executions 402 may in some instances operate to retrieve data from the stream data processing system 170, such as data enqueued in another message stream. Moreover, while retrieval is described in FIG. 4A as occurring during execution of the map tasks 402, in some instances, base data may be retrieved or obtained independent of the map task execution itself. For example, a user device 102 may configure an additional message stream of the stream data processing system 170 as data source “triggering” execution of a map task, and designate messages within that additional message stream as a payload to be passed to a corresponding map task. Thus, publishing data to the additional message stream may initiate MapReduce processing based on that data. Illustratively, this processing may be used to “repartition” a set of data. For example, an organization may configure each of its end user devices to report usage metrics as messages to a message stream, which may invoke map and reduce tasks in accordance with the present disclosure to repartition the metrics according to other attributes (e.g., per-user, per-account-type, per-hour, etc.). Thus, description of the data as stemming from the data storage service 108 is intended to be illustrative.

To better illustrate the parallelization of interactions possible in accordance with the present disclosure, FIG. 4B includes an alternative depiction of the interactions of FIG. 4A, as they pertain to individual executions of the map and reduce tasks. As shown in FIG. 4B, the data flow can begin at (1), where individual map task executions 402A-N obtain input data. The input data may illustratively be a portion or chunk of a data set. In one embodiment, each map task execution 402 operates to select and retrieve the input data (e.g., to cause the system 110 to retrieve the data) from the data set.

Thereafter, at (2), the map task executions 402 process their input data (their portion of the data set) to “map” the input data according to a user-designated attribute. The executions 402 then output a metric for each attribute value. For example, in the case of a word count, the map task executions 402 map the input data onto counters (e.g., the metric) for each word (e.g., the attribute value) detected within the input data. The count value of each word within the input data may be considered one output of the map task execution. The output of the map task executions 402 illustratively represents preliminary results of the MapReduce model.

At (3), these preliminary results are distributed to the stream data processing system 170, and particularly to a set of message shards 174A-N on the system. In one embodiment, the individual map task executions 402A-N are unaware of the shards 174, and simply publish their preliminary results to a message stream 172 as messages (e.g., each message representing a count value for a word within the data set). The stream data processing system 170 obtains these messages, and distributes the messages among the shards 174. Beneficially, the partition key used by the system 170 may correspond to the attribute value used by the map task executions to map the input data. For example, in the case of a word count, the system 170 may divide messages between the shards 174 according to a first letter value of the count reflected in each message. Thus, counts of words staring with ‘a’ may be distributed to shard 174A, counts of words starting with ‘b’ may be distributed to shard 174B, etc. In this manner, the outputs of the map task executions 402 may be “shuffled” among the message shards.

At (4), the messages from the map task executions 402A are enqueued into the sub-streams 174A-N. Illustratively, the messages are enqueued in a first-in, first-out (FIFO) ordering, such that message 1 is at the head of the queue, followed by message 2, 3, etc. In some instances, the stream data processing system 170 may be configured to “guarantee” a FIFO ordering. In other instances, the system 170 may generally provide FIFO ordering, but not guarantee such ordering. Guaranteed ordering may be beneficial in instances where a reduce function is order dependent, but may also increase computing costs on the system 170.

At (5), individual executions of the reduce task 404A-N retrieve the messages from the sub-streams 174A-N. In one embodiment, the on-demand code execution system 110 is configured to provide a poller device 132 for each sub-stream 174, which poller device 132 retrieves one or more messages to from the sub-stream 174, and invokes at least one reduce task execution to process the one or more messages. At (6), the reduce task executions implement reduce functionality with respect to the received messages.

Generally, a reduce function is stateful, in the sense that the state of the reduce function after processing each message is dependent in part on a prior state of the reduce function after processing a prior message. For example, in the case of a word count functionality, a total count of each word can be updated based on a combination of a prior total count with the subtotal counts of each new message processed. However, in some embodiments, the on-demand code execution system 110 may generally operate statelessly, such that there is little or no guarantee that a particular invocation of a task will maintain state information from a prior invocation of the task. One possibility to enable stateful operation of the reduce task executions 404 would be to record, after each invocation of the task, the state of the execution 404 in an external data store (e.g., data storage services 108). However, such external recording of state information may greatly increase the computing resources needed for each reduce task execution 404. This may be particularly true in the instance that the external data store is resilient to failure, as a single recordation of state (itself an operation) may result in numerous other operations to duplicate that state.

To address these issues, the on-demand code execution system 110 may implement techniques to enable low-cost (in terms of compute resource usage) stateful operation of the reduce task executions 404. In one embodiment, the on-demand code execution system 110 may be configured to enable a task execution 404 to maintain state within its own execution environment (e.g., software container or virtual machine instance). Further, the on-demand code execution system 110 may be configured to continue to initiate subsequent executions of the same task in an existing execution environment, so long as such an environment exists (e.g., it has not been torn down by the system 110 to recapture the resources used by the environment). Thus, for example, a second call to execute the reduce task may result in an execution of the task in an execution environment previously used for a first execution of the task. In this manner, the second execution would gain access to state information for a first execution written within the execution environment.

In the context of the present disclosure, one difficulty with local storage of state information within an execution environment is the routing of messages from a particular sub-stream (e.g., sub-stream 174A) to a particular reduce task execution (e.g., execution 404A). Illustratively, the messages of each sub-stream are partitioned according to an attribute of the data, such as a staring letter, and thus can in theory represent authoritative information for the portion of the data sharing that attribute. For example, so long as the reduce task execution 404A views all messages within the sub-stream 174A, the execution 404A might be expected to provide an accurate count of all words in a corpus starting with the letter ‘a’. However, if messages from the sub-stream 174A are passed to multiple different executions 404, no individual execution 404 would hold an authoritative reduction of the messages in the sub-stream 174A. This would result in each reduce function holding incomplete information as to the underlying data set, harming implementation of the MapReduce model.

To address this concern, embodiments of the present disclosure may enable consistent routing of messages from a particular sub-stream 174 to a particular reduce task execution 404. Specifically, the on-demand code execution system 110 may enable the creation of multiple “copies” of a task, each associated with a particular instance identifier. From an invocation perspective, the system 110 may generally treat each copy of the task as independent. Thus, instance 1 may be invoked separately from instance 2, etc. Moreover, the system 110 may treat environments of each instance as related only to that instance. Thus, a call to instance 1 of task A might result in reuse of an environment from a past execution of instance 1, but not reuse of a past execution of instance 2 of the same task. In some instances, the system 110 may enable a new copy of a task to be created simply by transmitting a call to execute the task while passing a new instance identifier for the task. For example, a call to execute “task a, instance 1” may create a first copy of the task, a call to execute “task a, instance 2,” may create a second copy of the task, etc. The system 110 may treat task instances as a common task for purposes of task creation, storage, and administration, thus reducing computing resource usage relative to actual creation of separate tasks.

Under this configuration, each poller device 130 reading from a particular sub-stream 174 can be configured to transmit invocations to a particular instance of the reduce task, such as with an instance identifier corresponding to an identifier of the sub-stream 174. Because the system 110 can treat each instance separately, messages from sub-stream 174A might be passed only to instance A of the reduce task execution, messages from sub-stream 174B might be passed only to instance B of the reduce task execution, etc. Due to this “affinity” between sub-streams 170 and executions 404, each reduce task execution 404 can be enabled to utilize its locally maintained state information in processing reduce functionality. Moreover, because invocations between executions 404 can be expected to occur relatively rapidly (e.g., on a frequency based on the speed of operation of the map task executions 402A), the likelihood of reuse of an existing environment for an instance of the task may be quite high.

Additionally or alternatively to maintaining state environment within each execution environment of a task instance, the on-demand code execution system 110 may in some instances be configured to provide a low-cost mechanism for maintaining state information for a particular reduce task execution 404 outside of that execution 404. For example, each polling device 130 corresponding to a sub-stream 174 may be configured to maintain state information for a reduce task execution 404 corresponding to that sub-stream 174. Illustratively, the polling device 130 may obtain messages from the sub-stream, and submit each message (or set of messages) as a synchronous hypertext transport protocol (HTTP) call (e.g., a “POST” call). A reduce task execution 404A can be configured to return, in response to the HTTP call, its state information subsequent to processing the message of the call. The polling device 130 may then, in subsequent calls, pass both the message to be processed, as well as state information for a prior execution. In this way, the reduce task executions 404 can be considered stateless, in that each execution is provided, in the call, all information needed to achieve reduce functionality. Further details regarding maintaining of state information at a poller device 130 are provided within U.S. Pat. No. 11,099,917 to Hussels et al., entitled “EFFICIENT STATE MAINTENANCE FOR EXECUTION ENVIRONMENTS IN AN ON-DEMAND CODE EXECUTION SYSTEM” (the “'917 patent”). The entirety of the '917 patent is incorporated by reference herein.

After implementing reduce functionality of the received messages, the reduce task executions 404A-N output reduce results, at (7). The results may be written, for example, to a data storage service 108. In one embodiment, the reduce task executions 404 may write a reduced result output only after detecting that all map task executions have completed. For example, a map task execution that processes a final portion of the data set may write a “complete” message to the message stream 172, which may be propagated to each sub-stream 172. The reduce task executions 404, on detecting the complete message, may then write their outputs. In another embodiment, the map task executions 404 may write their results periodically, such as every n seconds, every n messages, or a combination thereof. These periodically results may beneficially be used as “checkpoints” for a task execution 404, such that if the execution 404 fails during operation, the execution 404 can be resumed from the previous results. Moreover, in such an implementation, propagation of a “complete” message to each sub-stream 174 may be unnecessary, as a single execution (corresponding to a single sub-stream 174) may observe the complete message, and note that results (e.g., written to a data storage service 108) are complete.

As can be seen from FIG. 4B, embodiments of the present disclosure enable implementation of the MapReduce programming model on an on-demand code execution system 110 and stream data processing system 170, without the need for a dedicated framework, coordinator function, or the like. Moreover, embodiments of the present disclosure enable highly-scalable parallelization of the MapReduce model without requiring that computing resources be dedicated to MapReduce implementation. Rather, operation of the on-demand code execution system 110 provides for simple parallelization of a map function, by invoking a desired number of executions of a map task. Operation of the stream data processing system 170 then provides corresponding parallelization of a reduce function, by automatically partitioning a message stream according to a partition key chosen to correspond to an attribute measured under the MapReduce model.

With reference to FIGS. 5-8, block diagrams will be described depicting illustrative routines for implementing the MapReduce programming model by use of tasks executing on the on-demand code execution system 110 and a stream data processing system 170 intermediary. Specifically, the routine 500 of FIG. 5 is an illustrative routine for implementing a map function on an the on-demand code execution system 110, the routine 600 of FIG. 6 is an illustrative routine for utilizing a stream data processing system 170 to divide outputs of the map functions into sub-streams according to an attribute of the outputs relevant to a reduce function, and the routine 700 of FIG. 7 is an illustrative routine for reducing the messages in each sub-stream via a reduce function implemented on the on-demand code execution system 110.

The routine 500 begins at block 504, where the on-demand code execution system 110 initiates a set of map task executions according to parallelization parameters specified by a user. The number of executions may be selected by a user, with more executions generally processing the data set more quickly while incurring greater compute resource usage. In one embodiment, a user device 102 may invoke each map task execution. In another embodiment, the user device 102 may specify parallelization parameters to the system 110, and the system 110 may invoke the map task executions.

The map task illustratively corresponds to code executable to select a portion of a data set being analyzed, to analyze that data set according to a map function, and to output each mapped result (e.g., each word count) to a message stream. Thus, at block 506, the executions of the map tasks process portions of the data set to result in the mapping of the portion. Often, the number of concurrent map task executions is not sufficient to process the entire data set concurrently. Thus, each map task execution may process multiple portions of the data set in serial. Additionally or alternatively, each map task execution may include code to call an additional execution of the map task on completion of processing a portion, thus creating serial processing of portions of the data set.

At block 508, the outputs of the map task executions (e.g., the mapped data) is published to a data stream. Illustratively, the outputs may be divided according to one or more attributes measured by the MapReduce model, such as “words” in the instance of a word count application. In some instances, the output may be divided according to multiple attributes. For example, in the instance that MapReduce is used to process logs regarding computing resource usage of an organization, each message may be divided according to type of application used and the region in which the use occurred. Any combination of attributes may be used to divide outputs of the map task executions, which attribute combinations may be specified within the code of the map task. Thereafter, the outputs of the map tasks may be passed to executions of a reduce task via the stream data processing system 170, thus implementing the MapReduce programming model. The routine 500 then ends at block 510.

While the blocks of FIG. 5 are described above sequentially for ease of reference, it should be appreciated that the routine 500 or portions thereof may be implemented in parallel (e.g., among multiple task executions), in serial (e.g., with respect to each task execution), or a combination thereof. Thus, a map task execution may repeat processing of data portions and publishing outputs multiple times, a first task execution may publish its own output while a second task execution is still processing its respective portion of the data set, etc.

The routine 600 of FIG. 6 is an illustrative routine for utilizing a stream data processing system 170 to divide outputs of the map functions into sub-streams according to an attribute of the outputs relevant to a reduce function. The routine 600 begins at block 604, where map outputs are received on a message stream as messages to the stream. The outputs may be received, for example, based on implementation of the routine 500 of FIG. 5, discussed above. The outputs are illustrative divided according to one or more attributes measured by the MapReduce model, such as “words” in the instance of a word count application. Thus, for example, each output may reflect a count of a particular word in a portion of a data set analyzed by a map task executions.

At block 606, the map outputs are divided between a set of sub-streams (or “shards”) according to at least one of the attributes reflected in the message. For example, in the case of a word count functionality, the starting letter (or multiple starting letters) of the word reflected in a message may be used to assign the message to one of a number of sub-streams. This attribute may thus represent a partition key for the message stream, used to partition messages among the sub-streams. At block 608, the divided messages are enqueued into their relevant sub-streams and made available to downstream devices (e.g., a device of the on-demand code execution system 110 executing a reduce task). The routine 600 then ends at block 610.

As discussed above, the number of sub-streams may vary according to a volume of messages on the stream (e.g., up to a user- or system-specified maximum). Thus, an end-user may not be required to specify a particular portioning of messages. Instead, the user may specify a partition key, and the system 170 may operate to divide the messages as appropriate between a current configuration of sub-streams. Moreover, by utilizing an attribute measured during the MapReduce implementation as a partition key for the message stream, and by applying a reduce function to each sub-stream, the “shuffling” of map outputs to reduce inputs is accomplished readily and without the need for a controlling framework or coordinator.

The routine 700 of FIG. 7 is an illustrative routine for reducing the messages in each sub-stream via a reduce function implemented on the on-demand code execution system 110. The routine 700 begins at block 704, where messages of each sub-stream of a message stream are obtained at the on-demand code execution system 110. The messages may be enqueued on the stream and divided between sub-streams, for example, based on implementation of the routine 600 of FIG. 6, which in turn may reflect outputs of map task executions based on implementation of the routine 500 of FIG. 5.

At block 706, messages of each sub-stream are passed to respective reduce task executions, to apply a reduce function to those messages. Each reduce task execution then, at block 708, applies a reduce function (e.g., an aggregation function) to the messages of the respective sub-streams. In one embodiment, a single reduce task execution processes the message of a single sub-stream. For example, a poller device 130 of the system 110 may operate to pull messages from a sub-stream and pass them to an execution of an instance of a reduce task, which instance is executed separately from other instances of the reduce task (e.g., used to process messages of other sub-streams). In this manner, multiple executions of the reduce task may be maintained separately. In other embodiments, multiple reduce task executions may be used to process messages from a sub-stream, with state information being passed between the executions, such that a final execution reflects the processing of all messages in the sub-stream. For example, in accordance with embodiments of the '917 patent, incorporated by reference above, a poller device 130 may pass each messages of a sub-stream to a reduce task execution along with state information reflecting a state of a reduce task execution which processed the prior messages of the sub-stream. Thus, a final reduce task execution can generate state information reflecting processing of all messages on the sub-stream.

At block 710, a reduction output (e.g., the result of a reduce function applied to the messages of a sub-stream) is output. The output may be provided, for example, to an user device 102, or written to a data storage service 108. In one embodiment, the output is provided on completion of processing of all messages in a sub-stream (e.g., based on detection of a “complete” message in the sub-stream). In another embodiment, the output is provided periodically during processing of messages on a sub-stream, and updated as additional messages are processed. The reduction outputs provided by each reduce task executions can collectively represent a result of the MapReduce model as applied to the data set. The routine 700 then ends at block 712.

All of the methods and processes described above may be embodied in, and fully automated via, software code modules executed by one or more computers or processors. The code modules may be stored in any type of non-transitory computer-readable medium or other computer storage device. Some or all of the methods may alternatively be embodied in specialized computer hardware.

Conditional language such as, among others, “can,” “could,” “might” or “may,” unless specifically stated otherwise, are otherwise understood within the context as used in general to present that certain embodiments include, while other embodiments do not include, certain features, elements and/or steps. Thus, such conditional language is not generally intended to imply that features, elements and/or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements and/or steps are included or are to be performed in any particular embodiment.

Disjunctive language such as the phrase “at least one of X, Y or Z,” unless specifically stated otherwise, is otherwise understood with the context as used in general to present that an item, term, etc., may be either X, Y or Z, or any combination thereof (e.g., X, Y and/or Z). Thus, such disjunctive language is not generally intended to, and should not, imply that certain embodiments require at least one of X, at least one of Y or at least one of Z to each be present.

Unless otherwise explicitly stated, articles such as ‘a’ or ‘an’ should generally be interpreted to include one or more described items. Accordingly, phrases such as “a device configured to” are intended to include one or more recited devices. Such one or more recited devices can also be collectively configured to carry out the stated recitations. For example, “a processor configured to carry out recitations A, B and C” can include a first processor configured to carry out recitation A working in conjunction with a second processor configured to carry out recitations B and C.

Any routine descriptions, elements or blocks in the flow diagrams described herein and/or depicted in the attached figures should be understood as potentially representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or elements in the routine. Alternate implementations are included within the scope of the embodiments described herein in which elements or functions may be deleted, or executed out of order from that shown or discussed, including substantially synchronously or in reverse order, depending on the functionality involved as would be understood by those skilled in the art.

It should be emphasized that many variations and modifications may be made to the above-described embodiments, the elements of which are to be understood as being among other acceptable examples. All such modifications and variations are intended to be included herein within the scope of this disclosure and protected by the following claims.

Claims

1. A system for implementing a MapReduce programming model comprising: a serverless code execution system comprising a set of computing devices configured to: obtain a request to analyze a set of data, wherein the request designates: a map task corresponding to code executable by the serverless code execution system to process a portion of the set of data to result in an output; anda reduce task corresponding to code executable by the serverless code execution system to process a plurality of outputs from individual executions of the map task to result in an aggregated output;initiate individual executions of the map task to process respective portions of the set of data and to publish outputs of the individual executions into a partitioned message stream, the partitioned message stream comprising a plurality of sub-streams, individual sub-streams containing messages according to values of an attribute of the messages designated as a partition key, wherein the individual executions publish the outputs to the partitioned message stream without regard, by the individual executions, to the plurality of sub-streams;a streaming data processing system comprising a set of computing devices configured to: receive a collective output representing outputs of the individual executions of the map task in aggregate, wherein the collective output represents the outputs of the individual executions of the map task as messages to be published into the partitioned message stream;divide the messages among the plurality of sub-streams within the partitioned message stream, wherein dividing the messages comprises, for individual messages received to be published into the partitioned message stream: identifying a-sub-stream, of the plurality of sub-streams, according to an individual message's value for the attribute; andenqueuing the individual message onto the sub-stream identified according to the message's value for the attribute;wherein the set of computing devices of the serverless code execution system are further configured to: for individual sub-streams of the plurality of sub-streams: detect one or more messages within the individual sub-stream; andresponsive to detecting the one or more messages within the individual sub-stream, instantiate an execution of the reduce task on the serverless code execution system, the individual execution of the reduce task aggregating the outputs of the individual executions of the map task corresponding to the one or more messages of the individual sub-stream to result in an aggregate result; andoutput the aggregate results of the individual executions of the reduce task.

2. The system of claim 1, wherein the set of computing devices of the serverless code execution system comprises a plurality of poller devices, individual poller devices from the plurality of poller devices corresponding to individual sub-streams of the plurality of sub-streams, and wherein the set of computing devices of the serverless code execution system are configured to initiate executions of the reduce task for individual sub-streams of the plurality of sub-streams at least partly by: at an individual poller device, retrieve messages from the sub-stream corresponding to the individual poller device; andsubmit calls to the serverless code execution system to initiate an execution of the reduce task to process the retrieved messages.

3. The system of claim 2, wherein the individual poller device is configured to: obtain, in response to an individual call, state information reflecting a state of the execution of the reduce task after processing a message submitted with the individual call; andpass in a subsequent call the state information and an additional message from the retrieved messages.

4. The system of claim 1, wherein the code of the map task is further executable by the serverless code execution system to select the portion of the set of data processed during an execution of the map task.

5. A method implemented by a computing system comprising: obtaining a set of outputs published to a data stream, wherein the set of outputs represent results of application of a map function to a data set, wherein individual outputs reflect respective values of an attribute of the data set, wherein the set of outputs are published to the data stream without regard, by one or more publishing entities, to a plurality of sub-streams within the data stream;dividing the set of outputs among the plurality of sub-streams within the data stream, wherein dividing the set of outputs comprises, for individual outputs of the set of outputs published to the data stream: identifying a sub-stream, of the plurality of sub-streams within the data stream according to the individual output's value for the attribute; andenqueuing the individual output onto the sub-stream identified according to the individual output's value for the attribute;for individual sub-streams of the plurality of sub-streams: responsive to presence of one or more outputs within the sub-stream, instantiating an execution of a reduce function on a serverless code execution system, the execution of the reduce function aggregating the outputs of an individual sub-stream to result in an aggregate result; andoutputting the aggregate results of the executions of the reduce function.

6. The method of claim 5, wherein the map function is implemented as code executable by the serverless code execution system in response to an invocation of a map task.

7. The method of claim 6 further comprising initiating a plurality of invocations of the map task to process the data set at least partially in parallel, wherein the plurality of invocations results in a plurality of executions of the map task, which executions represent the applications of the map function to the data set.

8. The method of claim 7, wherein the data set includes a set of messages published to a second data stream, and wherein initiating the plurality of invocations of the map task includes passing individual messages from the second data stream to individual invocations of the map task as payload data.

9. The method of claim 7, wherein individual executions of the map task are implemented within distinct execution environments on the serverless code execution system, and wherein an individual execution of the map task causes the serverless code execution system to: select a portion of the data set to be processed by the individual execution of the map task;apply the map function to the portion of the data set; andpublish outputs of the map function to the data stream.

10. The method of claim 9, wherein subsequent to applying the map function to the portion of the data set, the individual execution of the map task causes the serverless code execution system to at least one of i) invoke an additional execution of the map task; or ii) select an additional portion of the data set to be processed by the individual execution of the map task.

11. The method of claim 5, wherein a number of sub-streams within the plurality of sub-streams is determined by a stream data processing system based at least partly on a volume of data within the data stream.

12. The method of claim 11, wherein division of the set of outputs among the plurality of sub-streams comprises applying a hashing operation to each individual output's respective value for the attribute and according to the number of sub-streams.

13. The method of claim 5, wherein instantiating executions of the reduce function on the serverless code execution system for an individual sub-stream of the plurality of sub-streams comprises invoking an instance of a reduce task on the serverless code execution system, the instance of the reduce task maintaining state information regarding application of the reduce function to outputs within the individual sub-stream.

14. Non-transitory computer-readable media comprising instructions executable by a computing system to: obtain a set of outputs published to a data stream, wherein the set of outputs represent results of application of a map function to a data set, wherein individual outputs reflect respective values of an attribute of the data set, wherein the set of outputs are published to the data stream without regard, by one or more publishing entities, to a plurality of sub-streams within the data stream;divide the set of outputs among the plurality of sub-streams within the data stream, wherein dividing the set of outputs comprises, for individual outputs of the set of outputs published to the data stream: identifying a sub-stream, of the plurality of sub-streams within the data stream according to the individual output's value for the attribute; andenqueuing the individual output onto the sub-stream identified according to the individual output's value for the attribute;for individual sub-streams of the plurality of sub-streams: responsive to presence of one or more outputs within the sub-stream, cause an execution of a reduce function on a serverless code execution system, the execution of the reduce function aggregating the outputs contained within an individual sub-stream to result in an aggregate result; andoutput the aggregate results of the executions of the reduce function.

15. The non-transitory computer-readable media of claim 14, wherein the instructions are executable by the computing system to cause the execution of the reduce function on the serverless code execution system at least partly by: retrieving a subset of the one or more outputs within the individual sub-stream; andfor individual outputs within the subset of outputs, calling for execution of the reduce function to process an individual output.

16. The non-transitory computer-readable media of claim 15, wherein calling for execution of the reduce function to process the individual output comprises making a synchronous hypertext transport protocol (HTTP) call to the serverless code execution system, the synchronous HTTP call comprising the individual output.

17. The non-transitory computer-readable media of claim 16, wherein the synchronous HTTP call further comprises state information for a prior execution of the reduce function generated based on a prior output of the subset of outputs.

18. The non-transitory computer-readable media of claim 15, wherein calling for execution of the reduce function to process the individual output comprises calling for execution of an individual instance of a reduce task on the serverless code execution system, the individual instance of the reduce task being executed by the serverless code execution system in an execution environment distinct from a second execution environment used to execute a second instance of the reduce task.

19. The non-transitory computer-readable media of claim 14, wherein the instructions are executable by the computing system to output the aggregate results at least partly by writing the aggregate result of each execution of the reduce function to a common data storage location.

20. The non-transitory computer-readable media of claim 14, wherein a number of sub-streams within the plurality of sub-streams is determined by a stream data processing system based at least partly on a volume of data within the data stream.

21. The non-transitory computer-readable media of claim 20, wherein the stream data processing system is configured to divide the outputs among the plurality of sub-streams at least partly by applying a modulus division operation to each individual output's respective value for the attribute and according to the number of sub-streams.