Big Data Processing: Serverless MapReduce on Azure

Introduction into MapReduce

MapReduce is a programming model that allows processing and generating big data sets with a parallel, distributed algorithm on a cluster.

A MapReduce implementation consists of a:

  • Map() function that performs filtering and sorting, and a

  • Reduce() function that performs a summary operation on the output of the Map() function

Both the input and output of the MapReduce implementation are <key, value> pairs: (input) <k1, v1> -> map() -> <k2, v2> -> combine() -> <k2, v2> -> reduce() -> <k3, v3> (output)

The model is a specialization of the split-apply-combine strategy for data analysis and allows for significant increased speed in a multi-threaded implementation.

Introduction into Durable Functions

Durable Functions is an extension of Azure Functions and Azure WebJobs that allows writing long-running, stateful function orchestrations in code in a serverless environment.

This extension enables a new type of function called the orchestrator function that allows you to do several new things that differentiates it from an ordinary, stateless function:

  • They are stateful workflows authored in code. No JSON schema or designers.
  • They can synchronously and asynchronously call other functions and save output to local variables.
  • They automatically checkpoint their progress whenever the function awaits so that local state is never lost if the process recycles or the VM reboots.

Durable Functions: Fan-out/fan-in

One of the typical application patterns that can benefit from Durable Functions is the fan-out/fan-in pattern.

With normal functions, fanning out can be done by having the function send multiple messages to a queue. However, fanning back in is much more challenging. You'd have to write code to track when the queue-triggered functions end and store function outputs. The Durable Functions extension handles this pattern with relatively simple code:

public static async Task Run(DurableOrchestrationContext ctx)
    var parallelTasks = new List<Task<int>>();

    // get a list of N work items to process in parallel
    object[] workBatch = await ctx.CallActivityAsync<object[]>("F1");
    for (int i = 0; i < workBatch.Length; i++)
        Task<int> task = ctx.CallActivityAsync<int>("F2", workBatch[i]);

    await Task.WhenAll(parallelTasks);

    // aggregate all N outputs and send result to F3
    int sum = parallelTasks.Sum(t => t.Result);
    await ctx.CallActivityAsync("F3", sum);

In the example above, the fan-out work is distributed to multiple instances of function F2, and the work is tracked by using a dynamic list of tasks. The .Net Task.WhenAll API is called to wait for all of the called functions to finish. Then the F2 function outputs are aggregated from the dynamic task list and finally passed on to the F3 function.

The automatic checkpointing that happens at the await call on Task.WhenAll ensures that any crash or reboot midway through does not require a restart of any already completed tasks.

Note: You can read more about Durable Functions here: /en-us/azure/azure-functions/durable-functions-overview


Before you open or build the solution in this repo, make sure you have the following:

If you don't (want to) install Visual Studio and deploy via 'Publish', check the Visual Studio-less deployment environment setup section below to see if you need to run the setup script included in this repo to grab other bits necessary for automatic deployment.

Visual Studio-less deployment environment setup

You can set up your PC environment to enable deployment of this solution without installing Visual Studio by executing the SetupEnvironment.ps1 PowerShell script from an Administrator PowerShell instance.

Note: If you've never run PowerShell scripts on your computer, you'll need to change the Execution Policy to allow script running by executing Set-ExecutionPolicy Bypass -Scope Process -Force before running the above command. Using -Scope Process here ensures ExecutionPolicy isn't permanently set to allow scripts on your machine, but rather just while you're doing the import

This performs the following permanent changes to your machine:

1. Serverless MapReduce on Azure

The above diagram shows the architecture a MapReduce implementation generally follows.

In this sample you will find handful of Azure Functions comprising a Mapper and Reducer to determine the average speed of New York Yellow taxi trips, per day over all of 2017.

You will use Durable Functions - specifically the Fan-out/Fan-in pattern - to distribute the workload across multiple instances of the Mapper then coordinate the outputs in to the Reducer and finally return the computed values for each day.

The source data you will be using for this MapReduce implementation can be found here: and contains 12 CSV files of roughly 800MB each (Yellow dataset) you need to make available in Azure Blob storage - total size is ~9.6GB.

2. Hands on: Implementing the MapReduce pattern step by step

2.1 Copy the dataset to an Azure Blob Storage instance

Open a new PowerShell window & execute TaxiDataImporter.ps1 from the repo directory to copy each file from the NYC Taxi site in to your Azure Storage Account

You'll be asked to provide the following:

  • A connection string to your Azure Storage account
  • The container name within the storage account into which to copy the data

Note: If you wish to specify a subfolder, you'll need to execute the cmdlet with the -subfolderName parameter; it won't prompt you for it because it's optional. You can get examples of usage with Get-Help Import-TaxiDatasetToAzureStorage

The copy will kick off and should only take a couple of minutes as you're using Azure's bandwidth to do the downloading, not your own.

2.2 The code

Open ServerlessMapReduce.sln in Visual Studio You'll first notice there are two projects in the solution. One is a Function v2 (.Net Standard) project, the other is a Function v1 (.Net 4.x aka "netfx") project. The code for each project is identical, and shared between the two via linked files.


This is where all our MapReduce logic lies. Let's have a look


The entry point to our Durable Orchestration, this method is triggered by an HTTP request to the Function which contains a single path query parameter specifying the URL to the blob storage account containing the files to process. Example:

POST /api/StartAsync?code=Pd459lsir2CILjc8jRAkO6TLy3pasuBDikYZMZRKAjaTgjh00OW2wg==&path= HTTP/1.1
Cache-Control: no-cache

Note the format of the path variable. It's not only used to denote the container in which to look, but also the prefix to use when searching for the files. You can, therefore, get as arbitrary or specific as you want but if you have the files in a subfolder you must specify it.

Once StartAsync is kicked off, it parses out the container name and prefix from the path parameter and kicks off a new orchestration with BeginMapReduce as the entry point


This is the actual orchestrator for the entire process. First, we retrieve all the blobs from the storage container which match the prefix, using the Activity function GetFileListAsync. We must do this as the queries to Blob Storage are asynchronous and therefore cannot live inside an orchestrator function.

After getting the list of files to include it then spins up a mapper for each, in parallel:

var tasks = new Task<double[]>[files.Length];
for (int i = 0; i < files.Length; i++)
    tasks[i] = context.CallActivityAsync<double[]>(

We add the resulting Task<T> object from these calls to an array of Tasks, and we wait for them all to complete using Task.WhenAll()

Once they've completed, we've got mappers created for each file and it's time to reduce them. We do this by calling out once to another activity function: Reducer. This function does the math to aggregate the average speed computed for each day of the week in each line of the files in to an overall average across all the files for each day of the week.

After this, we return the result as a string back to the Orchestrator (BeginMapReduce) who sets this as output for the entire orchestration which the caller can discover by issuing an HTTP GET to the status API:

GET /admin/extensions/DurableTaskExtension/instances/14f9ae24aa5945759c3bc764ef074912?taskHub=DurableFunctionsHub&amp;connection=Storage&amp;code=ahyNuruLOooCFiF6QB7NaI6FWCHGjukAdtP/JGYXhFWD/2lxI9ozMg== HTTP/1.1

    "instanceId": "14f9ae24aa5945759c3bc764ef074912",
    "runtimeStatus": "Completed",
    "input": {
        "$type": "<>f__AnonymousType0`2[[System.String, mscorlib],[System.String, mscorlib]], ServerlessMapReduce.v1",
        "containerUrl": "",
        "prefix": "2017/yellow_tripdata_2017-1"
    "customStatus": null,
    "output": "Sun: 14.7773480317865, Mon: 13.412296342264, Tue : 12.0879698292105, wed: 12.1794013570501, Thu: 12.0085449069425, Fri: 12.4839412962461, Sat: 12.8889347824278",
    "createdTime": "2018-08-29T17:52:41Z",
    "lastUpdatedTime": "2018-08-29T18:00:42Z"

Note: Give this Status API a hit while the orchestration is running and you'll get an idea of where it's at in the process due to the calls to SetCustomStatus throughout the code


There are two ways to choose from to deploy this solution to your own Azure subscription:

Visual Studio

Right-click either/both of the project(s) and choose 'Publish...'

  • Walk through the wizard to create
    • App service plan
    • Storage account
    • Function app instance


Important: Run deploy.ps1 in a new PowerShell window only after you've executed SetupEnvironment.ps1. This refreshes environment variables so the build & deploy commands will execute successfully

Execute deploy.ps1 providing:

  • Subscription ID
  • Base name for resources

Note: By default all resources will be provisioned in the West US 2 region of Azure. If you wish to have them somewhere else, provide the -region parameter to the deployment script

and the following will be provisioned for you

  • New Resource group
  • 2 Application Insights instances (one for each of v1 and v2 Function apps)
  • 2 Storage accounts
  • 1 Consumption App Service Plan
  • 2 Function Apps

Naming of the Function App and Application Insights instances will follow (base name)+v1/v2 Naming of the Storage Accounts will follow (base name)+storv1/storv2

Execution / testing

After deployment:

  • Visit your Function App in the Azure Portal

  • Click the StartAsync function

  • Click 'Get function URL' & copy it for usage in your favorite REST API testing program

  • Issue an HTTP POST to that endpoint with the path parameter populated from the output of the PowerShell script you ran in 2.1

You'll receive back a list of URLs you can use to check status, issue new events (not handled by this sample), or terminate the orchestration.


The implementation shown here utilizes a single reducer. If you needed multiple reducers, you would create a sub-orchestration per reducer, launch those in parallel, do the CallActivityAsync() calls within the sub-orchestrator, reduce the results, pass that up to the parent orchestrator to further reduce the results, and so on.

It's also important to remember while Serverless technologies allow you to scale "infinitely" we must use the right tool for the right job. Durable Functions will, in theory, scale out to run any number of jobs in parallel and come back to the reduce step so this approach may work very well for loads that can be highly parallelized however the machines on which Azure Functions run (in Consumption plan) are of limited specs (more detail can be found here in the docs). This means this approach may not work well for loads with very large sections to be processed by each mapper. In this case, you may look in to hosting this implementation on an App Service Plan with large VMs or, even better, a Functions Premium offering with a large VM size to process the data more quickly. Another thing to note is the way DF passes results and parameters around; these are done via Azure Storage Queues which may/may not add unacceptable latency in to your big data process.