Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Software License PRs Welcome unofficial Google Analytics for GitHub


A simple architecture to consume and process messages coming from video game clients (or servers) using the following Azure services:

  • Event Hubs
  • Functions
  • Cosmos DB
  • Data Lake Store
  • Data Lake Analytics

Click here to deploy to Azure

Once the deployment finishes, you need to connect the independent resources (TO BE DOCUMENTED).


The scenario is based on a hypothetical multiplayer online game. Gamers connect to multiple game servers and compete among themselves. Gamers can have a 'win' or a 'loss'. You can think of wins/losses as actions/achievements one player might have over another. E.g. in a first-person shooter, win could be a 'kill' and loss could be a 'death'.


After you deploy the resources on Azure and finalize the connections among them, you can use the sample Node.js app we have on the 'client folder'. First, you need to configure a '.env' file on the client folder that contains the following necessary environment variables

REGISTERGAME_FUNCTION_URL=url for your deployed 'register game' Azure Function

Then, you can run 'dbCreate.js' to create the database in your Cosmos DB account and then run 'sendData.js' to send some data. Feel free to modify the related variables in the 'sendData.js' to your liking.




  • Game session: a single round of a multiplayer game
  • Wins: a 'win' of a user versus another one. This could be a 'kill' in a First Person Shooter, a bypassing in a racing game, a goal in a soccer match
  • Losses: the exact opposite of the win, depending on the game scenario
  • winnerID: the user that accomplished the win
  • loserID: the user that suffered the loss from winnerID

Data flow

  • Game server registers the game by calling the registergamesession Function. Each game has a sessionID which is formattted like year-month-day_GUID. Game server also submits the list of players that will participate in this game session. For this demo, we suppose that users cannot enter the game after it has began.
const gameDocument = {
                gameSessionID: string, //the game SessionID, its format is "DATE_GUID" where date is YEAR-MONTH-DAY
                type: string, //random game type, for the demo its format is "type" + random integer
                map: string, //random map, for the demo its format is "map" + random integer
                players: array, //array of player objects
                startDate: number //time string in UTC format (e.g. 2018-04-03T06:19:25.151Z), denotes time the game session started
const player = {
        playerID: string, //playerID
        playerCountry: string //country the player connects from
  • Game server sends messages to Event Hub using the format:
const event = {
                    eventID: string, //a unique event ID, its format is "GUID_gameSessionID"
                    gameSessionID: string, //the game session in which this event took place
                    winnerID: string, //playerID of the winner
                    loserID: string, //player ID of the loses
                    special: string, //special attribute(s) for this win, like 'Low health'
                    eventDate: number //time string in UTC format (e.g. 2018-04-03T06:19:25.151Z), denotes time the event took place
  • Event Hub triggers the dataingesthot Function which receives the messages in batch. Messages are aggregated and sent to Cosmos DB (hot path) whereas a copy of them is stored on Blob Storage.
  • Blob Storage triggers the dataingestcold Function which receives the batched messages. Messages are finally ingested into Data Lake Store (cold path) without any processing.
  • Game server or client can call statistics Function passing the gameSessionID as argument and get game session related data
  • External service can use Data Lake Analytics jobs to query data in Data Lake Store (there is a relevant .usql script on the various folder). The output of these jobs can be ingested into other services (via Data Factory) or be used directly from a visualization platform such as PowerBI.

Functions' details


This Function is used to register a game session and store relevant metadata. Data is stored into Cosmos DB as well as two CSV files in Data Lake Store. On the first file, we store details about each unique game session. The filename is '/YEAR/MONTH/DAY/gamesessions.csv' whereas the data that's being appended to it has the format:

const line = `${gameDocument.gameSessionID},${gameDocument.type},${},${gameDocument.startDate}\n`;

The second file is '/YEAR/MONTH/DAY/playerspergamesession.csv'. On this file, we store data about each player that will participate in the current game session. Data has the format:

const line = `${gameDocument.gameSessionID},${player.playerID},${player.playerCountry}\n`;


This Function is triggered by Event Hub. It fetches messages in batch, stores an aggregation on Cosmos DB, like this:

    "gameSessionID": "2018-6-4_d2702447-b14d-42ee-849e-04fb3e5ee387",
    "type": "type3",
    "map": "map7",
    "startDate": "2018-06-04T08:17:19.980Z",
    "documenttype": "metadata",
    "players": [
            "playerID": "player_8zjM3",
            "playerCountry": "ESP"
            "playerID": "player_iI1jV",
            "playerCountry": "UK"
            "playerID": "player_c4s68",
            "playerCountry": "UK"
            "playerID": "player_Y39XX",
            "playerCountry": "FR"
            "playerID": "player_moZH9",
            "playerCountry": "ESP"
            "playerID": "player_f5eq6",
            "playerCountry": "ESP"
            "playerID": "player_NIXcA",
            "playerCountry": "FR"
            "playerID": "player_AyMtm",
            "playerCountry": "FR"
    "id": "49397a1f-d55c-edf6-f48f-8afce816efbe",
    "_rid": "oN5oALJEJwAHAAAAAAAAAA==",
    "_self": "dbs/oN5oAA==/colls/oN5oALJEJwA=/docs/oN5oALJEJwAHAAAAAAAAAA==/",
    "_etag": "\"0000892b-0000-0000-0000-5b1576bf0000\"",
    "_attachments": "attachments/",
    "_ts": 1528133311

Moreover, this Function stores a copy of the batched messages on Blob Storage, on the container cold eventbatches.


This Function is triggered by Blob Storage. It fetches the batched messages and appends their data to Data Lake Store. Filename is '/YEAR/MONTH/DAY/gameevents.csv' and data is appended in CSV format, with each line containing having the following format:

const line = `${message.eventID},${message.gameSessionID},${message.winnerID},${message.loserID},${message.special},${message.eventDate}\n`;


This Function is an HTTP triggered one. It accepts the gameSessionID as an argument and returns details (players, wins, losses, etc.) about the specific gameSession by connecting to Cosmos DB and fetching/aggregating the relevant documents.


Is this the best architecture / solution?

Of course not. It always depends on your requirements. For example, you could swap Event Hubs with Kafka, Functions with Stream Analytics, Data Lake Analytics with Databricks or HDInsight. This is one implementation of a data streaming and processing pipeline, it can certainly work well and scale, however you are encouraged to modify parts of this solution towards your objectives. For an overview of Data architectures on Azure, check here.

Are messages meant to be sent from the game client or server?

It depends. Clients can hack (especially on PC) the messages and send malicious data to your Event Hubs. You wouldn't want to have users cheat on their ranking, right? The best approach would be to have the game server send the important messages (e.g. user1 killed user2) whereas game clients can send less important messages, like behavioral ones (e.g. in an adventure game, user's response to the question was the second one or it took 2' for the user to solve this puzzle or reach the goal).

How can I modify the communication between Event Hubs and Azure Functions?

You should check the parameters listed here. Moreover, you are encouraged to read here regading Functions' scaling when using the Event Hubs trigger. Finally, check here for performance best practices for Azure Functions.

Are you doing anything specific to optimize parallel data ingestion into Data Lake Store?

Yes, we are using the ConcurrentAppend method of the Data Lake Store API, check here for details. For more information on how to fine-tune Azure Data Lake for performance, check here.

I want to read more information regarding Event Hubs.

Check out the excellent documentation here. For techniques on how to optimize event ingestion and consumption, check here. For in order event processing with Event Hubs and Azure Functions check here whereas you can check here for reliable event processing. For a comprehensive article that compares Event Hubs to other Azure messaging services, check here.

Can the blob retrieval from the dataingestcold Function be any faster?

Yes, you could use Event Grid trigger instead of Blob Storage trigger.


Big data reference architecture and implementation for an online multiplayer game








No releases published


No packages published