A good game is often built on the solid feedback from playtesters. The tighter the feedback loop for a game, whether it be in development or production, the faster the developers can collect and iterate on feedback and bug reports.

Most game devs in the beginning will do periodic builds after long periods of isolated development, asking their testers to download the whole game repeatedly to test fixes and updates. While this works initially, this takes signifigant focus from the developer to create builds and distribute them manually to the testers. This methodology is also flawed when the game is released. Bugs will inevitably pop up: telling players who have already spent time to download a potentially large (1-5GB+) game each time there is a minor update is simply unacceptable. For indie developers, you could probably use Steam or similar services to patch and deploy new changes; however, many of these services can be prohibitively expensive for small indie studios to use. For doujin developers who may or may not be creating derivative works without the luxury of these services, we’re more or less on our own: there is no exisitng support structure for this kind of service.

To remedy these problems, we will apply the principles of continuous integration and continuous deployment. The end goal is to produce an automated pipeline: when a developer makes a change, the game is built, tested, and deployed to player’s desktops and browsers. Spending time to manually do all of the following is simply draining time you could be spending improving your game.

Basic Setup

At it’s core, this automated pipeline needs four things: a queryable source control repository, an automated build system, a deployment web server , and a client side launcher/updater.

The automated build system will listen to source control for changes. When notified, it will automatically start building your game in accordance to your specifications. If the build and tests succeed, the goal is to notify the deployment server to make the built artifacts publicly available. The client launcher then checks if the game files are outdated at launch, and updates the game as needed.

Source Control

It all begins with the source control. Any good developer worth their salt, makes sure to version control their work. I won’t go too much into detail as there are entire books and methodologies written about good source control. For the sake of this article, the only hard requirement is that the source control service either can automatically notify 3rd parties about pushed updates, or can be periodically queried for changes. There are plenty of services out there that can satisfy requirement: GitHub, BitBucket, GitLab, AWS CodeCommit, etc. Simply choose one that suits your needs and push your source to a repository. For my case, my project lives in a Github repository.

NOTE: Most of these services are free for open source software, while putting a limit on private use. If you wish to keep your game closed source, there will be costs involved at this stage.

Automatic Builds (Continuous Integration)

The next step is to trigger automatic builds based on source control changes. This part also usually involves simply choosing the right service for your game. There is AppVeyor (Windows) and TravisCI (macOS and Linux) for free and open source projects. If you cannot use either for whatever reason, you can always roll your own instance via Jenkins or BuildBot. For Unity3D games, I strongly recommend Unity Cloud Build as setting up a Unity build enviroment in any of these other CI systems can be very slow and costly.

NOTE: Again most of these services are free for open source software, if your game is closed source, be prepared to pay up some money, or you may need to do additional configuration so that these services have the proper authentication to access your source control.

Automatic Deploymment

You now have a automatic stream of game builds coming in now. Now comes the hard part: exposing these files in a public manner for the client to query against. To my knowledge, there is no public free service that does this: so I rolled my own. HouraiDeploy is a tiny Flask web app that does the following:

  • Expose a HTTP endpoint for receiving notifications from automatic deployment
  • On being notified, it either:
    • Clones a copy of the source control for itself, and rebuilds the project locally using the same build process used in the CI build. This is for security, as it ensures others cannot upload malicious “build artifacts” that ultimately will end up executed on player’s computers.
    • Download copies of build artifacts from trusted parties. This is usually used when the build process is either too complex or too computationally intensive to do locally (i.e. Unity3D builds).
  • Upon obtaining the files, it places them in a location accessible by a nginx static file server, which then serves the files to remote clients.
  • It then generates an index file for clients to query. More details down below.

For added security, HouraiDeploy also checks for an privately generated access token in the HTTP request. These tokens are only stored on the host itself, and as encrypted enviroment variables on the continuous integration system’s end.

Generating a Index

The updater works effectively by checking each individual file against a remote index. This index includes the file’s path, size, and SHA-256 hash (though any hash would work, it’s advised to use SHA-2 or better. As MD5 and SHA-1 have been known to be cracked and/or faked before). The hash is a quick and simple way to check if files are identical. If the file exists at a path exists locally, and has a matching size and file hash, it’s almost 100% guarenteed that the files are identical, and thus does not need to be redownloaded. If the hash changes, we simply update the file by downloading the remote version and replacing the local one with it. For HouraiDeploy, we get the file hashes like so:

from hashlib import sha256
def hash_file(path, block_size=65536):
  hahser = sha256()
  with open(path, 'r') as file_obj:
    for block in iter(lambda: file_source.read(block_size), b''):
      hasher.update(block)
  return hasher.hexdigest()

This method also minimizes the amount of data needed to update the game: a index with ~120 files, irrespective of file size, is only 20kb, saving you the bandwidth cost and the player’s time from needing to serve or download unnecessary data.

We generate one of these indexes for the game each time a new build is completed, and it is saved as a file right next to the game data.

An example of such a index file can be seen here.

Client Side Launcher

This one is a tough one: we want to create an efficent cross-platform downloader with a responsive UI that can work with the index and files we generate. Obviously nothing will work off the bat with the freshly created Hourai Deploy, so I created Hourai Launcher, a sister project that reads Hourai Deploy compatible index files and updates local files. It’s created using Python: PyQt provides a cross platform GUI, and PyInstaller creates standalone executables so that players don’t need to have a python runtime to run it. At it’s core Hourai Launcher is a script for downloading and updating the game, everything else is to get it to work nicely with a responsive UI. A more barebones update script can be done as follows:

# For downloading files to disk
def download_file(url, file_path, block_size=1024*1024):
  with open(file_path, 'wb+') as file:
    # Remember to stream this
    reqponse = requests.get(url, stream=True)
    for block in iter(lambda: file_source.read(block_size), b''):
      file.write(block)

# Index local files
game_index = dict()
replacement = GAME_BASE_DIR + os.path.sep
for directory, _, files in os.walk(GAME_BASE_DIR):
  for file in files:
    full_path = os.path.join(directory, file)
    relative_path = full_path.replace(replacement, '')
    # Use same hashing function mentioned earlier
    game_index[file] = hash_file(full_path)

# Fetch remote index using requests library
import requests
index_json = requests.get(INDEX_URL)

# Download missing or improper files
for file, data in index_json['files']:
  url = build_url(game_index, data)
  path = os.path.join(GAME_BASE_DIR, file)
  if file not in game_index:
    download_file(url, path)
  elif game_index[file] != data['sha256']:
    download_file(url, path)

# Remove unneeded files
for file, hash in game_index:
  if file not in index_json['files']:
    os.remove(os.path.join(GAME_BASE_DIR, file)

Bonus: There’s some extra space available on the UI, so I added a news list to help list the news regarding the game from a RSS feed.

Refinement and Improvement

Now that we have a basic system set up, there are a number of smaller problems that should probably be addressed to ensure the system runs as smoothly as possible.

Supporting Multiple Branches

Why would we want to support multiple branches? Ask Blizzard: Overwatch and many other games have a closed beta, open beta, endless beta, public test region (PTR), etc. This allows you to isolate different deployment enviroments and lets players have multiple installations on your machine: a publicly released version, or the latest test beta. This supports development workflows like Gitflow where there are multiple historical branches of development, which are known to help avoid introducing momentary bugs when creating new features.

This is relatively simple, additional space and directories simply need to be allocated on both the players’ machines and on the deployment host, with additional configuration to Hourai Launcher and Hourai Deploy for this. When multiple branches are configured, the UI will expose a selector box for choosing which version of the game to use.

Caching and Cache Busting

There is one large caveat to our current setup: it can become exceptionally bandwidth heavy quickly, even with compression. A 5GB game with 1000 clean installs is 5TB of bandwidth used. This easily can get to $100 to $1000+ per month for just bandwith (depending on where and how you are hosting your game files).

The solution is to use a CDN (Content Delivery Network) to cache the game files. Instead of hammering your file server with heavy requests, requests for already cached files end at the CDN, signifigantly reducing For added benefit, good CDNs geographically replicate your data, meaning download speeds are more uniform, regardless of geographic location. Currently at time of writing, HouraiDeploy assumes the static file server is behind CloudFlare: a free CDN service that doubles as DDoS protection. In the case of Hourai Teahouse, we set up the Page Rules for https://patch.houraiteahouse.net to very aggressively cache files:

  • Cache Everything - every file served from the domain is cached
  • Browser Cache TTL: 1 year - browsers and clients should keep the file cached for a full year, and assume the file has not changed.
  • Edge Cache TTL: 1 month - If a file is cached on CloudFlare’s servers, it will not check for a change in the file until a month later.

The combination of these three drastically reduce the amount of bandwidth used: CloudFlare will likely fetch a file once and only once from the host. All other requests will be handled entirely by CloudFlare itself.

However, these aggressive caching rules presents a problem: updated files are not served if they are on the same path. For example, if the base binary fc.exe was fetched and cached on June 1st 2017, but then updated 7 days later, CloudFlare will assume nothing has changed until July 1st 2017, serving the old fc.exe, despite the base file changing. There are two ways of dealing with this issue: cache invalidation and cache busting.

Cache invalidation involves explicitly telling CloudFlare or any other CDN that the cached file is now invalid, forcing it to refetch and recache the file. CloudFlare has explicit REST API endpoints for this purpose. There are some caveats to this: this endpoint is limited to 2000 calls per 24 hours, so it’s not well suited for doing purges of large groups of files. It is however, incredibly useful for invalidating fixed API endpoints, like the generated index.json file. Hourai Deploy actively uses this calls this endpoint to invalidate these kinds of files.

If cache invalidation does not work on large groups of files, cache-busting does a much better job. The idea is not to invalidate the files server side, but rather force the client to fetch from a location that will intentionally miss the cache. To do this, we simply need to alter the URL for a file to include a static “change identifier”. This identifier should be the same so long as the binary content of the file is the same, and should be drastically different if there is even a small change in the content. A simple solution is to append the SHA-256 hash which we previously computed for index.json to the filename. For example, fc.exe now becomes fc.exe_142ac6014584d99e0a520145f142e176fe200bec431d9cc96029ef058ac34f01. As this hash will change signifigantly when the file changes even slightly, so will the URL, forcing a cache miss each time there is a new file. If the file has not changed, the URL remains the same, allowing full use of the cache. It should be noted that old cached versions of files will remain in the cache with this technique. Hourai Deploy automatically appends file hashes, and Hourai Launcher will build the URL accordingly to incorporate the hash.

The end strategy is thus to invalidate the index.json each time it’s updated, in which it includes the data needed to build the appropriate cache-busting URL to fetch new files.