A video introducing this project is available here.

The algorithm:

To retrieve the most YouTube video ids in order to retrieve the most video captions, we need to retrieve the most YouTube channels. So to discover the YouTube channels graph with a breadth-first search, we proceed as follows:

1. Provide a starting set of channels.
2. Given a channel, retrieve other channels thanks to its content by using YouTube Data API v3 and YouTube operational API and then repeat 1. for each retrieved channel.

A ready to be used by the end-user website instance of this project is hosted at: https://crawler.yt.lemnoslife.com

See more details on the Wiki.

The project structure:

• main.cpp contains the C++ multi-threaded algorithm proceeding to the YouTube channels discovery. It is notably made of the following functions:
  • main which takes into account the command line arguments, load variables from files (channels.txt, keys.txt, channels/ content) and start the threads as executing treatChannels function
  • treatChannels gets a YouTube channel to treat, treat it in treatChannelOrVideo function and compress the retrieved data
  • treatChannelOrVideo which provided a YouTube channel id or a video id, treats this resource. In both cases it treats comments left on this resource. In the case of a channel it also treats its CHANNELS, COMMUNITY, PLAYLISTS and LIVE tabs and downloads the captions of the channel videos.
  • markChannelAsRequiringTreatmentIfNeeded which provided a YouTube channel id marks it as requiring treatment if it wasn't already treated
  • execute which provided an yt-dlp command executes it in a shell
  • getJson which provided an API request returns a JSON structure with its result. In the case that the API requested is YouTube Data API v3 and a set of keys is provided (see below keys.txt), it rotates the keys as required
• channels.txt contains a starting set of channels which contains mostly the 100 most subscribed French channels
• keys.txt contains a set of YouTube Data API v3 keys (not provided) to have the ability to request this API (see an alternative to filling it in the section below with --no-keys command line argument)
• scripts/ contains Python scripts to:
  • generate the channels.txt as described above (retrieveTop100SubscribersFrance.py)
  • remove channels being treated before a restart of the algorithm as described in the main function documentation (removeChannelsBeingTreated.py)
• website/ is a PHP website using WebSocket to allow the end-user to proceed to requests on the retrieved dataset. When fetching the website, the end-user receives the interpreted index.php which upon making a request interacts with websocket.php which in the back-end dispatches the requests from various end-users to search.py (which treats the actual end-user request on the compressed dataset) by using users/ to make the inter-process communication.

Note that this project heavily relies on YouTube operational API which was modified for this project.

Running the YouTube graph discovery algorithm:

Because of the current compression mechanism, Linux is the only known OS able to run this algorithm.

To clone the repository, run:

git clone https://gitea.lemnoslife.com/Benjamin_Loison/YouTube_captions_search_engine

Move to the cloned repository by running:

cd YouTube_captions_search_engine/

To install the dependencies on an apt based Linux distribution of this project make sure to have pip and run:

sudo apt install nlohmann-json3-dev
pip install yt-dlp

To compile the YouTube discovery graph algorithm, run:

make

To see the command line arguments of the algorithm, run:

./youtubeCaptionsSearchEngine -h

To run the YouTube discovery graph algorithm, run:

./youtubeCaptionsSearchEngine

Except if you provide the argument --youtube-operational-api-instance-url https://yt.lemnoslife.com, you have to host your own instance of the YouTube operational API.

Except if you provide the argument --no-keys, you have to provide at least one YouTube Data API v3 key in keys.txt.

Hosting the website enabling users to make requests:

Move to the website/ folder by running:

cd website/

To install its dependencies make sure to have composer installed and run:

sudo apt install nginx
composer install
pip install webvtt-py

Add the following configuration to your Nginx website one:

    # Make the default webpage of your website to be `index.php`.
    index index.php;

    # Allow end-users to retrieve the content of a file within a channel zip.
    location /channels {
        rewrite ^(.*).zip$ /channels.php;
        rewrite ^(.*).zip/(.*).json$ /channels.php;
        rewrite ^(.*).zip/(.*).txt$ /channels.php;
        rewrite ^(.*).zip/(.*).vtt$ /channels.php;
        # Allow end-users to list `channels/` content.
        autoindex on;
    }

    # Disable end-users to access to other end-users requests.
    location /users {
        deny all;
    }

    # Configure the websocket endpoint.
    location /websocket {
        # switch off logging
        access_log off;

        # redirect all HTTP traffic to localhost
        proxy_pass http://localhost:4430;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header Host $host;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;

        # WebSocket support (nginx 1.4)
        proxy_http_version 1.1;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection "upgrade";

        # timeout extension, possibly keep this short if using a ping strategy
        proxy_read_timeout 99999s;
    }

Start the websocket worker by running:

php websockets.php