So, how they solved it? 𝗧𝗵𝗲𝘆 𝗳𝗶𝗿𝘀𝘁 𝘁𝗿𝗶𝗲𝗱 𝘁𝗼 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱 𝗵𝗼𝘄 𝘁𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀. They tracked what Elixir processes were doing, if they were stuck waiting on something, etc. They recorded the event types, how many of each kind of message they received, and their processing times. In addition, they tried to understand how much memory they use, the performances of garbage collectors, etc.

After the analysis, they 𝗰𝗿𝗲𝗮𝘁𝗲𝗱 𝘁𝗵𝗲 𝗳𝗼𝗹𝗹𝗼𝘄𝗶𝗻𝗴 𝘀𝘁𝗿𝗮𝘁𝗲𝗴𝘆:

𝟭. 𝗣𝗮𝘀𝘀𝗶𝘃𝗲 𝘀𝗲𝘀𝘀𝗶𝗼𝗻𝘀: Discord significantly reduced the amount of data processed and sent by differentiating between active and passive user connections, cutting the fanout work by 90% for large servers.

𝟮. 𝗥𝗲𝗹𝗮𝘆𝘀: Implementing a relay system (read - multithreading) allowed Discord to split the fanout process across multiple machines, enabling a single guild to utilize more resources and support more prominent communities. Relays maintain connections to the sessions instead of the guild and are responsible for doing fanout with permission checks.

𝟯. 𝗪𝗼𝗿𝗸𝗲𝗿 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗲𝘀 𝗮𝗻𝗱 𝗘𝗧𝗦: To maintain server responsiveness, Discord employed worker processes and Erlang Term Storage (ETS) for operations requiring iteration over large sets of members, thus avoiding bottlenecks in the guild process. ETS is an in-memory database that supports the ability of multiple Elixir processes to access it safely. This enables the creation of a new worker process and passes the ETS table so this process can run expensive operations and offload the central guild server.

🔗https://discord.com/blog/maxjourney-pushing-discords-limits-with-a-million-plus-online-users-in-a-single-server