This is a Proof of Concept of using the Raft Consensus Algorithm to create a highly available Discord bot that has its shards distributed across multiple servers, without a central manager server telling which node to run which shards.
In this PoC, one of the nodes is elected as the leader and, instead of running the regular bot, it becomes the manager for the shard cluster. The leader is responsible for telling each node which shards to run. If the leader suddenly becomes unavailable, another node will be elected as the new leader and will take over the manager role.
The amount of shards to run is stored in recommendedShards.txt, as a mock of Discord's recommended shards
endpoint, for ease of editing. The leader
will read this file and distribute the shards across the nodes.
Shards are semi-pinned to nodes, meaning that a shard will only be moved to another node if the node it is currently running on becomes unavailable, or if that node is over the shard limit. This is to avoid unnecessary shard migrations and downtime.
-
If the total amount of shards changes, all nodes will need to be restarted for the new shard distribution to take effect. This is due to the way that Discord's gateway works.
-
When a worker's shards change, all of the worker's shards will be disconnected and reconnected. This is also due the way that Discord's gateway works.
This PoC requires Bun and Redis to run. Once redis is running, use bun to start as many nodes as you want.
bun src/index.tsThe token for the bot is read from the DISCORD_TOKEN environment variable.