This is a very hacky project, so it might stop working if Discord changes their API core. This is unlikely, but keep that in mind while using the proxy.
This is a proxy for Discord gateway connections - clients can connect to this proxy instead of the Discord Gateway and interact with it just like they would with the Discord Gateway.
The proxy connects to Discord instead of the client - allowing for zero-downtime client restarts while the proxy keeps its connections to the gateway open. The proxy won't invalidate your sessions or disconnect you (exceptions below).
It connects all shards to Discord upfront and mimics to be the actual API gateway.
When a client sends an IDENTIFY
payload, it takes the shard ID specified and relays all events for that shard to the client.
It also sends you self-crafted, but valid READY
and GUILD_CREATE
/GUILD_DELETE
payloads at startup to keep your guild state up to date, just like Discord does, even though it doesn't reconnect when you do internally.
Because the IDENTIFY
is not actually controlled by the client side, activity data must be specified in the config file and will have no effect when sent in the client's IDENTIFY
payload.
It uses a minimal algorithm to replace the sequence numbers in incoming payloads with fake sequence numbers that are valid for the clients, but does not need to parse the JSON for that.
Create a file config.json
and fill in these fields as you wish:
{
"log_level": "info",
"token": "",
"intents": 32511,
"port": 7878,
"activity": {
"type": 0,
"name": "on shard {{shard}} with kubernetes"
},
"status": "idle",
"backpressure": 100,
"externally_accessible_url": "ws://localhost:7878",
"cache": {
"channels": false,
"presences": false,
"emojis": false,
"current_member": false,
"members": false,
"roles": false,
"stage_instances": false,
"stickers": false,
"users": false,
"voice_states": false
}
}
You can omit the token
key entirely and set the TOKEN
environment variable when running to avoid putting credentials in the configuration file.
By default, the total shard count will be calculated using the /api/gateway/bot
endpoint. If you want to change this, set shards
to the amount of shards. It will also launch all shards by default, you can customize this to launch only a range of shards using shard_start
and shard_end
(start inclusive, end exclusive).
If you're using twilight's HTTP-proxy, set twilight_http_proxy
to the ip:port
of the HTTP proxy.
Take special care when setting cache flags, only enable what you actually need. The proxy will tend to send more than Discord would, so double check what your bot depends on.
Compiling this from source isn't the most fun, you'll need a nightly Rust compiler with the rust-src component installed. Then run cargo build --release --target=MY_RUSTC_TARGET
, where MY_RUSTC_TARGET
is probably x86_64-unknown-linux-gnu
.
Instead, I recommend running the Docker images that are prebuilt by CI.
The Docker images are tagged based on the CPU microarchitecture that they are built and tuned for, currently either znver3
(Zen 3), znver2
(Zen 2), haswell
, sandybridge
or x86-64
(the only target with SIMD disabled, therefore the most compatible).
To run the image, mount the config file at /config.json
, for example:
docker run --rm -it -v /path/to/my/config.json:/config.json docker.io/gelbpunkt/gateway-proxy:haswell
Connecting is fairly simple, just hardcode the gateway URL in your client to ws://localhost:7878
. Make sure not to ratelimit your connections on your end.
If you have not configured a shard count manually, you can check the amount of shards you need to create on your client by requesting http://localhost:7878/shard-count
. The endpoint returns the number of shards running as plaintext.
Important: The proxy detects zlib-stream
query parameters and compress
fields in your IDENTIFY
payloads and will encode packets if they are enabled, just like Discord. This comes with CPU overhead and is likely not desired in localhost networking. Make sure to disable this if so.
The proxy exposes Prometheus metrics at the /metrics
endpoint. They contain event counters, cache size and shard latency histograms specific to each shard.
Voice support, while being present for a while, has been removed entirely. This is because the proxy would have to track voice sessions as sent by Discord, while also accounting for other caveats. I currently don't use this feature and would much prefer Discord to add a voice session API to their HTTP endpoints. The old implementation of this was ugly and very quickly hacked together; I would definitely appreciate a PR to implement this in a pretty and well-documented way, but won't do it myself for now.
In theory, the proxy is very fast for the reasons mentioned above. In practice, this shows. There is almost zero overhead in latency.
Using 225 shards, with almost full caching (members, guilds, channels, roles, voice states) the proxy uses 11.7GB of memory and sits around 2% CPU usage over all 4c/8t of my machine. This again shows that the processing overhead is negligible, the only thing you can and should optimize on is the cache configuration.
- Re-add voice support