/ɕʰinɡjʌ̹ŋ/ • sin-gyeong (try it with IPA Reader)
- nerve
/nərv/ • noun
- (in the body) a whitish fiber or bundle of fibers that transmits impulses of sensation to the brain or spinal cord, and impulses from these to the muscles and organs.
"the optic nerve"
신경 is the nerve-center of your microservices, providing a message bus + message queue with powerful routing, automatic load-balancing + failover, powerful HTTP request proxying + routing, and more. 신경 aims to be simple to get started with while still providing the features for a variety of use-cases.
신경 is a part of the amyware Discord server.
If you like what I make, consider supporting me on Patreon:
Language | Author | Link | Maintained? |
---|---|---|---|
Elixir | @queer | https://github.com/queer/singyeong-client-elixir | yes |
.NET | @FiniteReality | https://github.com/finitereality/singyeong.net | yes |
Python | @PanKlipcio | https://github.com/StartITBot/singyeong.py | yes |
Typescript | @cyyynthia | https://github.com/borkenware/singyeongjs | looks like yes (used in squirrelchat?) |
Language | Author | Link | Notes |
---|---|---|---|
Python | @PendragonLore | https://github.com/PendragonLore/shinkei | Owner said it's unmaintained |
Java | @queer | https://github.com/queer/singyeong-java-client | I don't write much Java anymore S: |
Typescript | @alula | https://github.com/KyokoBot/node-singyeong-client | Repo gone |
신경 was inspired by sekitsui, a project by the Ayana developers. 신경 was developed due to sekitsui seemingly having halted development (no release as far as I'm aware, no repo activity in the last 1-2 years).
신경 is ALPHA-QUALITY software. The core functionality works, but there's no guarantee that it won't break, eat your cat, ... Use at your own risk!
Configuration is done via environment variables, or via a custom configuration
file. See config.exs
for more information about config options.
Sometimes, it's necessary to include custom configuration files - such as for
something that the environment variables don't cover. In such a case, you can
add a custom.exs
file to config/
that includes the custom configuration you
want / need.
Example: Prove that custom config works:
use Mix.Config
IO.puts "Loading some cool custom config :blobcatcooljazz:"
Example: Always have debug-level logging, even in prod mode:
use Mix.Config
config :logger, level: :debug
Plugins belong in a directory named plugins
at the root directory. See the
plugin API and the
example plugin for more info.
신경 is capable of discovering cluster members automatically, using libcluster and the gossip strategy by default.
신경 is a metadata-oriented message bus + message queue + HTTP proxy. Clients connect over a websocket, and can send messages, queue messages, and send HTTP requests that can be routed to clients based on client metadata.
신경 clients are identified by three factors:
- Application id.
- Client id.
- Client metadata.
When sending messages or HTTP requests over 신경, you do not choose a target service instance directly, nor does 신경 choose for you. Rather, you specify a target application and a metadata query. 신경 will then run this query on all clients under the given application, and choose one that matches to receive the message or request.
For example, suppose you wanted to let users who had opted-in to a beta program
use beta features, but not all users. You could express this as a 신경 query,
and say something like "send this message to some service in the backend
application where version_number >= 2.0.0
."
Of course, something like that is easy, but 신경 lets you do all sorts of
things easily. For example, suppose you had a cluster of websocket gateways
that users connected to and received events over. Instead of having to know
which gateway a user is connected to, you could trivially express this as a
신경 query - "send this message to a gateway
node that has 123
in its
connected_users
metadata." Importantly, sending messages like this is done
in exactly the same way as sending any other message. 신경 tries to make it
very easy to express potentially-complicated routing with the same syntax as a
simple "send to any one service in this application group."
No. You should not try to route to a specific 신경 client by id; instead you should be expressing a metadata query that will send to the client you want. Generally speaking, clients should be capable of running statelessly, or you should use metadata to route messages effectively.
Nope.
신경 has masterless clustering support.
- No need for Kubernetes or something similar - anything that can speak websockets is a valid 신경 client.
- No configuration. 신경 is meant to be "drop in and get started" - a few options exist for things like authentication, but beyond that, no configuration should be needed (at least to start out).
- Fully dynamic. 신경 is meant to work well with clients randomly appearing and disappearing (ex. browser clients when using 신경 as a websocket gateway).
- No sidecars.
- Choose where messages / requests are routed at runtime; no need to bake exact targets into your application.
- Service discovery without DNS.
- Service discovery integrated into HTTP proxying / message sending.
- Query performance might be unacceptable.
- Websockets might not be acceptable.
- Development is still fairly early-stage; the alpha-ish quality of it may be nonviable.
Metadata-based routing is useful for all sorts of things:
- Storing documentation about what messages your services send/recv, what REST endpoints they expose, ... and querying on it to route to a service that accepts a specific format of a specific message.
- Sending a message to a subset of clients without doing a full pubsub and relying on clients to drop them properly.
- Discord bot message-passing between services. No more
pubsub or whatever, just "send this message to the service where
123 in guild_ids
." - Websocket gateway. "Send this message to the client where
user_id = 123
." - Container scheduling.
- Monitoring host stats.
- Routing messages to an audit-logging service and a handler service at the same time.
- Message queues that can only dispatch messages when a client is capable (ex.
"dispatch this message from the queue to a client where
latency < 10
") - Anything you can think of!
In general, 신경 can get messages to the right place with some very complicated conditions very easily.
I like Elixir 👍 Elixir is well-suited to the use-case of message-queuing, and imo is a lot friendlier to write scalable messaging code in with little effort relative to any of the above-mentioned languages.
Phoenix's socket abstraction is really really useful. Also I didn't want to have to build eg. HTTP routing from scratch; Phoenix does a great job of it already so no need to reinvent the wheel. While it is possible to use Plug or a similar library on top of Cowboy or another HTTP server, I just liked the convenience of getting it all out-of-the-box with Phoenix and being able to focus on writing my application-level code instead of setting up a ton of weird plumbing.
Check out PROTOCOL.md.
The code is intended to be pretty easy to read. The general direction that data flows is something like:
client -> decoder -> gateway -> dispatch -> process and dispatch response
-> identifier -> allow or reject connection
-> metadata updater -> apply or reject metadata updates
In terms of module structure, the way things go is something like:
|-> <phx/cowboy code>
| |
| V
| SingyeongWeb.Transport.Raw
| |
| V
| Singyeong.Gateway
| |
| |------------------------------------|-----------------------------------------|
| V V V
| Singyeong.Gateway.Handler.Identify Singyeong.Gateway.Handler.DispatchEvent Singyeong.Gateway.Handler.Heartbeat
| | | |
| V V V
| Singyeong.Gateway Singyeong.Gateway.Dispatch Singyeong.Store ---------|
| | | |
| V |--------------|-----------------|-----------------------------| |
|-- SingyeongWeb.Transport.Raw | V | | |
^ V (METADATA_UPDATE) V V |
| Singyeong.Gateway Singyeong.Store Singyeong.MessageDispatcher Singyeong.Queue |
|----------------| | | |
| | | |
|------------------------------------------------------|-----------------------------|----------|
Additionally, I aim to keep the server fairly small, ideally <5k LoC, but absolutely <10kLoC no matter what.. At the time of writing, the server is ~3.600 LoC:
git:(master) X | -> tokei lib/
===============================================================================
Language Files Lines Code Comments Blanks
===============================================================================
Elixir 48 4558 3621 267 670
===============================================================================
Total 48 4558 3621 267 670
===============================================================================
git:(master) X | ->
When you send a message, the server inspects its target (ie. its metadata query) and queries its own internal metadata store to find clients that can be routed to that match said query. In the case of a multi-node setup, each node is only aware of its own clients, and thus metadata queries are a parallel RPC across the cluster.
A very basic query looks like this:
{
"application": "api",
"ops": [],
}
That is, "route this message to any client that's a member of the api
application." More-complex queries may do things like specifying constraints
based on latency:
{
"application": "api",
"ops": [
{
"path": "/latency/http",
"op": "$lte",
"to": {"value": 100},
}
],
}
That is, "route this message to any client that's a member of the api
application, where the value at that application's /latency/http
key is less
than or equal to 100
."
As metadata is arbitrary JSON -- the only restriction being that the top-level JSON is an object -- you can query effectively-infinitely-nested structures. You can query paths with a syntax inspired by JSON Patch.
For more about how the query language works, see the relevant docs
Messages are delivered at-least-zero times. That is, speaking broadly, your messages will be delivered exactly once, but there are cases (network issues etc.) that can lead to either no delivery, or more than one delivery, of the same message.
You can run tests with mix test
. Note that the plugin tests WILL FAIL
unless you set up the test plugin
in priv/test/plugin/
. If you don't want to deal with the plugin-related code
when running tests (tho you REALLY should care...), you can skip those tests by
running mix test --exclude plugin
.
Note that the HTTP proxying tests use an echo server I wrote (echo.amy.gg
),
rather than using a locally-hosted one. If you don't want to run these tests,
set the DISABLE_PROXY_TESTS
env var.
Note that there is no ratelimit on authentication attempts. This means that a malicious client can constantly open connections and attempt passwords until it gets the correct one. It is highly recommended that you use a very long, probably-very-complicated password in order to help protect against this sort of attack.
신경 means nerve, and since the nervous system is how the entire body communicates, it seemed like a fitting name for a messaging system. I considered naming this something like 등뼈 (deungppyeo, "spine"/"backbone") or 회로망 (hoelomang, "network") or even 별자리 (byeoljali, "constellation), but I figured that 신경 would be easier for people who don't know Korean to pronounce, as well as being easier to find from GitHub search.