Real-time channels

pkg/realtime pushes messages to connected clients: a hub that owns topics, the two transports a browser understands, presence, and a relay for when the application runs in more than one process. Storage & background tasks introduces it in one screen; this page is the detail.

One thing to know before building on it: this package is not part of the API freeze. It is a transitional package, and no pkg/realtime symbol appears in the frozen baseline, so the no-removal guarantee that covers pkg/router or pkg/signals does not cover the names below. Everything described here is what ships today.

Hubs, topics and clients

A hub owns the connections this process holds and the topics they subscribed to. A topic is a string you choose — "orders", "room:42", "user:17:notifications". There is no registration step: a topic exists while at least one client is subscribed to it and disappears when the last one leaves.

import "github.com/jcsvwinston/nucleus/pkg/realtime"

hub := realtime.New(realtime.Config{Logger: logger})
defer func() { _ = hub.Close() }()

Config takes a Logger (slog.Default() when omitted), a ClientBuffer (64 when omitted — see A slow client) and an optional Relay. One hub per process is the usual arrangement: build it in your bootstrap, capture it in the closures that serve the routes, close it on shutdown. Close disconnects everybody and closes the relay if there is one.

A client is one connection. Subscribe(id, user, topics...) registers it: the id must be non-empty and unique among the connections the hub currently holds, and at least one topic is required (ErrNoTopics otherwise). The transports below call it for you and unsubscribe when the connection ends — you call it directly only when writing a transport of your own. The user is whatever the handler decided; the package never derives it. Topics(), Count(topic) and Unsubscribe(id) round out the surface, and all of them answer for this process alone.

Broadcasting

payload, err := json.Marshal(order)
if err != nil {
    return err
}
hub.Broadcast(ctx, realtime.Message{
    Topic: "orders",
    Event: "created",
    ID:    order.ID,
    Data:  payload,
})

Data is already encoded — the package does not marshal for you. Broadcast returns nothing: it delivers to this process's subscribers and then hands the message to the relay, and a relay failure is logged rather than returned.

Event and ID reach an SSE client as the event: and id: fields. The WebSocket transport writes Data alone, so a WebSocket client that needs to know which kind of message it received needs that inside the payload. This is the main place the two transports deliver different things; they also differ in what each one says when the hub drops a client (A slow client) and in how each refuses a duplicate ClientID. It is worth deciding once for the whole application.

Serving a channel over server-sent events

r.Get("/live", func(c *nucleus.Context) error {
    return realtime.ServeSSE(c.Writer, c.Request, realtime.SSEConfig{
        Hub:    hub,
        Topics: []string{"orders"},
        User:   currentUser(c),
        OnConnect: func(send func(realtime.Message)) {
            send(realtime.Message{Event: "snapshot", Data: currentOrders()})
        },
    })
})

ServeSSE sets Content-Type: text/event-stream, Cache-Control: no-cache, Connection: keep-alive and X-Accel-Buffering: no (nginx buffers proxied responses by default, which turns a live stream into one lump at the end), writes the 200 and flushes it, then holds the connection until the request context is cancelled or the hub disconnects the client. It refuses early in four cases: no hub is a 500, no topics a 400, a response writer that is not an http.Flusher a 500, and a ClientID already connected to this hub a 409. Leaving ClientID empty derives a fresh one.

OnConnect runs once the client is subscribed and before the first broadcast reaches it. That ordering is what makes it the place to send current state: a client that connects halfway through the day sees where things stand instead of waiting for the next change.

A payload with newlines gets one data: prefix per line, so a multi-line JSON document arrives whole. Every KeepAlive (25 seconds by default) the stream writes a : keep-alive comment, which clients ignore and which stops a proxy deciding the connection is dead. The timer runs on a fixed interval; a stream that has just delivered a message writes the comment anyway.

Message.ID is written as the id: field and a browser sends it back as Last-Event-ID when it reconnects. The framework does not read that header and the hub keeps no history, so resuming from an id is something you build on top — the usual answer is to send current state from OnConnect instead. StreamContext runs the same handler with the request's context replaced by the one you pass, so the stream ends when that context is done and no longer when the client disconnects. Derive it from r.Context() — with context.WithCancel or a deadline on top — when you want both.

Serving a channel over WebSocket

r.Get("/ws/rooms/{room}", func(c *nucleus.Context) error {
    room := c.Param("room")
    return realtime.ServeWS(c.Writer, c.Request, realtime.WSConfig{
        Hub:    hub,
        Topics: []string{"room:" + room},
        User:   currentUser(c),
        OnMessage: func(client *realtime.Client, data []byte) error {
            hub.Broadcast(c.Request.Context(), realtime.Message{
                Topic: "room:" + room,
                Data:  data,
            })
            return nil
        },
    })
})

ServeWS completes the handshake — including the Sec-WebSocket-Accept key computed from the client's — then subscribes, reassembles fragmented inbound messages, answers the peer's pings and writes every broadcast as one unmasked text frame. The handshake is refused with 400 when the request is not an upgrade or carries no Sec-WebSocket-Key, with 426 and a Sec-WebSocket-Version: 13 header when the client does not ask for version 13 — including when it sends no version header at all — with 403 when the origin check says no, and with 500 when the response writer cannot be hijacked.

One refusal arrives after the handshake rather than instead of it. ServeWS upgrades first and subscribes second, so a ClientID already connected to this hub — the case ServeSSE answers with a 409 — gets a completed 101 followed immediately by a close frame carrying 1000, the same code a normal close uses. There is no status and no reason for the client to read. Leave ClientID empty, or derive it per connection, not per user: two tabs of one person are two connections.

OnMessage receives what the client sends, and returning an error from it ends the connection. It receives the payload without the frame's opcode, so a handler cannot tell a binary message from a text one; text and binary frames both arrive as []byte. Leaving it nil ignores inbound messages, which is the right default for a one-way channel — the read loop still runs, because it is what answers the peer's pings. OnConnect works as it does for SSE, except that what it sends goes out as Data only.

Outbound, ServeWS writes every message — broadcasts and whatever OnConnect sends — as a text frame. Conn.WriteBinary exists but nothing in ServeWS reaches it, so broadcasting binary means calling realtime.Upgrade yourself and driving the *Conn it returns.

The connection is pinged every PingInterval (30 seconds by default), on a fixed timer that traffic does not reset. Note the limit: the server sends the ping but does not enforce a deadline on the pong, so a peer that vanished without closing the TCP connection is noticed when a write to it fails or when the request context is cancelled, not one interval later. Conn.SetReadDeadline is exported for applications that drive a connection themselves and want that bound.

Which transport

Reach for SSE unless something forces the other choice. It is plain HTTP, so it crosses proxies that mangle upgrades; browsers reconnect on their own; and it carries the event name and the id without you encoding them. Its limits are that it is one-way and text-only. Take WebSocket when the client has to send as well — a chat room, a collaborative cursor, an acknowledgement — or when it sends binary frames: they reach OnMessage, as []byte the handler cannot tell from text. It does not buy you binary in the other direction: ServeWS writes text. And you own the reconnection behaviour on the client side, because a browser WebSocket does not reconnect by itself the way an EventSource does.

Either way, authorisation happens in the handler, before the call, the way it does for any other route: a channel is a route, and giving it a second authorisation mechanism is how the two drift apart. What the handler decides, it passes in as User.

Presence

Presence(topic) reports who is connected, one entry per connection — {ID, User, Topics}, sorted by id:

for _, entry := range hub.Presence("room:42") {
    log.Printf("connection %s belongs to %s", entry.ID, entry.User)
}

Two browser tabs belonging to one person are two entries, which is what a "3 devices" badge needs; count distinct User values when the question is "who is here". The relay carries broadcasts between replicas and not presence, so on more than one process Presence answers for one replica, not for the cluster.

The origin check

This is the security boundary of the WebSocket transport, so it is worth being precise. UpgradeConfig.CheckOrigin decides whether a handshake is allowed. Left nil — the default — the rule is:

• No Origin header at all: allowed. Browsers always send one; native clients and tests do not.
• An Origin header: allowed only when it matches the request's Host. Everything after :// in the Origin is compared case-insensitively to Host. The port is part of both and therefore part of the comparison; the scheme is not compared.
• Anything else: 403, before the connection is hijacked.

The default is same-origin because a browser sends the session cookie with a WebSocket handshake and does not apply CORS to it. A handshake accepted from another origin is a cross-site request carrying the user's session, and whatever the socket then streams is readable by the page that opened it.

Setting CheckOrigin replaces that rule entirely: a function that always returns true accepts every origin on the internet. If your front end genuinely lives somewhere else, compare against the origins you wrote down:

allowed := map[string]bool{"https://app.example.com": true}

cfg := realtime.WSConfig{
    Hub:    hub,
    Topics: []string{"orders"},
    Upgrade: realtime.UpgradeConfig{
        MaxMessageBytes: 64 << 10,
        CheckOrigin: func(r *http.Request) bool {
            return allowed[r.Header.Get("Origin")]
        },
    },
}

ServeSSE performs no origin check of its own: an SSE stream is an ordinary HTTP response, covered by the authentication, the CORS policy and the authorisation the route already carries.

Message size and close codes

MaxMessageBytes bounds inbound messages; unset it is 1 MiB (DefaultMaxMessageBytes). It is applied to the frame header before any buffer is allocated, because that length is attacker-controlled: a client that announces a gigabyte is refused, not allocated for.

• A frame whose declared length is over the limit — including one that only announces it and sends nothing — is answered with close code 1009 (message too big). Fragments that cross the limit while being reassembled get the same code.
• A client data frame — text, binary or continuation — that is not masked is answered with close code 1002. RFC 6455 requires every frame from a client to be masked, and an unmasked one is either a broken client or something that is not a browser. Control frames are handled before that check, so an unmasked ping is still answered with a pong, an unmasked pong is still ignored, and an unmasked close still ends the connection normally.
• Conn.Close, the normal end of a connection, sends 1000.

In each refusal the close frame is written and the socket closed immediately after; ServeWS returns and the client is unsubscribed. Outbound messages are not bounded by that limit — the server writes what you broadcast, in one unmasked frame with no fragmentation. They are bounded in time instead: at most five seconds waiting for the connection's write lock, then a ten-second write deadline, so one peer that stopped reading cannot hold the connection for ever.

A slow client

Each client has a send buffer of ClientBuffer messages, 64 by default. Broadcast never blocks on a client whose buffer is full: it increments that client's dropped counter and disconnects it, logging a warning that names the connection and the topic. Blocking the broadcaster would make one stalled browser everybody's problem and an unbounded buffer would make it the process's, whereas disconnecting makes it the client's. An EventSource reconnects on its own and the new request resubscribes it; a browser WebSocket does not, and nothing in this package resubscribes it, so on that transport the reconnection is code you write.

What the client sees depends on the transport. An SSE stream writes one last event before it ends:

event: nucleus.disconnected
data: {"reason":"too slow","dropped":1}

A WebSocket connection is closed with the normal 1000 code and the count is not sent.

Read that count for what it is. Client.Dropped() counts the sends that failed, and the first failure is what disconnects the client — so after an eviction it is 1, whether the connection went on to miss ten messages or ten thousand. Treat it as a flag that this connection fell behind, not as how many messages it missed: the hub does not keep that number, and a badge built on it would always read one. Because the hub keeps no history, a client that reconnects receives what happens next and nothing from while it was away; sending current state from OnConnect is what closes that gap.

More than one process

A hub broadcasts to the clients this process holds, so with a second replica running half your users stop seeing updates — the failure that looks like a flake and is not. A Relay carries broadcasts between replicas, and RedisRelay is the implementation that ships:

relay, err := realtime.NewRedisRelay(realtime.RedisRelayConfig{
    URL:     cfg.RedisURL,
    Channel: "myapp:realtime",
})
if err != nil {
    return err
}

hub := realtime.New(realtime.Config{Logger: logger, Relay: relay})
if err := hub.StartRelay(ctx); err != nil {
    return err
}
defer func() { _ = hub.Close() }()

StartRelay starts this replica's subscriber in a goroutine and sleeps 10 milliseconds before returning. The pause is a fixed sleep, not a confirmation that the subscription landed: it gives the subscription a moment before the process starts work, and a broadcast another replica publishes inside that window can still be missed here. It says nothing about the other direction either — what this replica publishes goes straight to Redis from Broadcast, and whether the other replicas receive it depends on their own StartRelay. The subscription runs until ctx is done. If it fails earlier the error is logged; if the pub/sub channel simply closes, the subscriber stops without an error and without a line.

Neither call reports whether Redis is reachable, so neither one tells you the relay works. The example checks both errors because both signatures return one, but today NewRedisRelay fails only on a missing or unparseable URL and StartRelay returns nil on every path it has. A wrong host, a wrong password or a Redis that is down starts cleanly, and the first sign is the relay error Broadcast logs. Register health.NewRedisProbe among your health checks if you want that connection reported rather than discovered.

Channel defaults to nucleus:realtime, so two applications sharing one Redis database must set distinct channels — or better, use distinct databases.

What it does about its own messages. Broadcast delivers locally first and publishes second. Every published envelope carries the replica's Origin — a generated identifier when you do not set one — and the subscriber skips envelopes whose origin is its own, because delivering the relayed copy as well would show every client every message twice. Messages arriving from another replica are delivered to local subscribers and not published back out, so they cannot loop.

The skip is a string comparison, which makes Origin a value that must differ per replica. Set it to something stable and shared — an application name, a deployment name, anything every pod reads from the same config — and each replica discards every other replica's broadcasts: the relay goes quiet, with no error anywhere, and you are back to half your users not seeing updates. Leave it empty unless you have a per-instance identifier to hand.

Redis pub/sub rather than a stream, deliberately: nothing is persisted, so a message published while a subscriber is briefly disconnected is gone. If losing it is not acceptable then it is not a broadcast — use the transactional outbox or a background task, and broadcast from the consumer.

Timeouts and routing

A channel is a route and gets the default middleware stack, where one default matters: request_timeout (30 seconds) wraps handlers in http.TimeoutHandler, which buffers the whole response and hides http.Flusher — behind it a stream cannot flush a byte until it returns.

The router already exempts the two cases this page is about. A request whose Accept header asks for text/event-stream is exempt, which is what a browser's EventSource always sends, and so is a WebSocket upgrade; both work with no configuration. A stream served to a client that does not send that header — curl, a native client — needs its path prefix listed instead. Entries are matched with strings.HasPrefix, not compared as whole paths, so an entry of /api also exempts /api-internal and everything under it: list the narrowest prefix that covers the stream, or the 30-second timeout quietly comes off a whole subtree. An exempt prefix gets the raw response writer, with Flush and Hijack working, and no deadline:

request_timeout: 30s
timeout_exempt_paths:
  - /live

In tests, the kit opens a stream for you: srv.Stream("/live") returns an open connection, Next(limit) reads the next event and skips keep-alive comments, and Quiet(limit) asserts that nothing arrived — the assertion an authorisation bug slips past when nobody writes it.