pg-sse

Live updates from the Postgres you already run. No Redis, no broker, no WebSocket service: a table trigger sends NOTIFY, one listener per process wakes the open streams, and each stream re-reads from its cursor and sends server-sent events. Works with Starlette and FastAPI out of the box, and with any other framework through hub.stream.

pip install "pg-sse[starlette]"   # or "pg-sse[fastapi]"; plain "pg-sse" is enough for hub.stream
pip install "psycopg[binary]"     # on a machine without libpq; see the psycopg install docs for the alternatives

The one rule

A notification carries only a key (a room id, a conversation id) and means "something changed here, re-read from your cursor". It never carries the data. So a lost notification costs a delay, never an event, and there's no 8000-byte payload limit to hit.

Why not X

• Polling. Every client hits the database on a timer, so load grows with clients times frequency, and latency is half the interval. Here an idle stream costs one read a minute, and a write reaches clients within milliseconds.
• WebSockets. You usually end up running a second service and writing your own reconnect and catch-up logic. EventSource reconnects by itself and sends Last-Event-ID, and one-way updates are all most pages need.
• Redis pub/sub. It is another system to run, secure and keep consistent with your database, and a message missed during a disconnect is gone. Here the database is the source of truth and a missed wake-up costs a delay.
• Supabase Realtime. It ties you to one vendor's platform and its protocol. This runs against any Postgres 14+ you can open a direct connection to.

Use it

1. Once, in a migration: wire the table to a channel.

from pg_sse import trigger_sql

conn.execute(trigger_sql("messages", key_column="room", channel="room_change", order_column="id"))

That creates two triggers:

2. The app:

from typing import Annotated
from fastapi import FastAPI, Header
from pg_sse import Event, Hub

hub = Hub(DIRECT_DATABASE_URL, channel="room_change")  # a direct connection, not a pooled one
app = FastAPI(lifespan=hub.lifespan)


def read_after(room: str, after: int) -> list[Event]:
    rows = db.fetch("SELECT id, body FROM messages WHERE room = %s AND id > %s ORDER BY id LIMIT 500", (room, after))
    return [Event({"id": id, "body": body}, id=id) for id, body in rows]


def latest_id(room: str) -> int:
    return db.fetchval("SELECT coalesce(max(id), 0) FROM messages WHERE room = %s", (room,))


@app.get("/rooms/{room}/stream")
async def stream(room: str, last_event_id: Annotated[str | None, Header()] = None):
    cursor = int(last_event_id) if last_event_id and last_event_id.isdigit() else None  # client input: parse it
    return hub.sse(room, read=lambda after: read_after(room, after), cursor=cursor, start=lambda: latest_id(room))

3. The browser:

const source = new EventSource("/rooms/lobby/stream");
source.onmessage = (e) => render(JSON.parse(e.data));

EventSource reconnects by itself and sends Last-Event-ID; the stream catches up from there.

A full runnable app, with async reads on a connection pool and a browser page: examples/chat.py.

Other frameworks

hub.sse is a thin wrapper over hub.stream, which yields encoded SSE frames (str) and knows nothing about Starlette. It raises TooManySubscribers at the call, before any response exists, so turn that into your framework's 503. The Django and Litestar snippets below are illustrative, not tested in CI.

# Django (ASGI; run the hub's lifespan from your ASGI app or a lifespan wrapper)
from django.http import HttpResponse, StreamingHttpResponse


async def stream(request, room):
    try:
        frames = hub.stream(room, read=lambda after: read_after(room, after), start=lambda: latest_id(room))
    except TooManySubscribers:
        return HttpResponse(status=503, headers={"Retry-After": "30"})
    return StreamingHttpResponse(frames, content_type="text/event-stream", headers={"Cache-Control": "no-cache"})

# Litestar
from litestar import get
from litestar.response import Stream


@get("/rooms/{room:str}/stream")
async def stream(room: str) -> Stream:
    frames = hub.stream(room, read=lambda after: read_after(room, after), start=lambda: latest_id(room))
    return Stream(
        frames, media_type="text/event-stream", headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"}
    )

Aiohttp and Quart work the same way: write each frame to the response as it arrives.

Signal streams

A stream doesn't have to carry data. It can tell the page what to refetch: "this list changed", "only the typing indicator changed", "you may have missed something, refetch everything". For that, give read a second parameter named wake (or a required second parameter). It receives a Wake that says why the read is running; several causes can arrive together, since a burst is one read.

wake.data_may_have_changed is True unless the only causes are hub.wake signals.

from pg_sse import Event, Wake


def signals(cursor, wake: Wake) -> list[Event]:
    if wake.start:
        return [Event({}, event="ready")]  # the page loads everything once
    if wake.resync or wake.reread:
        return [Event({}, event="resync")]  # anything may have changed
    if wake.keys:
        return [Event({"keys": sorted(wake.keys)}, event="change")]
    if "typing" in wake.signals:
        return [Event({}, event="typing")]  # refetch only the typing indicator
    return []


@app.get("/inbox/stream")
async def inbox(user: User = Depends(current_user)):
    return hub.sse(None, read=signals, owner=user.id)  # every key on the channel


# elsewhere, from any thread:
hub.wake(room_id, signal="typing", all_keys=False)

A one-parameter read(cursor) works too, and so does one whose second parameter has a default under another name (read(after, limit=500)): only a parameter named wake, or a required one, receives the Wake. hub.wake(key) reaches the streams for key and the all-keys streams (key=None); all_keys=False keeps it on its key, which is right for ephemeral per-key state (typing, presence, a cursor position) that an all-keys view has nothing to show for. Database notifications always reach the all-keys streams: they report real writes.

What hub.sse does for you

• Subscribes before the catch-up read, so a write between the two is a queued wake-up, not a gap.
• Calls start() after subscribing when there's no cursor, and sends that cursor as the first frame's id, so a reconnect resumes from it even if no event arrived.
• Re-reads on every wake-up, and a burst of notifications costs one read.
• Pings every 15–25 s (ping_seconds) with an SSE comment, so proxies do not time an idle stream out. A ping costs no database read.
• Re-reads after 60 s of quiet (reread_seconds) even when nothing woke it. This is the safety net under every lost notification: the worst case is a one-minute delay.
• Wakes every stream after the listener connects or reconnects, since anything committed before its LISTEN existed (at startup, or while it was down) was missed. Reconnects with backoff; a SELECT 1 every minute and TCP keepalives catch a half-open connection.
• Bounded queue per stream (64). Identical wake-ups are queued once, so a hot key costs one slot, not one per write. A stream ends, rather than buffers, when more than queue_size distinct keys or signals arrive during one read (only an all-keys stream on a wide channel gets there); the browser reconnects and catches up.
• Caps open streams per process (500) and, when you pass owner, per key and owner (5). Past a cap, hub.sse raises a 503 with Retry-After: 30. Without owner, one client can hold every slot in a worker; pass owner (user id, session, or client address) on public endpoints.
• Ends streams after 15 minutes. A stream is authenticated once, when it opens; this bounds how long a revoked session keeps one.
• Ends every stream on SIGTERM, so a deploy doesn't wait out uvicorn's graceful-shutdown timeout. It chains onto the server's handler; with no handler installed (a plain script) the default action is left alone and ends the process as usual.
• Says goodbye before a planned close. The last frame carries retry:, which EventSource applies by itself, and, with end_event set, a named event with the reason:

There is no goodbye after Event(final=True), since the application already said what it meant, or when read raised, since the stream's state is unknown and an abrupt close is the honest signal.

// EventSource: nothing to write. `retry` is applied automatically; the reason is there if you want it.
source.addEventListener("end", (e) => console.debug("stream ended:", JSON.parse(e.data).reason));

// A fetch-based reader decides for itself.
if (event.event === "end") {
  const { reason } = JSON.parse(event.data);
  scheduleReconnect(reason === "busy" ? 30_000 : event.retry ?? 1_000);
}

API

Health check

import dataclasses

from fastapi.responses import JSONResponse


@app.get("/healthz")
async def healthz():
    stats = hub.stats()
    return JSONResponse(dataclasses.asdict(stats), status_code=200 if stats.listener_connected else 503)

Testing your app

The hub needs a real Postgres; there is nothing to mock, because the database is the source of truth. Point the tests at a scratch database and enter the hub's lifespan. Starlette's TestClient runs the lifespan off the main thread, where the signal hook is skipped anyway, but pass handle_signals=False in tests so a test never touches the process's signal handlers.

hub = Hub(TEST_DSN, channel="room_change", handle_signals=False)
with TestClient(app) as client:  # runs the lifespan
    ...

Multiple workers

Each process (each uvicorn or gunicorn worker) holds one LISTEN connection and its own caps and counters. Nothing is shared between workers and nothing needs to be: Postgres delivers every notification to every listener, and each worker wakes its own streams. Budget one connection per worker.

Limits

• LISTEN needs a direct connection. Behind a transaction-mode pooler (PgBouncer in transaction mode, the pooled connection strings Supabase and Neon give out) LISTEN succeeds and nothing ever arrives; streams then update only on the re-read. Give the hub the direct URL; your queries can still use the pooler.
• The cursor sees new rows, not updates. If a stream must show edits, have read return the current state alongside new rows, or use a cursor that moves on update (a version column with order_column).
• Sync reads run in a thread pool. A plain-function read or start runs in asyncio's default executor (min(32, cpu + 4) threads), and under Starlette other sync routes share anyio's pool of 40. Hundreds of streams re-reading at once queue behind it. Prefer async read and start on an psycopg_pool.AsyncConnectionPool; see examples/chat.py. Use functools.partial(read_after, room) rather than a lambda to keep an async function off the pool.
• order_column serialises inserts per key. The ordering trigger takes pg_advisory_xact_lock on the key and holds it until the transaction ends, so every other insert to that key waits for the whole transaction. Two transactions inserting into keys in opposite order can deadlock, and Postgres aborts one of them. Keep the transaction that inserts short, or skip order_column and tolerate a late id arriving on the next re-read. The lock is pg_advisory_xact_lock(hashtext(channel), hashtext(key)); advisory locks share one namespace per database, so an application taking the same two-int lock would serialise against these inserts, and since hashtext is 32-bit two different keys can collide and briefly wait on each other (a delay, never wrong ids).
• Per process. Each worker runs its own listener, which works, but the caps are per process.
• The database may be down at startup. lifespan waits up to listen_timeout for the first LISTEN and then starts anyway; until the listener connects, streams update on their re-read and listener_connected is false.
• Disconnect detection depends on the server. Starlette 1.7 under an ASGI server that advertises spec_version 2.4 no longer listens for http.disconnect during a streaming response: a dead client is noticed at the next write, at most ping_seconds[1] later, and its slot is held until then. uvicorn 0.54 advertises 2.3 and frees the slot at once.
• psycopg 3 for the listener; your own queries can use any driver.
• A framework for the response. hub.sse needs Starlette or FastAPI; any other framework uses hub.stream.

For AI coding agents

If you are generating code with this library, follow these rules:

1. Run trigger_sql(table, key_column, channel, order_column="id") once in a migration. Do not call pg_notify from application code, and do not put row data in a notification.
2. Give Hub a direct database URL (not a pooled / PgBouncer / Supabase port 6543 / Neon -pooler URL).
3. Register the hub with FastAPI(lifespan=hub.lifespan), or async with hub.lifespan(): inside an existing lifespan.
4. Call hub.sse(...) (Starlette/FastAPI) or hub.stream(...) (any framework; catch TooManySubscribers) from an async def route, after any 404 or permission check.
5. read(after) returns pg_sse.Event items after the cursor, oldest first, each with id set. Use WHERE id > %s ORDER BY id LIMIT n.
6. Parse Last-Event-ID before passing it as cursor; it is client input.
7. Pass start= (the current max id) when the stream should start from now. Do not compute it before calling hub.sse: that reintroduces a gap.
8. Prefer async read and start on a connection pool; plain functions run in a limited thread pool.
9. In the browser use EventSource; it resumes with Last-Event-ID by itself.
10. read(cursor, wake) is optional: use it only when the stream must tell the client what kind of change happened (see Signal streams). A stream that sends rows keeps read(cursor).
11. Use hub.wake(key, signal=..., all_keys=False) for ephemeral per-key state (typing, presence). The default also wakes the all-keys streams.
12. Set Hub(end_event="end") when the client is not EventSource, so it can read the reason and the retry before a planned close.

Development

See CONTRIBUTING.md. In short:

uv sync
PG_SSE_TEST_DSN=postgresql://sse:sse@127.0.0.1:55432/pg_sse_test uv run pytest --cov
uv run ruff check . && uv run mypy

The tests create and drop a schema per test in the database you point them at; use a scratch database.

License

MIT