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, IntCursor

hub = Hub(DIRECT_DATABASE_URL, channel="room_change")  # a direct connection, not a pooled one
app = FastAPI(lifespan=hub.lifespan)
cursors = IntCursor()  # a Last-Event-ID is client input: only digits get through


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 = cursors.parse_one(last_event_id)  # None for no header or one that is not a cursor: start from now
    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. The header is client input, so IntCursor lets only digits through, and the route starts anything else from now, as if there were no header. Not a 422: EventSource stops for good on any response that is not a 200, so one stale or malformed id (an old cursor format after a deploy) would end live updates in that tab. Do not hand-roll the check with str.isdigit(): it is true for ², which int() then rejects, so a client could turn the route into a 500.

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

Notify only for the writes a reader can see. By default every insert, update and delete notifies. A cursor-based read only finds new rows, so on a table with edits, read receipts or soft deletes, each update would wake every reader of that key for a read that returns nothing. Say which writes matter:

trigger_sql("messages", key_column="room", channel="room_change", order_column="id", on=("insert",))
# Inserts, deletes, and edits of body (the key column is watched too):
trigger_sql("messages", key_column="room", channel="room_change", update_of=("body",))

on takes any of "insert", "update" and "delete". update_of narrows updates to those whose SET list names one of the columns; the key column is always added, so a row that moves to another key still notifies. Run the SQL again with different options to change an installed trigger.

SQLAlchemy and Alembic

trigger_sql returns plain SQL, so it goes through whatever runs your migrations. With Alembic, install the trigger after the table exists in upgrade, and remove it before the table goes in downgrade (dropping a table drops its triggers but leaves the functions behind):

# alembic/versions/0001_messages.py
import sqlalchemy as sa
from alembic import op
from pg_sse import drop_trigger_sql, trigger_sql


def upgrade() -> None:
    op.create_table(
        "messages",
        sa.Column("id", sa.BigInteger, primary_key=True, autoincrement=True),
        sa.Column("room_id", sa.Integer, nullable=False),
        sa.Column("body", sa.Text, nullable=False),
    )
    op.execute(trigger_sql("messages", key_column="room_id", channel="room_change", order_column="id", on=("insert",)))


def downgrade() -> None:
    op.execute(drop_trigger_sql("messages", key_column="room_id", order_column="id"))
    op.drop_table("messages")

The SQL is several statements in one string. psycopg runs that; asyncpg refuses it ("cannot insert multiple commands into a prepared statement"). If your app uses asyncpg, point Alembic at postgresql+psycopg:// (psycopg 3 is already a dependency; the async migration template works with it too). Your app's own engine can stay on asyncpg.

read and start on an AsyncSession, wired the same way as the chat example. The room id is an integer here, and a key is always text, so the route passes str(room_id):

from functools import partial
from typing import Annotated

from fastapi import FastAPI, Header
from sqlalchemy import func, select
from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine

from pg_sse import Event, Hub, IntCursor

engine = create_async_engine("postgresql+asyncpg://...")  # the app's own connection, pooled however you like
Session = async_sessionmaker(engine, expire_on_commit=False)
hub = Hub(DIRECT_DATABASE_URL, channel="room_change")  # psycopg; a direct connection, not a pooled one
app = FastAPI(lifespan=hub.lifespan)
cursors = IntCursor()


async def read_after(room_id: int, after: int) -> list[Event]:
    async with Session() as session:
        query = select(Message).where(Message.room_id == room_id, Message.id > after).order_by(Message.id).limit(500)
        return [Event({"id": m.id, "body": m.body}, id=m.id) for m in await session.scalars(query)]


async def latest_id(room_id: int) -> int:
    async with Session() as session:
        return (
            await session.scalar(select(func.coalesce(func.max(Message.id), 0)).where(Message.room_id == room_id)) or 0
        )


@app.get("/rooms/{room_id}/stream")
async def stream(room_id: int, last_event_id: Annotated[str | None, Header()] = None):
    return hub.sse(
        str(room_id),  # not room_id: an int subscribes fine, never wakes, and pg_sse raises a TypeError for it
        read=partial(read_after, room_id),
        cursor=cursors.parse_one(last_event_id),
        start=partial(latest_id, room_id),
    )

Both blocks were run against a real database (an Alembic upgrade and downgrade, and a live stream on the psycopg and asyncpg drivers) but are not part of the test suite, so treat them as illustrative.

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
from pg_sse import TooManySubscribers


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.

When several causes arrive together, answer them in this order: start, then resync or reread, then keys, then signals. The first two already tell the page to refetch everything, so anything after them would repeat it. read may be async def: it then runs on the event loop with no thread hop, which suits a read that only looks at wake.

from pg_sse import Event, Wake


async 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
    events = []  # a write and a signal can arrive together: answer both
    if wake.keys:
        events.append(Event({"keys": sorted(wake.keys)}, event="change"))
    if "typing" in wake.signals:
        events.append(Event({}, event="typing"))  # refetch only the typing indicator
    return events


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


typing = hub.ephemeral("typing", ttl=8)  # at import time, next to the hub
# elsewhere, from any thread, on every keystroke:
typing.set(room_id, user_id)

signal_read builds that function for you, with the same order of causes:

from pg_sse import signal_read

read = signal_read(change="messages", signals={"typing": "typing"})
hub.sse(room, read=read)
# sends: "ready" first; "resync" after a reconnect or a quiet re-read; "messages" for a write; "typing" for that signal

A runnable version, with a typing indicator that lapses on its own and an inbox that watches every room: examples/inbox.py.

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.

Frames you will see

What hub.stream yields, and what EventSource does with each:

Ephemeral state

Typing and presence are true only while a client keeps saying so. hub.ephemeral keeps them for you: a signal per key and person that lapses unless renewed, and wakes the key's streams when it starts, is removed or lapses, so a closed tab never leaves a stale "typing" on another screen.

typing = hub.ephemeral("typing", ttl=8)

typing.set(room_id, user_id)  # a keystroke: wakes the room's streams the first time, then only renews
typing.set(room_id, user_id, on=False)  # they sent the message: removed, wakes
typing.live(room_id)  # ["ann", "bo"]: who has not lapsed, for the GET the page refetches

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. Each stream re-reads after its own random delay of up to resync_spread (2 s), so the reads do not arrive together. 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). It still sends what that read found, so its cursor moves on, then says behind; the browser reconnects and catches up.
• Caps open streams per process (500) and, when you pass owner (a user id, session, or client address, as text), per key and owner (5). Past a cap, hub.sse raises a 503 with Retry-After: 30. The per-owner cap counts one key at a time: an owner may hold 5 streams on every key it can name, up to the process cap. To bound one client's total, set max_total_per_owner (off by default). Without owner, a client is not counted at all and can hold every slot in a worker, so pass owner on public endpoints, and size the caps on the total rather than on the per-key number.
• Ends streams after 15 minutes. A stream is authenticated once, when it opens; this bounds how long a revoked session keeps one. To end it sooner, call hub.close(owner=...) when you revoke.
• 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(event.retry ?? 1_000);
}

API

Everything is importable from pg_sse. Each name below has a short description; the docstrings carry the rest.

Hub

Hub(conninfo, channel, *, max_subscribers=500, max_per_owner=5, max_total_per_owner=None, queue_size=64,
    max_stream_seconds=900, ping_seconds=(15, 25), reread_seconds=60, json_default=None, end_event=None,
    handle_signals=True, listen_timeout=5.0, resync_spread=2.0, application_name="pg-sse-listener",
    read_executor=None, on_error=None)

One per channel per process.

• conninfo: a DSN, or a function returning one. A function is called on every connect, so a rotating password or settings that are only readable at startup need no rebuilt hub.
• max_subscribers: open streams per process.
• max_per_owner: open streams per key and owner, so several tabs on one room are one client. It does not limit how many keys an owner opens streams on.
• max_total_per_owner: open streams per owner across all keys; None (the default) sets no total. This is the cap that keeps one client from filling max_subscribers.
• Both owner caps apply only to streams opened with an owner; anonymous streams are limited by max_subscribers alone. A refused stream is a TooManySubscribers, a 503 from hub.sse.
• queue_size: wake-ups queued per stream before it is ended as too far behind. At least 1. A resync takes no slot.
• max_stream_seconds: how long a stream lives before a planned close. math.inf turns the limit off.
• The durations here and in Ephemeral take any real number (int, float, Decimal, a numpy scalar) and are kept as float; a bool, a string or NaN raises ValueError.
• ping_seconds: the (low, high) range, jittered, between pings on an idle stream.
• reread_seconds: how long a stream may sit quiet before it re-reads anyway.
• json_default: json.dumps's default for Event data that is not JSON-native. json_default=str is the common choice for datetimes, UUIDs and Decimals.
• end_event: names the event sent with the reason before a planned close. Checked when it is given: an empty name, or one with a line break or NUL, raises ValueError (here and in stream/sse). Off by default; the retry: alone is always sent.
• handle_signals: chain onto SIGTERM and SIGINT so streams end at once on a deploy. Pass False in tests.
• listen_timeout: seconds lifespan waits for the first LISTEN before the app starts (None: do not wait). A finite number, 0 or more. It never fails startup.
• resync_spread: after the listener connects or reconnects, every stream re-reads. Each waits a random 0 to this many seconds first, so a database that has just come back is not hit by every open stream at once (up to max_subscribers reads queued behind your thread pool or connection pool). A notification for the stream's key during the wait starts the resync early (an all-keys stream keeps its time: every notification reaches it, so on a busy channel the first one would end all their waits at once). Any read in the meantime that re-reads everything counts as the resync: every read of a read(cursor), and a quiet re-read. A read(cursor, wake) woken by a hub.wake signal alone reads at once and still gets its resync at its time. A resync takes no queue slot, so it never ends a busy stream as behind. 0 reads at once. Pings and the periodic re-read are already jittered; this closes the one synchronised burst.
• read_executor: a thread pool (concurrent.futures.ThreadPoolExecutor; a ProcessPoolExecutor or InterpreterPoolExecutor raises ValueError, since a read runs with the caller's context variables) for plain (not async) read and start functions. By default they share asyncio's thread pool (at most 32 threads, also used by everything else that calls asyncio.to_thread), so a burst of stream re-reads can starve other work. Give the streams their own bounded pool: ThreadPoolExecutor(8, thread_name_prefix="sse-read"). Async reads do not use it, and the hub never shuts it down.
• on_error: on_error(key, exc), called when a stream fails, after the failure is logged: read or start raised, or read returned something the hub could not send (not an Event, data json.dumps refuses, a tuple id). Use it to report to an error tracker. A plain function runs in asyncio's default pool (not read_executor, where in an outage it would wait behind the failing reads), and an async def one on the loop, each as its own task, so a slow tracker never holds the failed response open, its slot taken, or the loop; leaving lifespan waits up to 5 seconds for those still running, then cancels them. An exception either raises is logged and ignored. The stream still ends without a goodbye.
• application_name: what pg_stat_activity shows for the listener's connection, to tell several services' listeners apart.

hub.lifespan

FastAPI(lifespan=hub.lifespan), or async with hub.lifespan(): inside your own lifespan. Chains onto an existing SIGTERM/SIGINT handler (a default action is left alone) and restores it when it exits. Starts without a database: the listener keeps retrying and listener_connected says whether it is live. On exit every stream says goodbye with reason shutdown, including one created but not read yet, even if it is first read after the hub has started again. After it, hub.stream and hub.sse refuse new streams as during shutdown (a 503 from hub.sse).

hub.stream

hub.stream(key, read, *, cursor=None, start=None, owner=None, end_event=None, max_stream_seconds=None) -> AsyncIterator[str]

The primary API: an async iterator of encoded SSE frames, for any framework. Raises TooManySubscribers when it is called, before any response exists: the stream takes its slot then, not when its first frame is read, and gives it back when it ends, or, for a response that was never sent (the route raised after making it, or the client left before the body started), when it is garbage-collected (at once on CPython) and 30 seconds after it was made at the latest. A server reads a stream the moment it sends the response, so one read after that ends at once, with no goodbye.

• key: what the stream watches, or None for every key on the channel. It must be a str, the text form of the key column (str(room_id) for an integer column): the notification payload is key::text, a NULL key notifies "", and any other type raises a TypeError.
• read: takes (cursor) or (cursor, wake) and returns a list of Event, not a generator, which would run its queries on the event loop. A plain function runs in a thread; an async one runs on the loop. Its type is pg_sse.Read.
• cursor: where to start, already parsed from the client (Last-Event-ID, a query parameter). IntCursor parses and checks an integer one. Text with a line break or NUL cannot be an SSE id: the stream then starts from now, with a warning at most once a minute, rather than answering the client with a 500. A cursor of the wrong type (a row, bytes, a bool) raises TypeError.
• start: gives the current cursor when cursor is None, so the stream starts from now. Plain or async. Its type is pg_sse.Start.
• owner: a user id or session, for the per-owner caps and hub.close(owner=...). A str, or an int, which is turned into text, so hub.close(owner=7) and hub.close(owner="7") find the same streams. Any other type (a UUID, a bool) raises TypeError: pass str(owner).
• end_event: overrides the hub's.
• max_stream_seconds: overrides the hub's age limit for this stream, so a visitor's stream can expire sooner than a dashboard session's. Must be greater than 0; math.inf for no limit.

hub.sse

hub.sse(key, read, *, cursor=None, start=None, owner=None, end_event=None, max_stream_seconds=None,
        headers=None, refuse_with=None)

stream as a Starlette/FastAPI response (needs the starlette extra). Past a cap it raises a 503 with Retry-After: 30; refuse_with(err) returns your own exception to raise instead, e.g. one your error tracker ignores. headers are added to the response.

Event

Event(data, id=None, event=None, final=False)

• data: a str as is, anything else as JSON.
• id: the cursor after this event, handed back to read unchanged.
• event: the SSE event name; None means the browser's message.
• final: end the stream after this event (a closed conversation).

Wake

Wake(start=False, keys=frozenset(), signals=frozenset(), resync=False, reread=False, signal_keys={})

Why read runs, for a read(cursor, wake). See Signal streams for the fields. wake.data_may_have_changed is True unless the only causes are hub.wake signals.

hub.close

hub.close(*, owner=None, key=None)

Ends this process's streams opened with owner and/or on key, so a removed operator, a blocked visitor or a revoked session does not keep a stream for up to max_stream_seconds. That includes a stream that hub.sse has created but whose first frame has not been read yet, a stream whose read is running (an async def read is cancelled, so the goodbye goes out at once; a plain one runs on in its thread and what it found is not sent), and a stream partway through sending a batch (the rest is not sent). Each stream sends a goodbye with reason revoked and ends. At least one argument is required; hub.close_all() ends every stream. owner is a str or an int, as for stream; any other type raises TypeError.

• When it acts. Safe from any thread. Called on the event loop it takes effect before it returns; from another thread it is handed to the loop. It does nothing, silently, while the hub is not running, when there are no streams to end.
• Who it reaches. A stream opened after the call (the user signed in again) is not ended with the old ones, even when the call came from another thread and reaches the loop later. A route that checked the session before the revoke but opens its stream after the call is not reached; max_stream_seconds bounds that window.
• Slots. An ended stream's slot is free once it has said goodbye, or 10 seconds after it ended, whichever comes first. Until then it still holds a connection and counts against the caps and in open_streams, so a client that stops reading cannot open one stream after another, while a half-open connection (whose send blocks until TCP gives up) holds it no longer than that. A reconnect within that time can be refused with a 503. The same holds for a stream dropped for falling behind or ended by shutdown.
• A dropped stream past the grace. Once its slot is free the hub forgets it. A dropped stream whose read is still running then is out of reach of a shutdown that starts later: it says behind, not shutdown, and reconnects a little sooner.
• Not for one tab per user. Do not call it from the stream route to keep one tab per user: the closed tab reconnects a second later and its route closes the other, back and forth for as long as both are open.

await db.revoke_session(session_id)  # first: the reconnect below must be refused
hub.close(owner=str(user_id))

• Revoke in the database first. The client reconnects after the goodbye, and the route must refuse it; EventSource stops for good on a 401 or 403.
• With both arguments, only streams matching both end. A stream watching every key (key=None) matches by owner only.
• Only streams opened with owner= can be found by owner. Pass owner on every authenticated stream.
• It reaches this process only. With several workers, call it in each one; nothing is broadcast.

hub.wake

hub.wake(key, signal="wake", *, all_keys=True)

Wakes one key's streams for a change that is not a database write (someone is typing). Safe from any thread, and reaches only this process's streams. signal shows up in wake.signals; all_keys=False leaves the all-keys streams alone. key and signal must be str: a signal is matched against text labels, so an Enum would never match.

hub.call_later

hub.call_later(seconds, callback, *args)

Runs callback(*args) on the hub's event loop after seconds, for follow-up work on ephemeral state (check that a typing signal lapsed, then hub.wake). Safe from any thread; does nothing while the hub has no running loop.

signal_read

signal_read(*, start="ready", resync="resync", change="change", change_data=None, signals=None) -> Read

An async read(cursor, wake) for streams that tell the page what to refetch instead of sending rows. It runs on the event loop with no thread hop. An empty event name raises ValueError (the browser would dispatch it as a plain message) and a name that is not a str raises TypeError; pass None to send nothing. With resync=None, a resync that arrives with notified keys or signals still sends their events. Every event's data is {}, except change when change_data is given and a signal mapped to (event_name, data_fn).

Causes are answered in this order, because the first two tell the page to refetch everything and anything after them would repeat it:

1. start: only the start event.
2. resync or reread: only the resync event.
3. Otherwise the change event if keys were notified, then one event per mapped signal that arrived, in the order of signals (so the result is deterministic). A write that coincides with a signal sends both.

A cause with no mapping, or mapped to None, contributes nothing, so the result may be [].

IntCursor

IntCursor(parts=1, *, sep=".", max_digits=15)

Parses and formats a cursor of parts non-negative integers joined by sep: 42, or 12.3 for a feed that spans two tables.

max_digits=15 is exact in a JavaScript number. For longer ids (snowflake ids; a bigint has up to 19 digits) pass max_digits=19; format raises ValueError for a value with more digits than max_digits. Both parse and format raise ValueError for a part above the bigint maximum (9223372036854775807), so a 19-digit cursor never reaches the query out of range.

The constructor raises ValueError for parts or max_digits that is not an int (a bool included), parts < 1, max_digits outside 1 to 19, or a sep that is not a str, is empty, or holds a digit, a line break or NUL.

distinct

distinct(read, key)  # -> a read

Wraps read so a snapshot is sent only when it changed. key(event) summarises what the event shows, as a hashable value; an event whose key equals the key of the last one sent is dropped. None means "always send" (the next event is then compared with nothing, so it is sent too), and a final event is always sent. read may be read(cursor) or read(cursor, wake), sync or async; the wrapper takes the same arguments and is always async. A sync read runs in the hub's read_executor inside a stream, and in asyncio's default pool when you call the wrapper yourself (from a hub.subscribe loop, say), so it never blocks the loop. key runs on the event loop either way, once the events are back from read, so keep it cheap.

@app.get("/rooms/{room}/stream")
async def stream(room: str):
    # a read that returns a snapshot every time it runs; the page is told only when it differs
    return hub.sse(room, read=distinct(partial(snapshot, room), key=lambda e: e.data["unread"]))

Three things to get right:

• The state is per stream. Call distinct inside the route, once per request. Built once at import time, one wrapper would be shared by every stream, and a stream that has never been sent a snapshot would be told nothing.
• A dropped event's id is never applied. The hub advances the cursor only on events it yields, and the browser records only ids it receives. Drop only events whose id would not have advanced the cursor: a snapshot with no id, or one that repeats the last. Never drop an event that carries a new row, or a reconnect would replay from before it: make key change with every new row (next point), or, failing that, return None for any event that carries rows.
• None forgets the last key. After an event that is always sent, the next snapshot is sent again even if it is unchanged. Where you can, make key cover everything the client is shown, including the newest message or row id: a new row then changes the key by itself and None is not needed.

hub.ephemeral and Ephemeral

hub.ephemeral(name, *, ttl, signal=None, grace=0.25, all_keys=False, clock=time.monotonic) -> Ephemeral

Signals per key that lapse unless renewed (typing, presence), announced with hub.wake. hub.ephemeral(...) is Ephemeral(hub, name, ...).

set and clear are safe from any thread; key must be a str. Each renewal schedules one lapse check with hub.call_later; a check that finds the signal renewed leaves it alone, and the renewal's own check handles it later. State lives in this process only, so with several workers a signal set on one is invisible to the streams on the others, and it is lost on restart. While the hub is stopped nothing checks, so entries linger until the next set or clear for their key (live still skips them).

hub.stats

A frozen HubStats: open_streams, refused_total, dropped_total, listener_connected, listener_reconnects_total, last_notification_at (Unix time or None), read_errors_total (streams that failed, as on_error reports them) and ended_total, a frozen EndedTotals of goodbyes sent by reason: max_age, shutdown, revoked and behind. "Streams dropped for being slow" is dropped_total; ended_total.behind counts the ones that then said so, which leaves out a dropped stream that a close or shutdown reached first, or whose client left. A goodbye is counted once the client took it, so a stream whose client left, or that ended with a final event or a failure, is in none of them. dataclasses.asdict(hub.stats()) gives all of it as a dict, ended_total as a nested one. The _total fields only grow, so a metrics exporter can read them on a timer and publish them as counters. They are per process and start at zero. pg-sse ships no exporter code, since people use different ones.

hub.streams

hub.streams() -> list[StreamInfo]

Every subscription still holding a slot, oldest first, as StreamInfo(key, owner, age_seconds, ended) (key is None for an all-keys stream, owner is None for one opened without it, and ended is true for one closed or dropped that is still saying goodbye: it counts against the caps until it has). For an admin view or a debug endpoint; this process only. It copies a set the event loop changes: safe from another thread under CPython's GIL, but on a free-threaded build call it on the event loop.

hub.listening

A threading.Event, set while the listener holds its LISTEN. hub.stats().listener_connected says the same.

hub.subscribe and hub.unsubscribe

hub.subscribe(key, owner=None) -> Subscription
hub.unsubscribe(subscription)

The low-level queue under stream, for a custom stream loop. key and owner are checked like stream's. Unsubscribe in a finally.

await subscription.wait(timeout) returns everything queued once there is anything, or None on timeout. An item is ("notify", key) for a database notification, ("wake", key, signal) for a hub.wake, or ("marker", name), which means re-read everything. Once subscription.ended is true the stream is over: it fell behind, the hub is shutting down (hub.closing), or hub.close ended it. After a listener connect or reconnect the resync marker arrives at once: resync_spread applies only to hub.stream and hub.sse, so a custom loop that serves many streams should add its own jitter.

Read and Start

The types of read and start: Read is Callable[..., Iterable[Event] | Awaitable[Iterable[Event]]] and Start is Callable[[], Any]. Use them to annotate a helper that wraps hub.sse:

from pg_sse import Hub, Read, Start


def room_stream(hub: Hub, room_id: int, read: Read, start: Start):
    return hub.sse(str(room_id), read=read, start=start)

trigger_sql

trigger_sql(table, key_column, channel, *, order_column=None, on=("insert", "update", "delete"), update_of=None)

The SQL that wires a table to a channel, to run once in a migration. Postgres 14+.

• Names are plain lower-case identifiers; table may be schema.table.
• order_column must be a serial, bigserial or identity column; installing fails with a clear error otherwise.
• on limits which writes notify: any of "insert", "update", "delete".
• update_of limits updates to those whose SET list names one of the columns (plus the key column). Needs "update" in on.
• It creates a function and a trigger named "pg_sse$notify$<table>$<key>" (and "pg_sse$order$<table>$<order_column>"), which is what you will see in \d. Running it again replaces them.
• The order trigger looks up the column's sequence once, when the SQL runs, and writes its name into the function, so an insert does no catalog search while it holds the key's lock. Dropping and recreating the sequence under the same name is fine; if you rename it or move it to another schema, run trigger_sql again.

drop_trigger_sql

drop_trigger_sql(table, key_column, *, order_column=None)

The SQL that removes what trigger_sql created, for a downgrade migration. It does not need channel, on or update_of.

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
    ...

A fixture keeps that in one place. It needs an async plugin: anyio, with the tests marked @pytest.mark.anyio, or pytest-asyncio with asyncio_mode = "auto" (in its default strict mode, decorate it with @pytest_asyncio.fixture instead):

@pytest.fixture
async def running_hub():
    async with hub.lifespan():
        yield hub

To read a stream without a server, run the hub on the test's event loop with async with hub.lifespan(): (leaving it forgets every subscription, so the next test starts clean) and read frames with pg_sse.testing:

from pg_sse.testing import goodbye, next_event


async def test_a_new_message_arrives():
    async with hub.lifespan():
        frames = hub.stream("lobby", read=read_after_lobby, cursor=0)
        insert_message("lobby", "hi")
        assert (await next_event(frames)).json() == {"id": 1, "body": "hi"}

The timings are plain attributes, so a test can shorten them after building the hub and its signals: hub.ping_seconds = (0.2, 0.2), hub.reread_seconds = 0.2, and typing.ttl = 0.2 on an Ephemeral. Set them before the first stream opens; the constructor's checks do not run on assignment, so keep ping_seconds a (low, high) tuple with 0 < low <= high. Call typing.reset() between tests that share an Ephemeral. Set resync_spread=0 if a test expects the re-read after a listener reconnect to come at once.

next_event skips pings and the id-only frame that opens a stream with a cursor. Keep that: a ping comes before any frame a re-read produces, because every idle timeout pings, even the one that leads to the re-read, and before a goodbye when the age limit and the ping timeout coincide. A test that expects the very next frame to be the event is flaky without it. Use next_frame when the ping or the id is what you are testing.

pg_sse.testing

Reading a stream in tests, with no test framework and no database: the helpers take any async iterator of frames (what hub.stream returns, or a response's body_iterator). See Testing your app.

Waiting past timeout raises TimeoutError naming the frames seen so far; a stream that ends early fails with an AssertionError that does the same. A timed-out read cancels the generator, so the stream is over afterwards.

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.

The exception is hub.wake: it reaches only the streams of the process that calls it, since nothing is written to the database. State that lives in one worker's memory, which includes everything in hub.ephemeral (the typing indicator in the inbox example), is invisible to readers on the other workers; keep it where every worker can see it before you run more than one.

Logging

pg-sse logs to the standard-library logger named pg_sse and adds no handlers, so configure it like any other. At WARNING it logs a stream refused at a cap or during shutdown (back-pressure, not a fault) and the listener losing its connection (with the traceback and the retry delay). It also warns, at most once a minute, about a client cursor with a line break or NUL (the stream starts from now) and about an id longer than an IntCursor's max_digits. At ERROR it logs, with a traceback, a stream whose read or start raised or returned something other than Event items (that stream ends and the client reconnects), and an on_error callback that failed.

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). If it does not need to show them, stop them from notifying: trigger_sql(..., on=("insert",)), or update_of=(...) to wake only for edits of the columns a reader shows. Otherwise every update wakes every reader of that key for a read that returns nothing. A write the trigger leaves out never wakes a stream; the stream finds it on its next re-read.
• A key is text. hub.stream, hub.sse, hub.subscribe and hub.wake raise a TypeError for a key that is not a str, because a notification carries the key column as text and an int would subscribe and never wake. Pass str(room_id).
• Sync reads run in a thread pool. A plain-function read or start runs in asyncio's default executor (min(32, cpu + 4) threads) unless you pass read_executor, 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).
• Ephemeral state is per process. hub.ephemeral keeps its signals in memory: with several workers a signal set on one is invisible to the streams on the others, and a restart loses it. That costs nothing for typing or presence; anything that must be seen everywhere belongs in a table with a trigger.
• 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 does not listen 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. llms.txt carries the same rules in a form meant to be fed to an agent.

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. A module-level IntCursor()'s parse_one(header) returns the int or None (start from now; a 422 stops EventSource for good); never str.isdigit() then int().
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.ephemeral(name, ttl=...) for ephemeral per-key state (typing, presence): set(key, who) on every keystroke or heartbeat, live(key) to answer the refetch. For other changes that are not database writes, hub.wake(key, signal=..., all_keys=False); 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.
13. The key is always a str: pass str(room_id) for an integer or UUID column. Anything else raises a TypeError.
14. On a table with edits, read receipts or soft deletes, use trigger_sql(..., on=("insert",)) so only new rows wake readers.
15. In an Alembic migration, op.execute(trigger_sql(...)) in upgrade and op.execute(drop_trigger_sql(...)) in downgrade, on the psycopg driver (asyncpg rejects the multi-statement SQL).

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