Repository navigation
bare websockets - #6932
bare websockets#6932benedikt-bartscher wants to merge 99 commits into
Conversation
Merging this PR will degrade performance by 8.1%3 warnings
Warning Please fix the performance issues or acknowledge them on CodSpeed. Performance Changes
Tip Investigate this regression by commenting Comparing Footnotes
|
[High impact] The changes since the previous review appear safe to merge. SummaryReplaces the default Socket.IO connection with a plain WebSocket and adds channels for application messages and binary data.
Reviews (84) · Last reviewed commit: "Build the hydrate-too-large error in one..." · Reviewed by Greptile |
There was a problem hiding this comment.
All reported issues were addressed across 19 files
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
…ace, awaited disconnect cleanup; revert server-wide ws_max_size
There was a problem hiding this comment.
All reported issues were addressed across 17 files
Requires human review: Auto-approval blocked because this review re-detected 1 unresolved issue already reported by Cubic.
Re-trigger cubic
There was a problem hiding this comment.
All reported issues were addressed across 4 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
… message shape, log protocol closes; add transport benchmarks
…cut client watchdog churn
There was a problem hiding this comment.
All reported issues were addressed across 4 files (changes from recent commits).
Tip: Review your code locally with the cubic CLI to iterate faster.
Re-trigger cubic
There was a problem hiding this comment.
All reported issues were addressed across 6 files (changes from recent commits).
Tip: Review your code locally with the cubic CLI to iterate faster.
Re-trigger cubic
…icorn] extra with gunicorn
There was a problem hiding this comment.
All reported issues were addressed across 6 files (changes from recent commits).
Tip: Review your code locally with the cubic CLI to iterate faster.
Re-trigger cubic
The benchmark socket client builds on reflex-bench's wire codec, and two unit tests that also run on Windows imported it. Importing any part of reflex-bench makes coverage count the whole Linux-only package, whose own tests are ignored off Linux, which pulled Windows coverage under the threshold. Run those two tests on Linux only and import the client inside them.
Both transports now open sessions through BaseEventNamespace: one connect-and-boot sequence and one handler failure policy. A failed boot event ends the session on either transport (the plain WebSocket closes with 1011), so the client retries rather than waiting for a hydrate that never comes. Undeserializable input ends a session quietly, and other handler failures are logged within the error budget; Socket.IO's on_event, on_ping and on_client_error no longer raise into python-socketio. Socket.IO packets are serialized like plain frames, so an unpaired surrogate is escaped instead of breaking the send. Also: one fan-out body for channel room and token sends, a socket policy read once per connection for both transports, a supports_channels flag instead of re-reading the config, the embedded test server capped like the Granian launchers, CORS strings parsed as the environment variable is, and an unreachable uvicorn warning removed. The Socket.IO namespace tests move from test_app.py to test_socketio_namespace.py, and the namespace tests share fixtures and a serve() helper.
websocket.js settles a closed connection in one place, flushes the channel and transport queues through one helper, and checks and reports oversized frames through one size check and one message. The single-caller emitChannel, which bypassed the channel's guards, is gone, and the generic JSON helpers move to json.js. state.js shares its warm-socket, socket disposal and connect-error helpers and drops an unreachable check. The client-only tests that ran websocket.js through the Python harness move to websocket_js.test.mjs, and the channels integration test now queues its frame on the transport, as its name says.
The benchmark socket client detects a backend's protocol once instead of per connection, so Socket.IO reconnect timings no longer include a failed WebSocket upgrade per client, and codecs carry their bare connect and ping frames instead of callers branching on the protocol. reflex-bench shares its process spawn, stop and report handling, its worker wrapper and small session helpers, and drift tests pin the protocol constants it copies from Reflex. The bench tests share their fixtures through factories.py.
There was a problem hiding this comment.
All reported issues were addressed across 31 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Re-trigger cubic
The echo server announced its generator's 16 MiB frame limit as its message size, while a default Reflex app (and its own Socket.IO greeting) announces 1 MB; the drift test only compared keys. It now announces Reflex's default on both protocols, and the test compares the whole handshake against a default-configured app. detect_protocol always waited up to 30 seconds to open its websocket, so the benchmark socket client's first connection to a hung backend ignored the caller's timeout. Detection now takes the timeout, and the client passes its own through.
Port upstream's repeated-header combining (reflex-dev#7588) from the removed reflex/app.py helper to _decode_asgi_headers in reflex/event_namespace.py, which both transports share, keeping the branch's Latin-1 fallback. Its tests move from tests/units/test_app.py to tests/units/test_event_namespace.py.
Granian sets TCP_NODELAY on listeners it binds itself, but the harness binds its own socket, so with the Redis state manager each update stalled ~40 ms on delayed ACKs and test_client_side_state raced it. Also poll the client storage reads that follow an event or navigation.
Port upstream's reconnect back-off (reflex-dev#7605) to the transport abstraction: retryConnect keeps the branch's addConnectError helper and caps the attempt count at MAX_CONNECT_ERRORS. Socket.IO reports a transport closed before the namespace connected as a disconnect, which now retries with back-off; the plain WebSocket reports it as connect_error, so it gets the same back-off. The state.js back-off tests run over both transports, with the fake socket reporting a close before connecting the way each transport does, and read connect errors through upstream's lazily applied updaters.
Port upstream's EventNamespace changes to the moved namespace modules: the trusted proxy hops for the client IP (reflex-dev#7614) apply in build_static_router_data, and the cancellation of a disconnected socket's page-load chains (reflex-dev#7415) runs in BaseEventNamespace.handle_disconnect, so both transports get them. The branch already decoded non-UTF-8 header values (reflex-dev#7602). Their tests move from tests/units/test_app.py to tests/units/test_event_namespace.py, and the transport benchmark's app double gains the event processor attribute the disconnect reads. Port the oversized hydrate check of reflex-dev#7551 to the Socket.IO transport only: state.js wraps the socket.io encoder where it creates the Socket.IO socket, since the plain WebSocket has no engine.io encoder. Its state.js tests run over Socket.IO, whose fake socket now encodes CONNECT packets, and the reconnect back-off of reflex-dev#7605 stays in place for both transports. uv.lock is upstream's, relocked for the branch's dependencies.
The plain WebSocket sends its hydrate with the connect frame, before any handshake could announce the server's message limit. A frame over the limit closed the socket with 1009, which the client reported as a generic connection failure and retried with back-off forever. The server now names the limit in that close's reason, and the transport reports such a close as the hydration_payload_too_large error reflex-dev#7551 added for Socket.IO, with the frame's size and the limit when the reason gives it. A reconnect over a limit an earlier handshake announced is refused before it is sent. retryConnect treats the error as final for both transports: it records it once, schedules no retry and keeps wait_connect set, so reconnect() does not rebuild the hydrate, and the toast and banner keyed on the error code show it.
When a token moved to another socket, the transport dropped the stale recipient but send_to_token still returned True; the sender now reports whether it accepted anyone and the fan-out returns that.
There was a problem hiding this comment.
1 issue found across 4 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="reflex/event_namespace.py">
<violation number="1" location="reflex/event_namespace.py:1136">
P2: `bool(recipients)` reports success for linked SIDs even when `_deliver` skips the connection or `_Connection.send` rejects it after close. Return the number of recipients actually accepted by `_deliver` (or otherwise check the live connection state) so `send_to_token` does not report a delivery that never reached a socket.</violation>
</file>
Reply with feedback, questions, or to request a fix.
View guided diff | Re-trigger cubic
| # teardown forgets its channel sessions. | ||
| connection.close(1008) | ||
| self._deliver(recipients, payload, event) | ||
| return bool(recipients) |
There was a problem hiding this comment.
P2: bool(recipients) reports success for linked SIDs even when _deliver skips the connection or _Connection.send rejects it after close. Return the number of recipients actually accepted by _deliver (or otherwise check the live connection state) so send_to_token does not report a delivery that never reached a socket.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At reflex/event_namespace.py, line 1136:
<comment>`bool(recipients)` reports success for linked SIDs even when `_deliver` skips the connection or `_Connection.send` rejects it after close. Return the number of recipients actually accepted by `_deliver` (or otherwise check the live connection state) so `send_to_token` does not report a delivery that never reached a socket.</comment>
<file context>
@@ -1130,6 +1133,7 @@ async def _send_channel_message(
# teardown forgets its channel sessions.
connection.close(1008)
self._deliver(recipients, payload, event)
+ return bool(recipients)
async def _send_channel_error(
</file context>
A linked session whose connection was closing, overflowed its send backlog, or was already gone still counted as a recipient; the connection now reports whether it queued the frame.
There was a problem hiding this comment.
All reported issues were addressed across 2 files (changes from recent commits).
Reply with feedback, questions, or to request a fix.
View guided diff | Re-trigger cubic
A frame that overflowed the send backlog was queued before the connection closed, so the writer could still deliver a send reported as a miss.
…g it An app that replaced EventNamespace and overrode on_event saw every event but the hydrate on connect. The boot event goes through on_event again, re-raised while booting so a failure still refuses the connect, and overriding on_event now warns in favor of handle_event, which runs for every event on both transports.
The added tests checked fragments of errors, warnings and log output; compare each in full, with session ids normalized to a placeholder and the expected traceback output derived from the raised exception. Capture the token-loss test's log on the reflex logger, so its check still sees debug records when that logger has a level of its own.
The websocket library and frame size warnings were checked by fragment; compare the captured records in full.
There was a problem hiding this comment.
1 issue found across 8 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="tests/units/test_app.py">
<violation number="1" location="tests/units/test_app.py:5438">
P2: These assertions inspect every captured log record, so an unrelated warning can fail the test despite correct channel behavior. Filter `caplog.records` by `record.name == "reflex.app"` and `record.levelno == logging.WARNING` before comparing the exact message in both tests.
(Based on your team's feedback about scoping app log assertions.) [6b5f533c-bd1d-446a-abb1-3cd757d9f8b0]</violation>
</file>
Reply with feedback, questions, or to request a fix.
View guided diff | Re-trigger cubic
| app.register_channel(_ProbeChannel()) | ||
|
|
||
| assert app._channels == {} | ||
| assert caplog.record_tuples == [ |
There was a problem hiding this comment.
P2: These assertions inspect every captured log record, so an unrelated warning can fail the test despite correct channel behavior. Filter caplog.records by record.name == "reflex.app" and record.levelno == logging.WARNING before comparing the exact message in both tests.
(Based on your team's feedback about scoping app log assertions.) [6b5f533c-bd1d-446a-abb1-3cd757d9f8b0]
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At tests/units/test_app.py, line 5438:
<comment>These assertions inspect every captured log record, so an unrelated warning can fail the test despite correct channel behavior. Filter `caplog.records` by `record.name == "reflex.app"` and `record.levelno == logging.WARNING` before comparing the exact message in both tests.
(Based on your team's feedback about scoping app log assertions.) [6b5f533c-bd1d-446a-abb1-3cd757d9f8b0]</comment>
<file context>
@@ -5434,7 +5435,17 @@ def test_register_channel_skips_without_the_event_websocket(
assert app._channels == {}
- assert "needs the event websocket" in caplog.text
+ assert caplog.record_tuples == [
+ (
+ "reflex.app",
</file context>
The app double's enqueue returned None, so in dev mode handle_event's add_done_callback raised AttributeError on every inbound event and both inbound benchmarks measured the failure policy instead of the transport. It now returns a resolved future, like the EventFuture enqueue returns, and every transport benchmark fails if anything was logged at error level. Walltime, 100 inbound events: websocket 146 us -> 140 us, Socket.IO 318-324 us -> 313-317 us.
The error code and message for a hydrate over the backend's limit were defined in both websocket.js and state.js. websocket.js now exports HYDRATION_PAYLOAD_TOO_LARGE and hydrationTooLargeError(size, limit, unit), which the socket.io path in state.js uses too; the messages are unchanged. The state.js harness builds the error with the real helper and pins the socket.io message exactly.
saves ~11.6 KB gz / ~41 KB raw frontend size (a small js chunk replaces whole socketio)
dropped python deps: wsproto, simple-websocket, python-socketio, python-engineio, bidict
new benchmarks in test_event_transport.py show about 2x performance increase compared to socketio
also migrates AppHarness to granian for consistency. otherwise downstream tests using AppHarness still need uvicorn and wsproto/websockets
related reflex-dev/xy#513