Best Practices for Instrumenting Phoenix in FastAPI Websockets

·Apr 01, 2025 10:49 PM

RunLLM What are some best pracitces for instrumenting phoenix in a fastapi backend connected to frontend via websocket i have 4 apis exposed from fast api create_session - http get_session - http delete_session - http process_session_messages - ws websocket stays open until one of 2 conditions is reached 1) inactivity for an extended period of time 2) user changes to a different chat/homepage general flow goes create -> get (to show user latest state) -> process_session_messages (when there is a new user message) -> keep ws open for response + send additional user messages -> user moves away and we close the connection I am connecting all this using a session_id that is generated in create now my issue is that the first 2 steps come as traces correctly the third one only shows up as a trace + connected to session ONLY after the websocket completely disconnects, this make sense since the backend is listening and the connection is open, however this is an issue for me since we can have long running sessions and i wanna see traces of process_session_messages after every turn is done, how can i achieve this? currently doing this for every route's method

@app.websocket(
    "/api/sessions/{session_id}/ws"
)
async def websocket_endpoint(
    websocket: WebSocket, organization_id: str, environment_id: str, session_id: str
):
    with tracer.start_as_current_span(name="process_session_messages"):
        current_span = trace.get_current_span()
        current_span.set_attribute(SpanAttributes.SESSION_ID, session_id)

        with using_session(session_id):
            more code

@app.websocket("/api/sessions/{session_id}/ws") async def websocket_endpoint( websocket: WebSocket, organization_id: str, environment_id: str, session_id: str ): await websocket.accept() while True: try: # Receive message from frontend message = await websocket.receive_text() # Start a new span per message with tracer.start_as_current_span("process_single_message") as span: span.set_attribute("session_id", session_id) span.set_attribute("message_length", len(message)) # Process the message here await handle_message(message, websocket, session_id) except WebSocketDisconnect: break

Since i'm using FastApiInstrumentor (cuz we want all spans under the parent span of the "route" i get the following : Say a user lands on chat page and is gonna be there for approx 10-15 mins sending multiple user messages back and forth This whole back and forth (all the spans it emits) comes under one parent span (this is fine and how we want it) Parent Span with name /api/sessions/{session_id}/ws - A process_session_messages - B process_one_message - C All A, B, and C are set with session_id attribute C correctly closes when one message is processed and emits a "trace" with "spans" B correctly shows up as one of the C's ends and shows up as a "trace" with "spans" and child "trace" (The trace from C) However B, C don't associate to the session when they close i.e on session page , i don't see B , C as traces only A (once it closes) therefore only if the user refreshes or moves to a different page does the websocket close which is when A closes and shows up on sessions page

10 comments