fix(observability): Phoenix 는 OTLP protobuf 만 받음(JSON 415) — opentelemetry-proto 로 직렬화, 로컬 Phoenix 실측 통과

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-22 15:43:26 +09:00
parent e60c5bc273
commit 68b55907f0
4 changed files with 47 additions and 9 deletions
@@ -3,6 +3,7 @@
 대상은 설정으로 고름(둘 다 켜도 됨):
  LANGFUSE_HOST → {host}/api/public/otel/v1/traces (Basic 인증, langfuse.* 속성)
  PHOENIX_HOST  → {host}/v1/traces (인증 없음, OpenInference 속성: openinference.span.kind, input.value …)
+                  Phoenix 는 JSON 을 안 받아서(415, 2026-09-22 실측) protobuf 로 바꿔 보냄 — opentelemetry-proto 필요
 속성은 두 벌을 같은 span 에 같이 실음 — 각자 자기 것만 읽고 나머진 metadata 로 떨어짐.

 v4 는 옛 /api/public/ingestion 이 막혀서(score 만) OTel 엔드포인트로 감. 트레이스 하나 = 루트 span(trace 속성) + 자식 span(generation).
@@ -186,6 +187,24 @@ def usage_of(inp: int | None, out: int | None) -> dict | None:
    return {"input": inp or 0, "output": out or 0, "total": (inp or 0) + (out or 0)}


+def _to_protobuf(body: dict) -> bytes:
+    """OTLP JSON → protobuf 바이트. traceId/spanId 는 JSON 이 hex, proto-JSON 은 base64 라 바꿔 넣음."""
+    import base64
+    import copy
+
+    from google.protobuf.json_format import ParseDict
+    from opentelemetry.proto.collector.trace.v1.trace_service_pb2 import ExportTraceServiceRequest
+
+    b = copy.deepcopy(body)
+    for rs in b["resourceSpans"]:
+        for ss in rs["scopeSpans"]:
+            for sp in ss["spans"]:
+                for k in ("traceId", "spanId", "parentSpanId"):
+                    if k in sp:
+                        sp[k] = base64.b64encode(bytes.fromhex(sp[k])).decode()
+    return ParseDict(b, ExportTraceServiceRequest()).SerializeToString()
+
+
 def _otlp_body(spans: list[dict]) -> dict:
    project = settings.LANGFUSE.get("project") or "codeassist"
    return {"resourceSpans": [{
@@ -200,17 +219,21 @@ async def send(spans: list[dict]) -> bool:
    if not spans or not enabled():
        return False
    body = _otlp_body(spans)
-    targets: list[tuple[str, str, dict, tuple | None]] = []
+    # (이름, url, headers, auth, protobuf 여부)
+    targets: list[tuple[str, str, dict, tuple | None, bool]] = []
    if cfg.get("host"):
        targets.append(("langfuse", cfg["host"].rstrip("/") + "/api/public/otel/v1/traces",
-                        {"x-langfuse-ingestion-version": "4"}, (cfg.get("public_key", ""), cfg.get("secret_key", ""))))
+                        {"x-langfuse-ingestion-version": "4"}, (cfg.get("public_key", ""), cfg.get("secret_key", "")), False))
    if cfg.get("phoenix"):
-        targets.append(("phoenix", cfg["phoenix"].rstrip("/") + "/v1/traces", {}, None))
+        targets.append(("phoenix", cfg["phoenix"].rstrip("/") + "/v1/traces", {"Content-Type": "application/x-protobuf"}, None, True))
    ok = False
-    for name, url, headers, auth in targets:
+    for name, url, headers, auth, pb in targets:
        try:
            async with httpx.AsyncClient(timeout=5.0, transport=TRANSPORT) as client:
-                resp = await client.post(url, json=body, headers=headers, auth=auth)
+                if pb:
+                    resp = await client.post(url, content=_to_protobuf(body), headers=headers, auth=auth)
+                else:
+                    resp = await client.post(url, json=body, headers=headers, auth=auth)
            if resp.status_code != 200:
                log.warning("%s ← %s %s", name, resp.status_code, resp.text[:200])
                continue
@@ -24,8 +24,12 @@ run.sh 가 하는 것: `PHOENIX_PORT=8915 PHOENIX_WORKING_DIR=/www/phoenix/data`

 ## 3. 백엔드 연결

- -12 `code-assistant/5_django_backend/.env`: `PHOENIX_HOST=http://127.0.0.1:8915` → `bash deploy.sh`
- -13 `abap-specgen/web/deploy/bare/.env`: `PHOENIX_HOST=http://10.196.81.34:8915` → `run.sh restart`
+Phoenix 는 protobuf 로만 받아서 백엔드 venv 에 `opentelemetry-proto`(+protobuf) 가 있어야 함 — wheel 은 같은 묶음에 들어 있음.
+
+- -12: `cd /www/abap-ito/code-assistant/5_django_backend && .venv/bin/pip install --no-index --find-links /www/phoenix-wheels opentelemetry-proto`
+  → `.env` 에 `PHOENIX_HOST=http://127.0.0.1:8915` → `bash deploy.sh`
+- -13: wheel 2개만 따로(`otel-proto-wheels.zip`, 400KB) → `cd /www/abap-ito/abap-specgen/web/BE/code && .venv/bin/pip install --no-index --find-links /www/otel-proto-wheels opentelemetry-proto`
+  → `bare/.env` 에 `PHOENIX_HOST=http://10.196.81.34:8915` → `run.sh restart`
 - 웹 UI: 고객사 PC 에서 `http://10.196.81.34:8915` (프로젝트 `codeassist` / `abap-specgen` 로 나뉨)

 ## 운영
@@ -10,3 +10,6 @@ psycopg2-binary==2.9.*
 pytest
 pytest-django
 pytest-asyncio
+
+# LLM 관측(Phoenix 는 OTLP protobuf 만 받음) — common/apps.gateway langfuse.py
+opentelemetry-proto
@@ -18,7 +18,15 @@ def lf_server(monkeypatch):
    got: list[dict] = []

    def handler(req: httpx.Request) -> httpx.Response:
-        got.append({"url": str(req.url), "auth": req.headers.get("authorization"), "body": json.loads(req.content)})
+        if req.headers.get("content-type", "").startswith("application/x-protobuf"):
+            from google.protobuf.json_format import MessageToDict
+            from opentelemetry.proto.collector.trace.v1.trace_service_pb2 import ExportTraceServiceRequest
+
+            m = ExportTraceServiceRequest(); m.ParseFromString(req.content)
+            body = MessageToDict(m)  # bytes 필드는 base64 로 나옴
+        else:
+            body = json.loads(req.content)
+        got.append({"url": str(req.url), "auth": req.headers.get("authorization"), "body": body, "ctype": req.headers.get("content-type", "")})
        return httpx.Response(200, json={})

    monkeypatch.setattr(langfuse, "TRANSPORT", httpx.MockTransport(handler))
@@ -115,7 +123,7 @@ async def test_phoenix_target_no_auth_and_openinference_attrs(lf_server):
    tool = langfuse.tool_span("t9", "read", parent=root, input={"filePath": "a.md"}, output="본문")
    assert await langfuse.send([root, tool]) and len(lf_server) == 1
    req = lf_server[0]
-    assert req["url"] == "http://px.test/v1/traces" and req["auth"] is None
+    assert req["url"] == "http://px.test/v1/traces" and req["auth"] is None and req["ctype"].startswith("application/x-protobuf")
    res = {a["key"]: a["value"]["stringValue"] for a in req["body"]["resourceSpans"][0]["resource"]["attributes"]}
    assert res["openinference.project.name"] == "codeassist"
    r, tl = _attrs(req["body"]["resourceSpans"][0]["scopeSpans"][0]["spans"][0]), _attrs(req["body"]["resourceSpans"][0]["scopeSpans"][0]["spans"][1])