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>

5_django_backend/apps/gateway/langfuse.py

@@ -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