refactor: MCP 공통 함수/httpx 싱글톤, CORS 수정, ML 이벤트루프 블로킹 해소 (#4~#7)

backend/app/services/ml_ensemble.py

@@ -9,6 +9,7 @@ ML 앙상블 모델 서비스
 앙상블 스코어: 0~100 (높을수록 매수 신호 강함)
 """
 
+import asyncio
 import logging
 import os
 from pathlib import Path
@@ -275,7 +276,6 @@ class MLEnsemble:
             metrics["lgbm_accuracy"] = float(lgbm_acc)
             logger.info(f"LightGBM 검증 정확도: {lgbm_acc:.4f}")
 
-            # 피처 중요도
             importance = self.lgbm_model.feature_importances_
             top_features = sorted(
                 zip(FEATURE_COLUMNS, importance),
@@ -302,7 +302,6 @@ class MLEnsemble:
                 X_seq_train, y_train,
                 epochs=50, lr=0.001, batch_size=32,
             )
-            # 검증
             lstm_val_probs = self.lstm_model.predict_proba(X_seq_val)
             lstm_val_pred = np.argmax(lstm_val_probs, axis=1)
             lstm_acc = np.mean(lstm_val_pred == y_val)
@@ -342,7 +341,6 @@ class MLEnsemble:
             metrics["ensemble_accuracy"] = float(ensemble_acc)
             logger.info(f"앙상블 검증 정확도: {ensemble_acc:.4f}")
 
-        # 모델 저장
         self._save_models()
         self.is_trained = True
         self.training_metrics = metrics
@@ -355,7 +353,6 @@ class MLEnsemble:
         probs_list = []
         weights_list = []
 
-        # LightGBM
         if self.lgbm_model is not None:
             try:
                 lgbm_probs = self.lgbm_model.predict_proba(X_tabular)
@@ -364,7 +361,6 @@ class MLEnsemble:
             except Exception:
                 pass
 
-        # LSTM
         if self.lstm_model is not None and self.lstm_model.model is not None:
             try:
                 lstm_probs = self.lstm_model.predict_proba(X_seq)
@@ -373,7 +369,6 @@ class MLEnsemble:
             except Exception:
                 pass
 
-        # RF
         if self.rf_model is not None:
             try:
                 rf_probs = self.rf_model.predict_proba(X_tabular)
@@ -385,7 +380,6 @@ class MLEnsemble:
         if not probs_list:
             return None
 
-        # 가중 평균
         total_weight = sum(weights_list)
         weighted_probs = sum(
             p * (w / total_weight) for p, w in zip(probs_list, weights_list)
@@ -393,19 +387,15 @@ class MLEnsemble:
         return weighted_probs
 
     def predict(self, chart_data: list) -> Dict[str, Any]:
-        """단일 종목 예측.
+        """단일 종목 예측 (동기).
+
+        이벤트 루프 내 async 핸들러에서는 predict_async() 를 사용하세요.
 
         Args:
             chart_data: KIS API 차트 데이터 (최신순)
 
         Returns:
-            {
-                "ensemble_score": int (0~100),
-                "confidence": float (0~1),
-                "signal": str ("BUY"/"HOLD"/"SELL"),
-                "model_scores": {"lgbm": ..., "lstm": ..., "rf": ...},
-                "probabilities": {"up": ..., "hold": ..., "down": ...}
-            }
+            {ensemble_score, confidence, signal, model_scores, probabilities}
         """
         if not self.is_trained:
             return {"error": "모델이 학습되지 않았습니다", "ensemble_score": 50, "confidence": 0.0, "signal": "HOLD"}
@@ -414,25 +404,18 @@ class MLEnsemble:
         if tabular is None or sequence is None:
             return {"error": "피처 추출 실패 (데이터 부족)", "ensemble_score": 50, "confidence": 0.0, "signal": "HOLD"}
 
-        # 스케일링
         tabular_scaled = self.scaler.transform(tabular.reshape(1, -1))
         sequence_batch = sequence.reshape(1, 60, len(LSTM_SEQUENCE_FEATURES))
 
-        # 앙상블 예측
         probs = self._ensemble_predict_proba(tabular_scaled, sequence_batch)
         if probs is None:
             return {"error": "모든 모델 예측 실패", "ensemble_score": 50, "confidence": 0.0, "signal": "HOLD"}
 
         prob_down, prob_hold, prob_up = probs[0]
-
-        # 앙상블 스코어: UP 확률 기반 (0~100)
         ensemble_score = int(prob_up * 100)
-
-        # 신뢰도: 최고 확률과 나머지의 차이
         max_prob = max(prob_down, prob_hold, prob_up)
         confidence = float(max_prob)
 
-        # 신호 결정
         if prob_up > prob_down and prob_up > prob_hold and prob_up > 0.4:
             signal = "BUY"
         elif prob_down > prob_up and prob_down > prob_hold and prob_down > 0.4:
@@ -440,7 +423,6 @@ class MLEnsemble:
         else:
             signal = "HOLD"
 
-        # 개별 모델 스코어
         model_scores = {}
         if self.lgbm_model is not None:
             try:
@@ -473,6 +455,22 @@ class MLEnsemble:
             },
         }
 
+    # ─── 비동기 래퍼 (이슈 7: 이벤트 루프 블로킹 방지) ──────────────────
+    async def predict_async(self, chart_data: list) -> Dict[str, Any]:
+        """asyncio.to_thread로 CPU 추론을 워커 스레드에서 실행.
+
+        async 핸들러/스케줄러에서는 이 메서드를 사용하세요.
+        무거운 RandomForest 배치 학습에는 ProcessPoolExecutor도 고려 가능.
+        """
+        return await asyncio.to_thread(self.predict, chart_data)
+
+    async def train_async(
+        self, X_tabular: np.ndarray, X_seq: np.ndarray, y: np.ndarray
+    ) -> Dict[str, Any]:
+        """asyncio.to_thread로 CPU 학습을 워커 스레드에서 실행."""
+        return await asyncio.to_thread(self.train, X_tabular, X_seq, y)
+    # ─────────────────────────────────────────────────────────────────────
+
     def _save_models(self):
         """모델 파일 저장"""
         self._ensure_dir()
@@ -486,7 +484,6 @@ class MLEnsemble:
         if self.scaler is not None:
             joblib.dump(self.scaler, self.model_dir / "scaler.pkl")
 
-        # 메트릭 저장
         import json
         with open(self.model_dir / "metrics.json", "w") as f:
             json.dump(self.training_metrics, f, indent=2, ensure_ascii=False, default=str)
@@ -519,7 +516,6 @@ class MLEnsemble:
                 )
                 self.lstm_model.load(lstm_path)
 
-            # 메트릭 로드
             metrics_path = self.model_dir / "metrics.json"
             if metrics_path.exists():
                 import json