From abfab0fa5738974a41227432cd054ae037299116 Mon Sep 17 00:00:00 2001
From: I Luk Kim <yirugi@gmail.com>
Date: Tue, 14 Apr 2026 17:02:23 -0700
Subject: [PATCH] =?UTF-8?q?fix:=20OOM=20kill=20=EB=B0=A9=EC=A7=80=20?=
 =?UTF-8?q?=E2=80=94=20Phase=204=20=EA=B2=BD=EB=9F=89=ED=99=94=20+=20Semap?=
 =?UTF-8?q?hore=203=EC=9C=BC=EB=A1=9C=20=EC=A0=9C=ED=95=9C?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Worker가 OOM kill로 죽는 것을 확인 (docker inspect: OOMKilled=true).
호스트 메모리: 물리 RAM ~68MB 여유, 스왑 97% 사용 상태.

원인: Phase 4에서 SELECT AlpacaPriceData (전체 ORM 객체) 로드.
- 75 ticker × 78 5분봉 = 5,850개 ORM 객체/요청
- SQLAlchemy ORM instrumentation으로 dict 대비 ~10x 메모리 소비
- Semaphore(10) → 최대 10개 요청 동시 처리 = 피크 수백 MB ~ 1GB

수정:
- Phase 4: SELECT * → SELECT 7 columns only (ticker, date, OHLCV)
  → result.all()로 경량 Row namedtuple 반환, ORM 객체 생성 없음
  → SQLAlchemy Row는 속성 접근(row.date, row.open 등) 지원 — 엔드포인트 호환
- Semaphore(10) → Semaphore(3): 동시 처리 3개로 제한
  → 피크 메모리 ~70% 감소

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 app/api/v1/endpoints/alpaca.py       |  2 +-
 app/services/alpaca_price_service.py | 20 +++++++++++++++-----
 2 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/app/api/v1/endpoints/alpaca.py b/app/api/v1/endpoints/alpaca.py
index 72a9536..439494a 100644
--- a/app/api/v1/endpoints/alpaca.py
+++ b/app/api/v1/endpoints/alpaca.py
@@ -15,7 +15,7 @@ _MARKET_CLOSE_HOUR = 16  # 4:00 PM ET
 # Limit concurrent Alpaca intraday processing to prevent event-loop saturation
 # under bulk backfill workloads. Callers beyond this limit wait on the semaphore
 # (cheap asyncio wait) rather than flooding httpx connections and DB sessions.
-_INTRADAY_SEMAPHORE = asyncio.Semaphore(10)
+_INTRADAY_SEMAPHORE = asyncio.Semaphore(3)
 
 
 def _market_closed_for(d: date) -> bool:
diff --git a/app/services/alpaca_price_service.py b/app/services/alpaca_price_service.py
index f187a69..07757a5 100644
--- a/app/services/alpaca_price_service.py
+++ b/app/services/alpaca_price_service.py
@@ -191,10 +191,20 @@ class AlpacaPriceService:
                     f"Alpaca multi-bars: stored {len(rows)} rows for {len(need_fetch)} tickers"
                 )
 
-        # Phase 4: read back (short session) — filtered by interval to avoid mixing 1d/5m/etc.
+        # Phase 4: read back (short session, lightweight column query).
+        # Select only the 7 columns the endpoints need instead of loading full ORM
+        # objects — avoids SQLAlchemy instrumentation overhead (~10x lighter per row).
         async with AsyncSessionLocal() as db:
             result = await db.execute(
-                select(AlpacaPriceData)
+                select(
+                    AlpacaPriceData.ticker,
+                    AlpacaPriceData.date,
+                    AlpacaPriceData.open,
+                    AlpacaPriceData.high,
+                    AlpacaPriceData.low,
+                    AlpacaPriceData.close,
+                    AlpacaPriceData.volume,
+                )
                 .where(
                     and_(
                         AlpacaPriceData.ticker.in_(upper_tickers),
@@ -205,10 +215,10 @@ class AlpacaPriceService:
                 )
                 .order_by(AlpacaPriceData.ticker, AlpacaPriceData.date)
             )
-            db_rows = result.scalars().all()
-            await asyncio.sleep(0)  # yield after bulk ORM object creation
+            db_rows = result.all()  # lightweight Row namedtuples, not ORM objects
+            await asyncio.sleep(0)  # yield after bulk row creation
 
-        data: Dict[str, List[AlpacaPriceData]] = {t: [] for t in upper_tickers}
+        data: Dict[str, List] = {t: [] for t in upper_tickers}
         for row in db_rows:
             data[row.ticker].append(row)
         return data