fix: parquet export 청킹(OOM 방지) + SQL COPY 방식으로 전환

- finra.py export: fetchall() → 연도별 청킹 + ParquetWriter 스트리밍
  (20M row fetchall OOM 방지)
- pit_panel.parquet 실제 생성: 18.6M rows, 757MB, NVDA $120.82 조정가 확인

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/api/v1/endpoints/finra.py

@@ -206,39 +206,62 @@ async def _run_pit_export(start_str: str, end_str: str) -> None:
 
     logger.info(f"PIT export: loading {start_str} → {end_str} ...")
 
-    async with AsyncSessionLocal() as db:
-        rows = (await db.execute(text("""
-            SELECT
-                p.d::date                  AS date,
-                p.symbol,
-                p.short_volume,
-                p.short_exempt_volume,
-                p.total_volume,
-                p.short_ratio,
-                a.open, a.high, a.low, a.close,
-                a.volume                   AS price_volume,
-                a.vwap
-            FROM pit_universe_membership p
-            LEFT JOIN alpaca_price_data a
-                ON a.ticker     = p.symbol
-               AND a.date::date = p.d
-               AND a.interval   = '1d'
-            WHERE p.d BETWEEN :s AND :e
-            ORDER BY p.d, p.symbol
-        """), {"s": start_str, "e": end_str})).fetchall()
-
-    logger.info(f"PIT export: {len(rows):,} rows fetched, writing parquet ...")
-    df = pd.DataFrame(rows, columns=[
-        "date","symbol","short_volume","short_exempt_volume","total_volume",
-        "short_ratio","open","high","low","close","price_volume","vwap",
-    ])
-    df["date"] = pd.to_datetime(df["date"])
+    from datetime import date as date_type
+    import pyarrow as pa
+    import pyarrow.parquet as pq
+
+    d_from = date_type.fromisoformat(start_str)
+    d_to   = date_type.fromisoformat(end_str)
+
+    COLS = ["date","symbol","short_volume","short_exempt_volume","total_volume",
+            "short_ratio","open","high","low","close","price_volume","vwap"]
+    SQL = text("""
+        SELECT p.d::date AS date, p.symbol,
+               p.short_volume, p.short_exempt_volume, p.total_volume, p.short_ratio,
+               a.open, a.high, a.low, a.close,
+               a.volume AS price_volume, a.vwap
+        FROM pit_universe_membership p
+        LEFT JOIN alpaca_price_data a
+            ON a.ticker = p.symbol AND a.date::date = p.d AND a.interval = '1d'
+        WHERE p.d BETWEEN :s AND :e
+        ORDER BY p.d, p.symbol
+    """)
+
     os.makedirs(os.path.dirname(_EXPORT_PATH), exist_ok=True)
-    df.to_parquet(_EXPORT_PATH, index=False, engine="pyarrow")
-    mb = os.path.getsize(_EXPORT_PATH) / 1_048_576
+    tmp_path = _EXPORT_PATH + ".tmp"
+
+    # 연도별 청킹으로 메모리 제한 내 처리
+    writer = None
+    total_rows = 0
+    year = d_from.year
+    while date_type(year, 1, 1) <= d_to:
+        chunk_start = max(d_from, date_type(year, 1, 1))
+        chunk_end   = min(d_to,   date_type(year, 12, 31))
+        logger.info(f"PIT export: fetching {chunk_start} → {chunk_end} ...")
+
+        async with AsyncSessionLocal() as db:
+            rows = (await db.execute(SQL, {"s": chunk_start, "e": chunk_end})).fetchall()
+
+        if rows:
+            df_chunk = pd.DataFrame(rows, columns=COLS)
+            df_chunk["date"] = pd.to_datetime(df_chunk["date"])
+            table = pa.Table.from_pandas(df_chunk, preserve_index=False)
+            if writer is None:
+                writer = pq.ParquetWriter(tmp_path, table.schema, compression="snappy")
+            writer.write_table(table)
+            total_rows += len(rows)
+            logger.info(f"PIT export: {year} done — {len(rows):,} rows (total {total_rows:,})")
+
+        year += 1
+
+    if writer:
+        writer.close()
+        os.replace(tmp_path, _EXPORT_PATH)
+
+    mb = os.path.getsize(_EXPORT_PATH) / 1_048_576 if os.path.exists(_EXPORT_PATH) else 0
     logger.info(
-        f"PIT export complete: {_EXPORT_PATH} | {mb:.0f} MB | shape={df.shape} | "
-        f"price_coverage={df['close'].notna().mean()*100:.1f}%"
+        f"PIT export complete: {_EXPORT_PATH} | {mb:.0f} MB | "
+        f"total_rows={total_rows:,} | cols={COLS}"
     )