提交

backend/app/backtest/events.py

@@ -11,6 +11,7 @@ match=all(连续满足)/any(曾经满足)，多条件之间取 AND。
 """
 from __future__ import annotations
 
+import asyncio
 from datetime import date, datetime, timedelta
 
 import numpy as np
@@ -120,6 +121,69 @@ def _stats_block(trades: list[dict]) -> dict:
     }
 
 
+def _scan_batch(
+    candle_rows: list,
+    code_by_symbol: dict[str, str],
+    f_map: dict,
+    name_map: dict[str, str],
+    spec: EventBacktestSpec,
+    start_ts: datetime,
+    trades_limit: int,
+) -> list[dict]:
+    """纯 CPU：一批 bar -> 交易明细（放线程池跑，不阻塞事件循环）。"""
+    trades: list[dict] = []
+    if not candle_rows:
+        return trades
+    bars = pd.DataFrame(
+        candle_rows, columns=["symbol", "ts", "open", "high", "low", "close"]
+    )
+    for symbol, g in bars.groupby("symbol", sort=False):
+        if len(g) < 30:
+            continue
+        g = g.reset_index(drop=True)
+        ts_code_l = code_by_symbol[symbol]
+        cache: dict = {"_families": set()}
+        mask = _signal_mask(g, spec, cache)
+        if not mask.any():
+            continue
+        for sig_i in np.flatnonzero(mask.to_numpy()):
+            ts_sig = g.at[sig_i, "ts"]
+            # 信号必须落在回测窗口内（buffer 区只用于指标配热）
+            if ts_sig < start_ts:
+                continue
+            ie = _entry_exit_indices(int(sig_i), spec, len(g))
+            if ie is None:
+                continue
+            entry_i, exit_i = ie
+            e_row, x_row = g.iloc[entry_i], g.iloc[exit_i]
+            e_price = _price_at(e_row, "open" if spec.entry_timing == "next_open" else "close")
+            x_price = _price_at(x_row, "open" if spec.exit_timing == "open" else "close")
+            if not e_price or not x_price:
+                continue
+            f_in = f_map.get((ts_code_l, e_row["ts"].date()), 1.0)
+            f_out = f_map.get((ts_code_l, x_row["ts"].date()), 1.0)
+            ret_pct = (x_price * f_out) / (e_price * f_in) * 100 - 100
+            trades.append({
+                "ts_code": ts_code_l,
+                "name": name_map.get(ts_code_l),
+                "entry_date": e_row["ts"], "entry_price": round(e_price, 3),
+                "exit_date": x_row["ts"], "exit_price": round(x_price, 3),
+                "ret_pct": round(float(ret_pct), 3),
+            })
+            if len(trades) >= trades_limit:
+                return trades
+    return trades
+
+
+def _summarize(trades: list[dict]) -> tuple[dict, list[dict]]:
+    """纯 CPU：汇总统计 + 最好/最差样本（同样下线程池）。"""
+    stats = _stats_block(trades)
+    # 明细样本：最好 100 + 最差 100（其余统计已覆盖）
+    trades_sorted = sorted(trades, key=lambda t: t["ret_pct"], reverse=True)
+    sample = trades_sorted[:100] + (trades_sorted[-100:] if len(trades_sorted) > 100 else [])
+    return stats, sample
+
+
 async def run_event_backtest(
     session: AsyncSession,
     spec: EventBacktestSpec,
@@ -195,53 +259,15 @@ async def run_event_backtest(
         )).all()
         f_map = {(r[0], r[1].date()): float(r[2]) for r in adj_rows if r[2]}
 
-        bars = pd.DataFrame(
-            candle_rows, columns=["symbol", "ts", "open", "high", "low", "close"]
-        )
-        for symbol, g in bars.groupby("symbol", sort=False):
-            if len(g) < 30:
-                continue
-            g = g.reset_index(drop=True)
-            ts_code_l = code_by_symbol[symbol]
-            cache: dict = {"_families": set()}
-            mask = _signal_mask(g, spec, cache)
-            if not mask.any():
-                continue
-            for sig_i in np.flatnonzero(mask.to_numpy()):
-                ts_sig = g.at[sig_i, "ts"]
-                # 信号必须落在回测窗口内（buffer 区只用于指标配热）
-                if ts_sig < start_ts:
-                    continue
-                ie = _entry_exit_indices(int(sig_i), spec, len(g))
-                if ie is None:
-                    continue
-                entry_i, exit_i = ie
-                e_row, x_row = g.iloc[entry_i], g.iloc[exit_i]
-                e_price = _price_at(e_row, "open" if spec.entry_timing == "next_open" else "close")
-                x_price = _price_at(x_row, "open" if spec.exit_timing == "open" else "close")
-                if not e_price or not x_price:
-                    continue
-                f_in = f_map.get((ts_code_l, e_row["ts"].date()), 1.0)
-                f_out = f_map.get((ts_code_l, x_row["ts"].date()), 1.0)
-                ret_pct = (x_price * f_out) / (e_price * f_in) * 100 - 100
-                trades.append({
-                    "ts_code": ts_code_l,
-                    "name": name_map.get(ts_code_l),
-                    "entry_date": e_row["ts"], "entry_price": round(e_price, 3),
-                    "exit_date": x_row["ts"], "exit_price": round(x_price, 3),
-                    "ret_pct": round(float(ret_pct), 3),
-                })
-                if len(trades) >= MAX_TRADES:
-                    break
-            if len(trades) >= MAX_TRADES:
-                break
+        # pandas 全市场扫描是同步 CPU 重计算，丢线程池跑（await 期间事件循环可服务其他请求）
+        trades.extend(await asyncio.to_thread(
+            _scan_batch, candle_rows, code_by_symbol, f_map, name_map,
+            spec, start_ts, MAX_TRADES - len(trades),
+        ))
         if len(trades) >= MAX_TRADES:
             break
 
-    stats = _stats_block(trades)
-    # 明细样本：最好 100 + 最差 100（其余统计已覆盖）
-    trades_sorted = sorted(trades, key=lambda t: t["ret_pct"], reverse=True)
-    sample = trades_sorted[:100] + (trades_sorted[-100:] if len(trades_sorted) > 100 else [])
+    stats, sample = await asyncio.to_thread(_summarize, trades)
     return {
         "spec": spec,
         "universe": ts_code or "all",