perf(data): min/max改逐对双seek——INDEXED BY批IN形态VPS实测仍不达标(30只冷8.1s/热0.25s,大池冷态分钟级),ORDER BY ±datetime LIMIT 1三列全等前缀每股恰2次寻位冷热皆毫秒级——上一发(80fc517)生产探针续测定形态:①原IN+GROUP BY劣化计划101s冷;②钉复合索引后计划翻转(TEMP B-TREE消失)但仍逐股扫全部interval条目(d+15m+5m),30只冷8.1s→momentum池~2000只外推冷态≈9min过不了<60s验收;③行值IN变体在本SQLite对(row-value IN+INDEXED BY)走全索引扫描7min不归(弃);终态=逐对双seek,symbol/exchange/interval三列全等直落d子区间两端,连15m/5m条目都不碰;无日线对不落minmax(下游.get默认(None,None)同旧GROUP BY语义);测试5个重写到seek API(双索引名/回退/EXPLAIN计划签名interval=?直落+无TEMP B-TREE/batch==逐只等价);628绿;待VPS探针终验双seek计时 [vps]
CI/CD / test (push) Successful in 9s
CI/CD / nas-deploy (push) Successful in 25s
CI/CD / nas-verify (push) Successful in 1s

This commit is contained in:
2026-08-24 14:14:11 +08:00
parent 96a1e6b413
commit aafa5baa67
2 changed files with 65 additions and 44 deletions
@@ -640,15 +640,25 @@ class LocalUnifiedProvider(DataProvider): # type: ignore[misc]
self._cached_minmax_index = name
return name
def _minmax_sql(self, conn: sqlite3.Connection, n_symbols: int) -> str:
"""min/max 批量 SQL; 复合索引可用时 INDEXED BY 钉死(防计划劣化)。"""
ph = ",".join("?" * n_symbols)
def _minmax_seek_sql(self, conn: sqlite3.Connection) -> tuple:
"""min/max 双 seek SQL(ORDER BY ±datetime LIMIT 1); 复合索引可用时钉死。
2026-08-24 计划劣化治本(VPS 生产实测选形态): ①原 IN+GROUP BY 批量 SQL 被
选去单列 interval 索引整扫 ~21M 日线行(30 只冷 101s, 选股 38-51min 主根因);
②INDEXED BY 钉复合索引后仍要逐股扫该股**全部 interval** 条目(30 只冷
8.1s/热 0.25s, 大池冷态分钟级); ③双 seek = 三列全等前缀直落
(symbol,exchange,'d') 子区间两端, 每股恰 2 次索引寻位, 冷热皆毫秒级。
索引名跨环境不同(ensure_schema 长名/NAS 副本 uq_dbbardata)→运行时探测,
探不到回退裸 SQL(三列全等下 planner 本就倾向复合索引)。
"""
idx = self._composite_minmax_index(conn)
hint = f" INDEXED BY {idx}" if idx else ""
return (
f"SELECT symbol, exchange, MIN(datetime), MAX(datetime) FROM dbbardata{hint} "
f"WHERE interval='d' AND symbol IN ({ph}) GROUP BY symbol, exchange"
base = (
f"SELECT datetime FROM dbbardata{hint} "
f"WHERE symbol=? AND exchange=? AND interval='d' "
f"ORDER BY datetime"
)
return base + " LIMIT 1", base + " DESC LIMIT 1"
def get_security_info_batch(
self,
@@ -670,17 +680,21 @@ class LocalUnifiedProvider(DataProvider): # type: ignore[misc]
syms = list({sym for sym, _ in (p[1] for p in pairs)})
conn = self._connect()
# 1. dbbardata min/max per (symbol, exchange); chunk 防 >999 参数;
# INDEXED BY 钉复合索引(2026-08-24 计划劣化治本, 见 _minmax_sql)
# 1. dbbardata min/max per (symbol, exchange): 逐对双 seek(ORDER BY
# ±datetime LIMIT 1, 见 _minmax_seek_sql); 无日线对不落 minmax,
# 下游 .get((sym,exc),(None,None)) 同旧 GROUP BY 语义
minmax: Dict[tuple, tuple] = {}
for i in range(0, len(syms), 400):
chunk = syms[i:i + 400]
sql = self._minmax_sql(conn, len(chunk))
for r in conn.execute(sql, chunk):
minmax[(r[0], r[1])] = (
r[2][:10] if r[2] else None,
r[3][:10] if r[3] else None,
)
sym_exc = list(dict.fromkeys(pe for _, pe in pairs)) # 去重保序
sql_lo, sql_hi = self._minmax_seek_sql(conn)
for sym, exc in sym_exc:
lo = conn.execute(sql_lo, (sym, exc)).fetchone()
if lo is None or not lo[0]:
continue
hi = conn.execute(sql_hi, (sym, exc)).fetchone()
minmax[(sym, exc)] = (
lo[0][:10],
hi[0][:10] if hi and hi[0] else None,
)
# 2. constituent_unified code_name(碰撞股名; 缺则回退 jq_code)
names: Dict[str, str] = {}
+35 -28
View File
@@ -603,11 +603,12 @@ class TestGetSecurityInfoBatch:
class TestSecurityInfoBatchIndexHint:
"""2026-08-24 P0 计划劣化治本: VPS 生产 EXPLAIN 实锤 min/max 批量 SQL 被选去
单列 interval 索引, 整扫 ~21M 日线行 + USE TEMP B-TREE(30 只冷 101s,
开盘选股 38-51min 主根因) = INDEXED BY 钉死 (symbol,exchange,interval,
datetime) 复合唯一索引; 索引名跨环境不同(ensure_schema 长名 / NAS 副本
merge_increment uq_dbbardata)运行时探测+实例缓存, 探不到回退裸查询"""
"""2026-08-24 P0 计划劣化治本(VPS 生产实测定形态): ①原 IN+GROUP BY 批量 SQL
被选去单列 interval 索引整扫 ~21M 日线行(30 只冷 101s, 选股 38-51min 主根因);
INDEXED BY 钉复合索引后仍逐股扫全部 interval 条目(30 只冷 8.1s/ 0.25s);
终态 = 逐对双 seek(ORDER BY ±datetime LIMIT 1), 三列全等前缀直落 d 子区间
两端, 每股恰 2 次索引寻位复合索引名跨环境不同(ensure_schema 长名 / NAS
副本 merge_increment uq_dbbardata)运行时探测+实例缓存, 探不到回退裸 SQL"""
@staticmethod
def _db_with_index(tmp_path, index_name):
@@ -621,34 +622,38 @@ class TestSecurityInfoBatchIndexHint:
c.close()
return db
def test_minmax_sql_pins_composite_index(self, tmp_path):
def test_seek_sql_pins_composite_index(self, tmp_path):
db = self._db_with_index(tmp_path, "dbbardata_symbol_exchange_interval_datetime")
p = LocalUnifiedProvider({"db_path": str(db), "data_dir": str(tmp_path)})
conn = p._connect()
sql = p._minmax_sql(conn, 3)
assert "INDEXED BY dbbardata_symbol_exchange_interval_datetime" in sql
sql_lo, sql_hi = p._minmax_seek_sql(conn)
assert "INDEXED BY dbbardata_symbol_exchange_interval_datetime" in sql_lo
assert "INDEXED BY dbbardata_symbol_exchange_interval_datetime" in sql_hi
assert sql_lo.endswith("ORDER BY datetime LIMIT 1")
assert sql_hi.endswith("ORDER BY datetime DESC LIMIT 1")
# 探测结果实例缓存(None 与未探测用哨兵区分, 见 _composite_minmax_index)
assert p._composite_minmax_index(conn) == "dbbardata_symbol_exchange_interval_datetime"
assert p._composite_minmax_index(conn) == \
"dbbardata_symbol_exchange_interval_datetime"
def test_minmax_sql_detects_nas_replica_index_name(self, tmp_path):
def test_seek_sql_detects_nas_replica_index_name(self, tmp_path):
"""NAS 副本复合索引叫 uq_dbbardata(硬编码索引名在此环境必炸)。"""
db = self._db_with_index(tmp_path, "uq_dbbardata")
p = LocalUnifiedProvider({"db_path": str(db), "data_dir": str(tmp_path)})
sql = p._minmax_sql(p._connect(), 3)
assert "INDEXED BY uq_dbbardata" in sql
sql_lo, _ = p._minmax_seek_sql(p._connect())
assert "INDEXED BY uq_dbbardata" in sql_lo
def test_minmax_sql_fallback_without_composite_index(self, tmp_path):
def test_seek_sql_fallback_without_composite_index(self, tmp_path):
db = _make_security_info_fixture(tmp_path)
p = LocalUnifiedProvider({"db_path": str(db), "data_dir": str(tmp_path)})
conn = p._connect()
sql = p._minmax_sql(conn, 3)
assert "INDEXED BY" not in sql # 探不到回退裸查询(不劣于现状)
rows = list(conn.execute(sql, ("600519", "000001", "000002")))
assert rows # 裸查询仍可执行
sql_lo, sql_hi = p._minmax_seek_sql(conn)
assert "INDEXED BY" not in sql_lo # 探不到回退裸 SQL(三列全等仍高效)
assert conn.execute(sql_lo, ("600519", "SSE")).fetchone()[0][:10] == "2024-06-18"
assert conn.execute(sql_hi, ("600519", "SSE")).fetchone()[0][:10] == "2024-06-20"
def test_pinned_plan_avoids_temp_btree(self, tmp_path):
"""钉死复合索引后 GROUP BY (symbol,exchange) 走索引序, 无 TEMP B-TREE
(劣化计划的特征签名); 单列 interval 索引同库共存也不受其扰"""
def test_seek_plan_hits_d_subrange_directly(self, tmp_path):
"""三列全等前缀 + 索引序 ORDER BY: 无 TEMP B-TREE、直落复合索引 d 子区间;
单列 interval 索引同库共存也不受其扰"""
db = self._db_with_index(tmp_path, "dbbardata_symbol_exchange_interval_datetime")
c = sqlite3.connect(str(db))
c.execute("CREATE INDEX idx_dbbardata_interval ON dbbardata (interval)")
@@ -656,16 +661,18 @@ class TestSecurityInfoBatchIndexHint:
c.close()
p = LocalUnifiedProvider({"db_path": str(db), "data_dir": str(tmp_path)})
conn = p._connect()
sql = p._minmax_sql(conn, 3)
plan = "\n".join(
r[3] for r in conn.execute(
"EXPLAIN QUERY PLAN " + sql, ("600519", "000001", "000002"))
)
assert "TEMP B-TREE" not in plan
assert "dbbardata_symbol_exchange_interval_datetime" in plan
sql_lo, sql_hi = p._minmax_seek_sql(conn)
for sql in (sql_lo, sql_hi):
plan = "\n".join(
r[3] for r in conn.execute(
"EXPLAIN QUERY PLAN " + sql, ("600519", "SSE"))
)
assert "TEMP B-TREE" not in plan
assert "dbbardata_symbol_exchange_interval_datetime" in plan
assert "interval=?" in plan # 三列全等(d 子区间直落, 不扫 15m/5m)
def test_batch_equivalence_with_index_hint(self, tmp_path):
"""带复合索引(=生产形态)下 batch 与逐只结果一致(hint 不改语义)。"""
"""带复合索引(=生产形态)下 batch 与逐只结果一致(seek 不改语义)。"""
db = self._db_with_index(tmp_path, "uq_dbbardata")
p = LocalUnifiedProvider({"db_path": str(db), "data_dir": str(tmp_path)})
codes = ["600519.XSHG", "000001.XSHE", "000002.XSHE", "999999.XSHG"]