feat(metrics): 指标改由 SQLite 提供(重启不清零)+ 月度用量导出

- 新增 _usage_snapshot():调用数/供应商/token/耗时(p50,p95) 走 api_calls 聚合,
  运行时长/在途/熔断仍取内存;库不可用自动回落内存态
- 影响 /metrics、/metrics/prometheus、/health.models;口径从「启动以来」变「最近 90 天」
- token 落库:表加 4 列(老库自动 ALTER),用 rid contextvar + token 侧信道关联
- 新增 /logs/usage/monthly 与 /logs/usage/monthly/export(仅管理员)+ 前端月度用量弹窗
- 模型统计排除无模型行(匿名早失败不再显示 ?)
- 版本 2.8.0 → 2.9.0
This commit is contained in:
jidekin committed 2026-10-23 11:28:53 +08:00
1 parent bb4c9281f7
commit 46f10f4594
6 files changed
+320 -7

No files matched your search

+138 -1
View File
@@ -69,6 +69,10 @@ CREATE TABLE IF NOT EXISTS api_calls (
ua TEXT,
prompt TEXT,
completion TEXT,
tok_prompt INTEGER DEFAULT 0,
tok_completion INTEGER DEFAULT 0,
tok_cached INTEGER DEFAULT 0,
tok_reasoning INTEGER DEFAULT 0,
ver TEXT
);
CREATE INDEX IF NOT EXISTS idx_calls_ts ON api_calls(ts);
@@ -81,7 +85,8 @@ CREATE INDEX IF NOT EXISTS idx_calls_rid ON api_calls(rid);
_COLUMNS = ("ts", "rid", "account", "key_id", "kind", "method", "path", "model",
"served", "provider", "caller", "stream", "ok", "status", "latency_ms",
"ip", "ua", "prompt", "completion", "ver")
"ip", "ua", "prompt", "completion",
"tok_prompt", "tok_completion", "tok_cached", "tok_reasoning", "ver")
_INSERT = f"INSERT INTO api_calls ({', '.join(_COLUMNS)}) VALUES ({', '.join('?' * len(_COLUMNS))})"
@@ -124,6 +129,7 @@ class CallLogStore:
db.execute("PRAGMA journal_mode=WAL")
db.execute("PRAGMA synchronous=NORMAL")
db.executescript(SCHEMA)
self._migrate(db)
db.commit()
try:
os.chmod(self.path, 0o600)
@@ -133,6 +139,15 @@ class CallLogStore:
log.info("调用日志已就绪: %s(保留 %d 天)", self.path, self.retention_days)
self.cleanup()
@staticmethod
def _migrate(db: sqlite3.Connection) -> None:
"""轻量迁移:老库缺 token 列时补上(2026-10-23 加)。"""
have = {r[1] for r in db.execute("PRAGMA table_info(api_calls)")}
for col in ("tok_prompt", "tok_completion", "tok_cached", "tok_reasoning"):
if col not in have:
db.execute(f"ALTER TABLE api_calls ADD COLUMN {col} INTEGER DEFAULT 0")
log.info("调用日志表已补列: %s", col)
async def start(self) -> None:
self.open()
if not self.enabled or self._task is not None:
@@ -353,6 +368,94 @@ class CallLogStore:
out["queued"] = len(self._q)
return out
# ── 聚合(供 /metrics、/health、月度用量导出使用)────────────────────
def usage_snapshot(self, *, account: str | None = None, days: int = 0) -> dict[str, Any]:
"""按模型/供应商/账号聚合调用数、耗时与 token(形状对齐内存态 `Metrics.snapshot()`)。
account 非空 → 只算该账号(成员视角);days=0 → 不限时间(用库内全部,即保留期内)。
"""
if not self.enabled or self._db is None:
return {"enabled": False}
since = time.time() - days * 86400 if days and days > 0 else 0
conds: list[str] = []
params: list[Any] = []
if since:
conds.append("ts >= ?")
params.append(since)
if account:
conds.append("account = ?")
params.append(account)
where = (" WHERE " + " AND ".join(conds)) if conds else ""
try:
with self._lock:
mcond = f"{where} AND model IS NOT NULL" if where else " WHERE model IS NOT NULL"
calls = {r[0]: {"ok": r[1] or 0, "fail": r[2] or 0}
for r in self._db.execute(
f"SELECT model, SUM(CASE WHEN ok=1 THEN 1 ELSE 0 END), "
f"SUM(CASE WHEN ok=0 THEN 1 ELSE 0 END) "
f"FROM api_calls{mcond} GROUP BY model", params)}
pcond = f"{where} AND provider IS NOT NULL" if where else " WHERE provider IS NOT NULL"
providers = {r[0] or "?": {"ok": r[1] or 0, "fail": r[2] or 0}
for r in self._db.execute(
f"SELECT provider, SUM(CASE WHEN ok=1 THEN 1 ELSE 0 END), "
f"SUM(CASE WHEN ok=0 THEN 1 ELSE 0 END) "
f"FROM api_calls{pcond} GROUP BY provider", params)}
latency: dict[str, Any] = {}
lcond = (f"{where} AND " if where else " WHERE ") + "model IS NOT NULL AND latency_ms IS NOT NULL"
for r in self._db.execute(
f"SELECT model, COUNT(*), AVG(latency_ms) FROM api_calls{lcond} GROUP BY model",
params):
latency[r[0]] = {"count": r[1], "avg": round(r[2] or 0, 3)}
# p50 / p95:表很小,用 offset 取分位值
for mid, d in list(latency.items()):
for key, frac in (("p50", 0.5), ("p95", 0.95)):
off = max(0, int((d["count"] - 1) * frac))
c2 = (f"{where} AND " if where else " WHERE ") + "model IS ? AND latency_ms IS NOT NULL"
row = self._db.execute(
f"SELECT latency_ms FROM api_calls{c2} ORDER BY latency_ms LIMIT 1 OFFSET ?",
[*params, mid, off]).fetchone()
d[key] = row[0] if row else None
tok = self._db.execute(
f"SELECT COALESCE(SUM(tok_prompt),0), COALESCE(SUM(tok_completion),0), "
f"COALESCE(SUM(tok_cached),0), COALESCE(SUM(tok_reasoning),0), COUNT(*) "
f"FROM api_calls{where}", params).fetchone()
except Exception as e: # noqa: BLE001
log.warning("用量聚合失败: %s", e)
return {"enabled": False, "error": str(e)}
return {"enabled": True, "calls": calls, "providers": providers, "latency": latency,
"tokens": {"prompt": tok[0], "completion": tok[1],
"cached": tok[2], "reasoning": tok[3]},
"total": tok[4], "since": since}
def usage_monthly(self, month: str | None = None) -> dict[str, Any]:
"""按月统计每个账号的用量(调用数 / 成功失败 / 平均耗时 / 内容字节 / token 不提供)。"""
if not self.enabled or self._db is None:
return {"enabled": False, "month": month, "items": []}
if not month:
month = time.strftime("%Y-%m")
try:
start = time.mktime(time.strptime(month + "-01", "%Y-%m-%d"))
except ValueError:
return {"enabled": False, "month": month, "error": "月份格式应为 YYYY-MM", "items": []}
# 下月 1 号
y, m = int(month[:4]), int(month[5:7])
ny, nm = (y + 1, 1) if m == 12 else (y, m + 1)
end = time.mktime(time.strptime(f"{ny:04d}-{nm:02d}-01", "%Y-%m-%d"))
try:
with self._lock:
rows = self._db.execute(
"SELECT account, COUNT(*), SUM(CASE WHEN ok=1 THEN 1 ELSE 0 END), "
"SUM(CASE WHEN ok=0 THEN 1 ELSE 0 END), AVG(latency_ms), "
"COUNT(DISTINCT model) FROM api_calls WHERE ts >= ? AND ts < ? "
"GROUP BY account ORDER BY COUNT(*) DESC", (start, end)).fetchall()
except Exception as e: # noqa: BLE001
return {"enabled": False, "month": month, "error": str(e), "items": []}
items = [{"account": r[0] or "anonymous", "calls": r[1] or 0, "ok": r[2] or 0,
"fail": r[3] or 0, "avg_ms": round(r[4], 1) if r[4] else 0,
"models": r[5] or 0} for r in rows]
return {"enabled": True, "month": month, "items": items,
"total_calls": sum(i["calls"] for i in items)}
def export_rows(self, **kw: Any) -> list[dict[str, Any]]:
"""导出用:不带分页,最多 50000 条。"""
kw.setdefault("limit", 50000)
@@ -387,6 +490,12 @@ def _trim(d: dict[str, dict]) -> None:
def _row(entry: dict[str, Any], meta: dict[str, Any]) -> dict[str, Any]:
rid = str(entry.get("rid") or meta.get("rid") or "")
tok = entry.get("tokens")
if not tok and rid:
with _tok_lock:
tok = TOKENS_BY_RID.pop(rid, None)
tok = tok or {}
return {
"ts": entry.get("ts") or time.time(),
"rid": entry.get("rid") or meta.get("rid"),
@@ -406,11 +515,39 @@ def _row(entry: dict[str, Any], meta: dict[str, Any]) -> dict[str, Any]:
"ip": meta.get("ip", ""),
"ua": _clip(meta.get("ua", ""), 200),
"prompt": _clip(entry.get("body"), CONTENT_MAX),
"tok_prompt": int(tok.get("prompt_tokens") or 0),
"tok_completion": int(tok.get("completion_tokens") or 0),
"tok_cached": int(tok.get("cached_tokens") or 0),
"tok_reasoning": int(tok.get("reasoning_tokens") or 0),
"completion": _clip(entry.get("response"), CONTENT_MAX),
"ver": meta.get("ver", ""),
}
# token 侧信道:`_record_usage_metrics()` 只知道 token、不知道 rid;请求中间件只知道 rid、
# 不知道 token。用一个按 rid 索引的小字典把两半接起来(AccessLog.add 落库时取走并清空)。
TOKENS_BY_RID: dict[str, dict[str, int]] = {}
_tok_lock = threading.Lock()
_TOK_MAX = 2000
def note_tokens(rid: str, prompt: int, completion: int,
cached: int = 0, reasoning: int = 0) -> None:
"""记下某次请求的 token 用量(由 `_record_usage_metrics` 调用)。"""
if not store.enabled or not rid:
return
with _tok_lock:
TOKENS_BY_RID[rid] = {"prompt_tokens": int(prompt or 0),
"completion_tokens": int(completion or 0),
"cached_tokens": int(cached or 0),
"reasoning_tokens": int(reasoning or 0)}
while len(TOKENS_BY_RID) > _TOK_MAX:
try:
TOKENS_BY_RID.pop(next(iter(TOKENS_BY_RID)), None)
except (StopIteration, KeyError):
break
def on_access(entry: dict[str, Any]) -> None:
"""`AccessLog.add()` 调用:拿到访问日志条目;若响应元数据已就绪则立即落库。"""
if not store.enabled: