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:
+75 -6
View File
@@ -45,7 +45,7 @@ from app.config import (ConfigError, GatewayConfig, load_config, validate_warnin
PROTOCOL_CHAT, PROTOCOL_MESSAGES, PROTOCOL_RESPONSES, PROTOCOLS,
ALIAS_PROTOCOLS)
from app import calllog
from app.metrics import (Metrics, AccessLog, RateLimiter, BODY_BY_RID, _BODY_MAX,
from app.metrics import (Metrics, AccessLog, RateLimiter, BODY_BY_RID, RID, _BODY_MAX,
video_poll_push, video_polls,
ACCOUNT, ROLE)
from app.models import (ChatRequest, ImageRequest, VideoCreateRequestExtended, BatchRequest,
@@ -351,7 +351,7 @@ _docs_on = config.server.docs_enabled # 安全默认关(report: /docs、/ope
# 2026-09-30: 2.6.1 → 2.7.0(按老板要求整个删除「火山方舟 Coding Plan」供应商:配置/模型/代码/文档/前端下拉/环境变量全清)。
# 2026-09-30: 2.7.0 → 2.7.1(界面提示白话化:熔断器→供应商自动保护、引擎→运行环境、slug→标识;删 120 个未引用死键)。
# ★ 版本号纪律:每次改动都要往上加,小改动加小版本号、大改动加大版本号(README 有专章)。
VERSION = "2.8.0"
VERSION = "2.9.0"
app = FastAPI(
title="API Gateway",
@@ -606,6 +606,7 @@ def _flush_calllog(request: Request, rid: str, status: int, latency_ms: float,
async def add_request_id(request: Request, call_next):
rid = _sanitize_rid(request.headers.get("X-Request-ID")) or str(uuid.uuid4())
request.state.rid = rid
RID.set(rid) # 2026-10-23:token 记账靠这个 rid 关联到调用日志那条记录
t0 = time.perf_counter()
status = 500
is_stream = False
@@ -1441,6 +1442,11 @@ def _record_usage_metrics(usage) -> None:
p, c, cached, reasoning = _usage_detail(usage)
if p or c or cached or reasoning:
_metrics.record_tokens(p, c, cached=cached, reasoning=reasoning)
# 2026-10-23:同时告诉调用日志(按 rid 暂存,落库时带上 token 数)
try:
calllog.note_tokens(RID.get(), p, c, cached=cached, reasoning=reasoning)
except Exception: # noqa: BLE001
pass
class _UsageScanner:
@@ -1711,6 +1717,12 @@ async def _passthrough_stream(request, rid, caller, protocol, adapter, cb,
_metrics.record_tokens(counts["prompt_tokens"], counts["completion_tokens"],
cached=counts["cached_tokens"],
reasoning=counts["reasoning_tokens"])
try:
calllog.note_tokens(RID.get(), counts["prompt_tokens"], counts["completion_tokens"],
cached=counts["cached_tokens"],
reasoning=counts["reasoning_tokens"])
except Exception: # noqa: BLE001
pass
_access.add({"method": "POST", "kind": "chat", "path": f"/v1/{protocol}", "rid": rid,
"model": model_in, "served": model_in, "provider": pkey,
"caller": caller, "ok": ok, "stream": True, "response": ""})
@@ -2993,6 +3005,33 @@ async def media_download(request: Request, url: str = "", filename: str = "",
# ─────────────────────────────────────
# GET /metrics + /metrics/prometheus(P1-3/8)
# ─────────────────────────────────────
def _usage_snapshot(account: str | None = None) -> dict:
"""指标快照(2026-10-23 起口径变化)。
**调用数 / 供应商分布 / token / 耗时** 来自 SQLite(`api_calls`,保留 90 天)——
所以**重启不再清零**,Dashboard 看到的是历史累计;
**运行时长 / 在途请求** 仍是进程内实时值(DB 里没有意义)。
库不可用时自动回落到内存态,行为与之前一致。
"""
rt = _metrics.snapshot()
base: dict = {"uptime_seconds": rt.get("uptime_seconds"),
"in_flight": rt.get("in_flight", 0),
"in_flight_peak": rt.get("in_flight_peak", 0)}
db = calllog.store.usage_snapshot(account=account) if calllog.store.enabled else {"enabled": False}
if db.get("enabled"):
base.update({"calls": db.get("calls") or {},
"providers": db.get("providers") or {},
"tokens": db.get("tokens") or {},
"latency": db.get("latency") or {},
"total_calls": db.get("total", 0),
"source": "db"})
else:
mem = _metrics.snapshot(account)
base.update({k: mem.get(k) for k in ("calls", "providers", "tokens", "latency") if k in mem})
base["source"] = "memory"
return base
@app.get("/metrics")
async def metrics(request: Request, _auth=Depends(require_auth)):
if not _cfg().server.metrics_enabled:
@@ -3001,10 +3040,10 @@ async def metrics(request: Request, _auth=Depends(require_auth)):
# 2026-10-22(老板):成员首页要看自己调用的次数统计,所以**不**把本端点收成仅管理员;
# 按账号聚合的逻辑保持原样(全局字段的收敛在 /health 里做)。
if request.state.role == "admin":
snap = _metrics.snapshot()
snap = _usage_snapshot()
snap["circuits"] = _circuits.statuses() # P2-16 熔断状态
else:
snap = _metrics.snapshot(request.state.account)
snap = _usage_snapshot(request.state.account)
snap["circuits"] = []
return snap
@@ -3025,7 +3064,7 @@ async def metrics_prometheus(request: Request, _auth=Depends(require_auth)):
raise HTTPException(403, "仅管理员可抓取指标")
if not _cfg().server.metrics_enabled:
raise HTTPException(404, "metrics 已禁用")
snap = _metrics.snapshot()
snap = _usage_snapshot()
lines = [
"# API Gateway Prometheus 指标",
f'api_gateway_uptime_seconds {snap["uptime_seconds"]}',
@@ -3203,6 +3242,36 @@ async def logs_calls(request: Request, start: str | None = None, end: str | None
limit=limit, offset=offset, with_content=bool(content))
@app.get("/logs/usage/monthly")
async def logs_usage_monthly(request: Request, month: str | None = None,
_auth=Depends(require_auth)):
"""按月统计每个账号的用量(调用数/成功失败/平均耗时/覆盖模型数)。仅管理员。"""
_require_admin_logs(request)
return calllog.store.usage_monthly(month)
@app.get("/logs/usage/monthly/export")
async def logs_usage_monthly_export(request: Request, month: str | None = None,
_auth=Depends(require_auth)):
"""月度用量 CSV 导出(按账号一行)。仅管理员。"""
_require_admin_logs(request)
data = calllog.store.usage_monthly(month)
if not data.get("enabled", True) and data.get("error"):
raise HTTPException(400, data["error"])
import csv as _csv
import io as _io
buf = _io.StringIO()
w = _csv.writer(buf)
w.writerow(["月份", "账号", "调用数", "成功", "失败", "失败率", "平均耗时ms", "涉及模型数"])
for it in data.get("items", []):
rate = f'{it["fail"] / it["calls"] * 100:.1f}%' if it.get("calls") else "-"
w.writerow([data.get("month"), it.get("account"), it.get("calls"), it.get("ok"),
it.get("fail"), rate, it.get("avg_ms"), it.get("models")])
fname = f"usage-{data.get('month')}.csv"
return Response(content=buf.getvalue().encode("utf-8-sig"), media_type="text/csv",
headers={"Content-Disposition": f'attachment; filename="{fname}"'})
@app.get("/logs/calls/{call_id}")
async def logs_call_detail(request: Request, call_id: int, _auth=Depends(require_auth)):
_require_admin_logs(request)
@@ -3930,7 +3999,7 @@ async def health(request: Request, _auth=Depends(require_auth_or_open)):
if not getattr(request.state, "health_authed", False):
return {"status": "ok"}
cfg = _cfg()
snap = _metrics.snapshot() if cfg.server.metrics_enabled else {}
snap = _usage_snapshot() if cfg.server.metrics_enabled else {}
# 2026-10-22(老板):成员首页需要"运行时长 / 并发 / 供应商列表"这些运行态,
# 但**全局** token 总量、各模型的全局调用量与失败率、供应商熔断状态属运维数据,
# 只应收口给管理员(此前任何有效 Bearer 都能读到全站用量)。