mirror of
https://gitcode.com/JianFeeeee/HomeAgent.git
synced 2026-09-28 05:13:27 +00:00
test(webui): WebAPI 只读端点压测 + 实测 2160 请求零失败
打的是**生产实例** 127.0.0.1:8080。 ## 为什么只压只读端点 压测绝不能改状态。18 个端点**逐个探测确认**是 GET + 只读: 小(/status…/proxy/services,104B–1.8KB)、中(/plugins…/knowledge, 4.4KB–53KB)、大(/kernel 157KB、/memory/graph 338KB、/knowledge/tree 425KB)。 **排除**:/chat(真调 LLM)、/chat/interrupt(中断在跑的任务)、 /settings/* 与 /plugins/*(改配置)、/login /logout(改会话)、 knowledge/memory 写接口、/device/*(控制真实设备)。 ## ★ 两条判据(都不是"看有没有报错") ### 1. 必须先 `--probe`:webui 无认证时返回 200 + 登录页 HTML 含 `THEME_PLACEHOLDER`,**状态码是 200**。只看 `http_code` 会把登录页 当成健康响应 ⇒「全部 200」是假的。脚本因此额外校验响应体。 ### 2. 端点表漏了 `/api/v1` 前缀 ⇒ 18 个端点全 404 第一版把端点存成路径后半段(`"/status"`),拼出 `http://127.0.0.1:8080/status`,而真实路径是 `/api/v1/status` ⇒ **全部 404**,同一时刻 curl `/api/v1/status` 却是 200。 ⇒ **压测脚本必须先 probe 再压。** 不 probe 的话那一跑的结论会是 「webui 全挂」,完全错误。这条已写进手册。 ## 实测(2026-09-28) | scale | 请求 | 吞吐 | p50 | p95 | p99 | max | 成功 | | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | 360 | 779 req/s | 6.7ms | 26.8ms | 45.5ms | 51.6ms | 360/360 | | 2 | 720 | 788 req/s | 12.2ms | 52.6ms | 163.0ms | 213.5ms | 720/720 | | 3 | 1080 | 1038 req/s | 16.4ms | 60.4ms | 79.7ms | 123.0ms | 1080/1080 | **2160 请求零失败**;压测期间服务端 `active`、0 个 5xx、0 个 webui 错误, QQ/agent 链路未受影响,homed CPU 仅 1.2%。 ### 重响应不是瓶颈(并发 1 → 16) | 端点 | 大小 | 并发 1 | 并发 16 | 吞吐 | | --- | --- | --- | --- | --- | | /status | 0.2KB | 104 req/s | **1781 req/s** | 320KB/s | | /kernel | 153KB | 126 req/s | 375 req/s | 19 → **57 MB/s** | | /memory/graph | 330KB | 70 req/s | 410 req/s | 23 → **135 MB/s** | | /knowledge/tree | 415KB | 79 req/s | 717 req/s | 33 → **297 MB/s** | 大 JSON 端点的 p50 **不随并发上升**(/knowledge/tree 12.3ms → 8.9ms, 排队更充分、效率更高)⇒ 瓶颈不在 JSON 序列化。 ## ⚠ 一处未下结论的观察 scale=2 的 p99(163ms)反而**高于** scale=3(79.7ms)。看着反常,但当时 机器上另一个 agent 的 chromium 占 **766% CPU**(另有 cjpm/cjc 在编译) ⇒ 是**环境噪声**。 **未在可比条件下重测就不下结论** —— 不拿这一组当性能特征。要判定需先 固定负载条件再跑。这条也写进手册。 ## 顺带 ruff 报的三条 blocker 都修了:`pct()` 空样本不再抛错(调用方直接拿去做 f-string 格式化)、写 JSON 失败明确报错而非静默丢报告、错误体读取用 `contextlib.suppress` 免得二次异常盖掉真正的状态码。
This commit is contained in:
@ -488,3 +488,77 @@ CGO_ENABLED=0 GOOS=linux GOARCH=amd64 go build \
|
||||
另注:`aaafaac`/`94c74b2`(修复设备反复掉线/静默失联:ping 路径断连 +
|
||||
bind 结果无人处理)正是 106 此前长期无 `online` 日志的成因,
|
||||
2026-09-26 已进 main,今天部署的 `1.4.0` 包含它。
|
||||
|
||||
---
|
||||
|
||||
## 7. webui WebAPI 压测
|
||||
|
||||
脚本:`scripts/kernel-stress/webui-bench.py`,打的是**生产实例**
|
||||
`127.0.0.1:8080`。
|
||||
|
||||
```bash
|
||||
K=$(sqlite3 /home/newqqagent/config.db "select value from config_webui where key='api_key';")
|
||||
python3 scripts/kernel-stress/webui-bench.py --key "$K" --probe # 先探测
|
||||
python3 scripts/kernel-stress/webui-bench.py --key "$K" --scale 3 --json /tmp/w.json
|
||||
```
|
||||
|
||||
### ★ 必须先 `--probe`
|
||||
|
||||
webui 无认证时返回 **200 + 登录页 HTML**(含 `THEME_PLACEHOLDER`)。
|
||||
**状态码是 200**,只看 `http_code` 会把登录页当成健康响应 ⇒
|
||||
「全部 200」是假的。脚本因此额外校验响应体(`looks_like_login_page()`)。
|
||||
|
||||
### ★ 只压只读端点
|
||||
|
||||
压测绝不能改状态。18 个端点**逐个探测确认**是 GET + 只读:
|
||||
|
||||
| 规模 | 端点 | 备注 |
|
||||
| --- | --- | --- |
|
||||
| 小 | `/status` `/network` `/terminals` `/tracker` `/adapters` `/agents` `/config` `/persona` `/proxy/services` | 104B–1.8KB |
|
||||
| 中 | `/plugins` `/runtime` `/memory/context` `/memory/text` `/memory` `/knowledge` | 4.4KB–53KB |
|
||||
| 大 | `/kernel` `/memory/graph` `/knowledge/tree` | 157KB / 338KB / **425KB** |
|
||||
|
||||
**排除**:`/chat`(真调 LLM)、`/chat/interrupt`(中断在跑的任务)、
|
||||
`/settings/*` `/plugins/*`(改配置)、`/login` `/logout`(改会话)、
|
||||
`/knowledge/*` `/memory/*` 写接口、`/device/*`(控制真实设备)。
|
||||
|
||||
### 实测(2026-09-28)
|
||||
|
||||
| scale | 请求 | 吞吐 | p50 | p95 | p99 | max | 成功 |
|
||||
| --- | --- | --- | --- | --- | --- | --- | --- |
|
||||
| 1 | 360 | 779 req/s | 6.7ms | 26.8ms | 45.5ms | 51.6ms | **360/360** |
|
||||
| 2 | 720 | 788 req/s | 12.2ms | 52.6ms | 163.0ms | 213.5ms | **720/720** |
|
||||
| 3 | 1080 | 1038 req/s | 16.4ms | 60.4ms | 79.7ms | 123.0ms | **1080/1080** |
|
||||
|
||||
**2160 请求零失败**,压测期间服务端 `active`、0 个 5xx、0 个 webui 错误,
|
||||
QQ/agent 链路未受影响。
|
||||
|
||||
#### 重响应不是瓶颈(并发 1 → 16)
|
||||
|
||||
| 端点 | 大小 | 并发 1 | 并发 16 | 吞吐 |
|
||||
| --- | --- | --- | --- | --- |
|
||||
| `/status` | 0.2KB | 104 req/s | **1781 req/s** | 320KB/s |
|
||||
| `/kernel` | 153KB | 126 req/s | 375 req/s | 19 → **57 MB/s** |
|
||||
| `/memory/graph` | 330KB | 70 req/s | 410 req/s | 23 → **135 MB/s** |
|
||||
| `/knowledge/tree` | 415KB | 79 req/s | 717 req/s | 33 → **297 MB/s** |
|
||||
|
||||
⇒ 大 JSON 端点的 p50 **不随并发上升**(`/knowledge/tree` 12.3ms → 8.9ms,
|
||||
即排队更充分、效率更高),瓶颈不在 JSON 序列化。
|
||||
|
||||
### ⚠ 一次未下结论的观察
|
||||
|
||||
scale=2 的 p99(163ms)反而**高于** scale=3(79.7ms)。看着像反常,但当时
|
||||
机器上另一个 agent 的 chromium 占 **766% CPU**(另有 cjpm/cjc 在编译)
|
||||
⇒ 是**环境噪声**,不是 webui 特性。
|
||||
|
||||
**未在可比条件下重测就不下结论** —— 别拿这一组数据当性能特征。
|
||||
要判定需先固定负载条件(停掉占 CPU 的进程)再跑。
|
||||
|
||||
### 踩过的坑
|
||||
|
||||
脚本第一版把端点表存成路径后半段(`"/status"`)而漏了 `/api/v1` 前缀 ⇒
|
||||
拼出 `http://127.0.0.1:8080/status` ⇒ **18 个端点全 404**,而同一时刻
|
||||
`curl /api/v1/status` 是 200。
|
||||
|
||||
⇒ **压测脚本必须先 probe 再压。** 若不 probe,那一跑会得出
|
||||
「webui 全挂」的错误结论。
|
||||
|
||||
256
scripts/kernel-stress/webui-bench.py
Normal file
256
scripts/kernel-stress/webui-bench.py
Normal file
@ -0,0 +1,256 @@
|
||||
#!/usr/bin/env python3
|
||||
"""webui WebAPI 只读端点压测。
|
||||
|
||||
## 为什么只压只读端点
|
||||
|
||||
这是**生产实例**(192.168.2.60:8080)。压测绝不能改状态:
|
||||
`/chat` 会真调 LLM、`/chat/interrupt` 会中断正在跑的任务、
|
||||
`/settings/*` 与 `/plugins/*` 会改配置、`/device/*` 会控制真实设备。
|
||||
|
||||
⇒ 端点表**按实测逐个确认过**是 GET + 只读,不是照路由表抄的。
|
||||
探测方式见下方 `--probe`。
|
||||
|
||||
## 判据(与 kernel-stress 那套一致)
|
||||
|
||||
不能只看"没报错"或墙钟时间。每一项都记:
|
||||
- 成功率(HTTP 200 占比)
|
||||
- 延迟分布(p50 / p95 / p99 / max)
|
||||
- 失败状态码与响应体样本
|
||||
|
||||
**失败必须带样本**,否则「200 但返回空 JSON」这类退化看不见。
|
||||
|
||||
## 认证
|
||||
|
||||
webui 认两种:`X-API-Key` 头,或 `homeagent_session` cookie。
|
||||
无认证时返回 **200 + 登录页 HTML**(含 `THEME_PLACEHOLDER`)——
|
||||
**状态码是 200 但内容是登录页**,只统计状态码会得出「全部健康」的错误结论。
|
||||
本脚本因此额外校验响应体,见 `looks_like_login_page()`。
|
||||
|
||||
用法:
|
||||
python3 webui-bench.py --key <API_KEY> [--scale 1|2|3] [--json out.json]
|
||||
python3 webui-bench.py --key <API_KEY> --probe # 只探测端点可用性
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import concurrent.futures as cf
|
||||
import contextlib
|
||||
import json
|
||||
import statistics
|
||||
import sys
|
||||
import time
|
||||
import urllib.error
|
||||
import urllib.request
|
||||
|
||||
BASE = "http://127.0.0.1:8080"
|
||||
# ★ 必须带 /api/v1 前缀。端点表里存的是**路径后半段**
|
||||
# ("/status"),拼 BASE + "/status" 会得到 "/status" ⇒ 全部 404。
|
||||
# 首次跑 --probe 时正是这样:18 个端点全 404,而同一时刻 curl
|
||||
# /api/v1/status 是 200。⇒ 压测脚本**必须先 probe** 再压。
|
||||
API_PREFIX = "/api/v1"
|
||||
|
||||
# 实测确认的只读端点(GET,200,不改状态)
|
||||
READONLY = [
|
||||
("/status", "小", "首页 KPI 之一"),
|
||||
("/network", "小", "网络信息"),
|
||||
("/terminals", "小", "终端列表"),
|
||||
("/tracker", "小", "跟踪器"),
|
||||
("/adapters", "小", "LLM 适配器"),
|
||||
("/agents", "小", "agent 列表"),
|
||||
("/config", "小", "配置"),
|
||||
("/persona", "小", "人格"),
|
||||
("/proxy/services", "小", "反代服务"),
|
||||
("/plugins", "中", "插件清单"),
|
||||
("/runtime", "中", "运行时"),
|
||||
("/memory/context", "中", "记忆上下文"),
|
||||
("/memory/text", "中", "文本记忆"),
|
||||
("/memory", "中", "记忆"),
|
||||
("/knowledge", "中", "知识库"),
|
||||
("/kernel", "大", "★ 总览 8 个 KPI 的数据源"),
|
||||
("/memory/graph", "大", "记忆图谱"),
|
||||
("/knowledge/tree", "大", "★ 知识树(实测最大 425KB)"),
|
||||
]
|
||||
|
||||
|
||||
def looks_like_login_page(body: bytes) -> bool:
|
||||
"""webui 无认证时返回 200 + 登录页 HTML。
|
||||
|
||||
状态码骗人:只看 http_code 会把「登录页」当成健康响应。
|
||||
"""
|
||||
head = body[:4096]
|
||||
return (
|
||||
b"THEME_PLACEHOLDER" in head
|
||||
or "统一门户登录".encode() in head
|
||||
or (b"<title>" in head and b"login" in head)
|
||||
)
|
||||
|
||||
|
||||
def one_request(path: str, key: str, timeout: float) -> dict:
|
||||
url = BASE + API_PREFIX + path
|
||||
req = urllib.request.Request(url, headers={"X-API-Key": key})
|
||||
t0 = time.perf_counter()
|
||||
try:
|
||||
with urllib.request.urlopen(req, timeout=timeout) as r:
|
||||
body = r.read()
|
||||
dt = (time.perf_counter() - t0) * 1000
|
||||
return {
|
||||
"ok": r.status == 200 and not looks_like_login_page(body),
|
||||
"status": r.status,
|
||||
"bytes": len(body),
|
||||
"ms": dt,
|
||||
"login_page": looks_like_login_page(body),
|
||||
}
|
||||
except urllib.error.HTTPError as e:
|
||||
dt = (time.perf_counter() - t0) * 1000
|
||||
# ★ 读取错误体可能二次抛错(连接已断、body 为空),不能让它
|
||||
# 盖掉真正的状态码 —— 用 suppress 只吞这一处的异常。
|
||||
sample = ""
|
||||
with contextlib.suppress(Exception):
|
||||
sample = e.read()[:200].decode("utf-8", "replace")
|
||||
return {"ok": False, "status": e.code, "bytes": 0, "ms": dt, "err": sample}
|
||||
except Exception as e:
|
||||
dt = (time.perf_counter() - t0) * 1000
|
||||
return {"ok": False, "status": 0, "bytes": 0, "ms": dt, "err": repr(e)[:200]}
|
||||
|
||||
|
||||
def pct(vals, p):
|
||||
"""百分位。空列表 / 非法入参都返回 None 而不是抛错。
|
||||
|
||||
调用方会把它直接拿去做 f-string 格式化(可能收到 None),
|
||||
所以**这里不能抛** —— 一个端点的空样本不该让整份报告崩掉。
|
||||
"""
|
||||
if not vals:
|
||||
return None
|
||||
try:
|
||||
s = sorted(float(v) for v in vals)
|
||||
k = min(len(s) - 1, max(0, int(round((float(p) / 100) * (len(s) - 1)))))
|
||||
except (TypeError, ValueError):
|
||||
return None
|
||||
return s[k]
|
||||
|
||||
|
||||
def probe(key: str) -> int:
|
||||
"""逐个确认端点真的可 GET、返回 JSON 而非登录页。"""
|
||||
print("端点探测(GET / 只读)")
|
||||
bad = 0
|
||||
for path, size, note in READONLY:
|
||||
r = one_request(path, key, 15)
|
||||
flag = "✓" if r["ok"] else "✗"
|
||||
extra = ""
|
||||
if r.get("login_page"):
|
||||
extra = " ← 返回登录页(认证有问题)"
|
||||
bad += 1
|
||||
elif r["status"] != 200:
|
||||
extra = f" ← {r.get('err', '')[:80]}"
|
||||
bad += 1
|
||||
print(f" {flag} {path:24} {r['status']:>3} {r['bytes']:>7}B {r['ms']:>7.1f}ms [{size}] {note}{extra}")
|
||||
return 1 if bad else 0
|
||||
|
||||
|
||||
def bench(key: str, scale: int, rounds: int, conc: int, timeout: float, out: str | None):
|
||||
jobs = []
|
||||
for _ in range(rounds):
|
||||
for path, _, _ in READONLY:
|
||||
jobs.append(path)
|
||||
print(f"\n压测:{len(READONLY)} 端点 × {rounds} 轮 = {len(jobs)} 请求,"
|
||||
f"并发 {conc}(scale={scale})")
|
||||
print(" 只读端点,不触碰 /chat、/settings、/plugins 写接口、/device\n")
|
||||
|
||||
t0 = time.perf_counter()
|
||||
results = []
|
||||
with cf.ThreadPoolExecutor(max_workers=conc) as ex:
|
||||
futs = {ex.submit(one_request, p, key, timeout): p for p in jobs}
|
||||
for f in cf.as_completed(futs):
|
||||
results.append((futs[f], f.result()))
|
||||
wall = time.perf_counter() - t0
|
||||
|
||||
by_path = {}
|
||||
for path, r in results:
|
||||
by_path.setdefault(path, []).append(r)
|
||||
|
||||
print(f"{'端点':24} {'成功':>10} {'p50':>8} {'p95':>8} {'p99':>8} {'max':>8} {'平均KB':>8}")
|
||||
print("-" * 78)
|
||||
all_ms = []
|
||||
total_ok = total = 0
|
||||
failures = []
|
||||
for path, _, _ in READONLY:
|
||||
rs = by_path.get(path, [])
|
||||
if not rs:
|
||||
continue
|
||||
oks = [r for r in rs if r["ok"]]
|
||||
ms = [r["ms"] for r in oks] or [r["ms"] for r in rs]
|
||||
all_ms += ms
|
||||
total_ok += len(oks)
|
||||
total += len(rs)
|
||||
for r in rs:
|
||||
if not r["ok"]:
|
||||
failures.append((path, r))
|
||||
avg_kb = sum(r["bytes"] for r in rs) / len(rs) / 1024
|
||||
print(f"{path:24} {len(oks):>4}/{len(rs):<5} "
|
||||
f"{pct(ms,50):>8.1f} {pct(ms,95):>8.1f} {pct(ms,99):>8.1f} {max(ms):>8.1f} {avg_kb:>8.1f}")
|
||||
|
||||
print("-" * 78)
|
||||
rps = total / wall if wall else 0
|
||||
print(f"总计 {total} 请求 / {wall:.2f}s ⇒ {rps:.0f} req/s "
|
||||
f"成功 {total_ok}/{total} = {total_ok/total*100:.2f}%")
|
||||
print(f"全局 p50={pct(all_ms,50):.1f}ms p95={pct(all_ms,95):.1f}ms "
|
||||
f"p99={pct(all_ms,99):.1f}ms max={max(all_ms):.1f}ms")
|
||||
if all_ms:
|
||||
print(f"平均 {statistics.mean(all_ms):.1f}ms "
|
||||
f"stdev {statistics.pstdev(all_ms):.1f}ms "
|
||||
f"min {min(all_ms):.1f}ms")
|
||||
|
||||
if failures:
|
||||
print(f"\n★ 失败 {len(failures)} 次(带样本,便于定位):")
|
||||
seen = set()
|
||||
for path, r in failures[:10]:
|
||||
key2 = (path, r["status"])
|
||||
if key2 in seen:
|
||||
continue
|
||||
seen.add(key2)
|
||||
print(f" {path:24} status={r['status']:>3} "
|
||||
f"{r.get('err', 'login_page' if r.get('login_page') else '')[:90]}")
|
||||
|
||||
if out:
|
||||
try:
|
||||
with open(out, "w", encoding="utf-8") as f:
|
||||
json.dump({
|
||||
"wall_s": wall, "rps": rps, "total": total, "ok": total_ok,
|
||||
"p50": pct(all_ms, 50), "p95": pct(all_ms, 95), "p99": pct(all_ms, 99),
|
||||
"max": max(all_ms) if all_ms else None,
|
||||
"per_path": {p: {"n": len(v),
|
||||
"ok": sum(1 for r in v if r["ok"]),
|
||||
"p50": pct([r["ms"] for r in v], 50),
|
||||
"p95": pct([r["ms"] for r in v], 95),
|
||||
"max": max((r["ms"] for r in v), default=None),
|
||||
"avg_bytes": sum(r["bytes"] for r in v) / len(v)}
|
||||
for p, v in by_path.items()},
|
||||
}, f, ensure_ascii=False, indent=2)
|
||||
except OSError as e:
|
||||
# ★ 写不出去就**说清楚**,不能默默丢掉整份报告。
|
||||
print(f"\n★ 写入 {out} 失败: {e}", file=sys.stderr)
|
||||
return 1
|
||||
print(f"\n已写 {out}")
|
||||
|
||||
return 0 if total_ok == total else 1
|
||||
|
||||
|
||||
def main():
|
||||
ap = argparse.ArgumentParser()
|
||||
ap.add_argument("--key", required=True, help="webui API key")
|
||||
ap.add_argument("--scale", type=int, default=1)
|
||||
ap.add_argument("--conc", type=int, default=0, help="0=按 scale 自动")
|
||||
ap.add_argument("--timeout", type=float, default=30.0)
|
||||
ap.add_argument("--json", default=None)
|
||||
ap.add_argument("--probe", action="store_true")
|
||||
a = ap.parse_args()
|
||||
|
||||
if a.probe:
|
||||
sys.exit(probe(a.key))
|
||||
|
||||
rounds = 20 * a.scale
|
||||
conc = a.conc or min(64, 8 * a.scale)
|
||||
sys.exit(bench(a.key, a.scale, rounds, conc, a.timeout, a.json))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user