chore(ops): reload 卡死三层预防 + F1-5a 完整走查报告

reload 卡死三层预防(走查中遭遇 uvicorn graceful shutdown 死等触发):
- Layer 1 (apps/backend/start_uvicorn.py 新): 把 reload-excludes
  封装在 Python 字符串内,ps1 命令行只有字面路径,根治 PowerShell
  PSNativeCommandArgumentPassing 在不同 profile 下 wildcard 展开
  行为差异(数组 splatting 和 --% 都不稳)。同时显式设
  timeout-graceful-shutdown=5,5 秒强杀防死等
- Layer 2 (scripts/ops/backend-watchdog.ps1 新): 自主 socket 探针
  (TcpClient + 手写 HTTP/1.1 GET,Connection: close)规避 .NET
  HttpClient pool 复用 + 系统代理误报;3s × 3 = 9s 触发重启;
  进程链 kill 至 pwsh 后端窗口(关闭原窗口);3 次/小时上限自停
- Layer 3 (scripts/ops/start-admin.ps1): 启动时拉起 watchdog,
  菜单 [4] 仅重启后端选项,主菜单退出时一并 kill 看门狗

CLAUDE.md: 新增"后端 reload 卡死预防(强制)"章节,
分级文件风险表 + SOP + 启动菜单速查

走查报告(应查尽查严肃版):
- 后端 6 个改造点 PASS(P1-P4 + GUC + ai_run_logs runtime 字段)
- admin-web 7 页 Playwright 实地走查 → 5 项 UI 不完整登记 F1-5b
- 小程序看板 tab 7 页 weixin-devtools-mcp 实地 + DB 数据核对 →
  board-finance 5/6 项上界裁剪吻合;board-customer 业务日生效;
  board-coach 月度聚合表设计盲区;5 项 sandbox 覆盖盲区登记 F1-5b
- 8 张走查截图归档 docs/audit/changes/screenshots/2026-05-05_f1_5a_walkthrough/

audit_dashboard 刷新到 153 条审计

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Neo
2026-05-05 11:53:08 +08:00
parent 1baa21222b
commit 95a4500c75
14 changed files with 745 additions and 5 deletions

View File

@@ -0,0 +1,258 @@
# -*- coding: utf-8 -*-
# 后端 FastAPI 看门狗:探针 /health,卡死自动重启
#
# 触发条件:
# - 连续 $FailThreshold 次 /health 探针失败(默认 3 次,30 秒)
#
# 行为:
# - kill 占用 8000 端口的所有进程(uvicorn worker + reloader 父进程)
# - 等待端口释放
# - 启动新 uvicorn 进程(同 start-admin.ps1 的命令格式)
# - 启动后 $StartupGrace 秒(默认 45)不再探针,等 lifespan 完成
#
# 由 start-admin.ps1 在主启动后调用,主菜单退出时一并 kill 看门狗。
#
# CHANGE 2026-05-05 | F1-5a 走查发现 reload 卡死问题:
# 修改 services/* main 等核心文件 + admin-web 浏览器开着 WS,
# uvicorn graceful shutdown 等不到 WebSocket / asyncio task / DB 连接释放,
# 死等在 "Waiting for background tasks to complete"。
# 看门狗作为 Layer 2 保护:reload 卡死 → 探针失败 → 强杀 → 重启。
param(
[string]$ProjectRoot = $env:NEOZQYY_ROOT,
[int]$CheckInterval = 5, # 探针间隔(秒)
[int]$FailThreshold = 12, # 连续失败次数阈值(达到则重启)3s × 3 = 9s 触发
[int]$StartupGrace = 60, # 启动后宽限期(秒,期间不探针)
[int]$MaxRestartsPerHour = 3, # 每小时最大重启次数(超出则停止看门狗,避免误报死循环)
[int]$Port = 8000,
[string]$HealthPath = "/health"
)
$ErrorActionPreference = "Continue"
# ── 解析项目根 ─────────────────────────────────────────
if (-not $ProjectRoot -or -not (Test-Path $ProjectRoot)) {
if ($env:NEOZQYY_LAUNCH_DIR -and (Test-Path $env:NEOZQYY_LAUNCH_DIR)) {
$ProjectRoot = $env:NEOZQYY_LAUNCH_DIR.TrimEnd('\')
} elseif ($PSScriptRoot) {
$ProjectRoot = Split-Path -Parent (Split-Path -Parent $PSScriptRoot)
}
}
if (-not $ProjectRoot -or -not (Test-Path $ProjectRoot)) {
Write-Host "[watchdog] 无法定位 ProjectRoot,退出" -ForegroundColor Red
exit 1
}
$venvPython = Join-Path $ProjectRoot ".venv\Scripts\python.exe"
$backendDir = Join-Path $ProjectRoot "apps\backend"
if (-not (Test-Path $venvPython)) {
Write-Host "[watchdog] .venv Python 不存在: $venvPython" -ForegroundColor Red
exit 1
}
$healthUrl = "http://127.0.0.1:$Port$HealthPath"
$psExe = if (Get-Command pwsh -ErrorAction SilentlyContinue) { "pwsh" } else { "powershell" }
# PS 5.1 + 传统控制台 UTF-8
if ($PSVersionTable.PSVersion.Major -lt 7 -and -not $env:WT_SESSION) {
chcp 65001 | Out-Null
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
}
function Write-Log {
param([string]$Msg, [string]$Color = "Gray")
$ts = Get-Date -Format "HH:mm:ss"
Write-Host "[watchdog $ts] $Msg" -ForegroundColor $Color
}
function Test-BackendAlive {
# CHANGE 2026-05-05 v3 | 自主管理 socket,完全规避 .NET HttpClient pool 复用 + 系统代理:
# - TcpClient 短连接(每次新建,不池化)
# - 手写 HTTP/1.1 GET request,Connection: close 强制无复用
# - send/receive 各 3 秒超时,稳定且能识别 worker 卡死(端口在但不响应 read)
$tcp = $null
try {
$tcp = [System.Net.Sockets.TcpClient]::new()
$tcp.SendTimeout = 1500
$tcp.ReceiveTimeout = 3000
$iar = $tcp.BeginConnect("127.0.0.1", $Port, $null, $null)
$ok = $iar.AsyncWaitHandle.WaitOne(1500, $false)
if (-not $ok -or -not $tcp.Connected) { return $false }
$stream = $tcp.GetStream()
$stream.ReadTimeout = 3000
$stream.WriteTimeout = 1500
$req = "GET $HealthPath HTTP/1.1`r`nHost: 127.0.0.1:$Port`r`nConnection: close`r`nUser-Agent: NeoZQYY-Watchdog/1.0`r`n`r`n"
$bytes = [System.Text.Encoding]::ASCII.GetBytes($req)
$stream.Write($bytes, 0, $bytes.Length)
$buf = New-Object byte[] 256
$n = $stream.Read($buf, 0, 256)
if ($n -lt 12) { return $false }
$statusLine = [System.Text.Encoding]::ASCII.GetString($buf, 0, [Math]::Min($n, 80))
return $statusLine -match "HTTP/1\.. 200"
} catch {
return $false
} finally {
if ($tcp) { try { $tcp.Close() } catch { } }
}
}
function Get-BackendPids {
# 找 8000 端口监听者 + 沿进程树往上找,直到 pwsh/powershell 后端窗口
# 链:uvicorn worker(python) → uvicorn reloader(python) → pwsh(后端窗口)
# 关闭原窗口需要把 pwsh 也加入 kill 列表
$pids = @()
$listeners = Get-NetTCPConnection -LocalPort $Port -State Listen -ErrorAction SilentlyContinue
foreach ($l in $listeners) {
$cur = $l.OwningProcess
$pids += $cur
for ($i = 0; $i -lt 5; $i++) {
$proc = Get-CimInstance Win32_Process -Filter "ProcessId=$cur" -ErrorAction SilentlyContinue
if (-not $proc -or -not $proc.ParentProcessId) { break }
$parent = Get-CimInstance Win32_Process -Filter "ProcessId=$($proc.ParentProcessId)" -ErrorAction SilentlyContinue
if (-not $parent) { break }
# 链上是 python(reloader) 或 pwsh/powershell(后端窗口) 都加入
if ($parent.Name -in @("python.exe", "pwsh.exe", "powershell.exe")) {
$pids += $parent.ProcessId
$cur = $parent.ProcessId
# 到了 pwsh/powershell 就停(再往上是 watchdog/explorer 不该杀)
if ($parent.Name -in @("pwsh.exe", "powershell.exe")) { break }
} else {
break
}
}
}
return $pids | Sort-Object -Unique
}
function Stop-StuckBackend {
Write-Log "强杀 8000 端口相关进程..." Yellow
$targetPids = Get-BackendPids
if (-not $targetPids -or $targetPids.Count -eq 0) {
Write-Log "未发现 8000 端口监听进程,可能已退出" DarkGray
return $true
}
foreach ($id in $targetPids) {
Write-Log " taskkill /T /F PID=$id"
taskkill /PID $id /T /F 2>$null | Out-Null
}
# 等端口释放
$waited = 0
while ($waited -lt 20) {
Start-Sleep -Seconds 1
$waited++
$still = Get-NetTCPConnection -LocalPort $Port -State Listen -ErrorAction SilentlyContinue
if (-not $still) {
Write-Log " 端口 $Port 已释放(等待 ${waited}s)" Green
return $true
}
}
Write-Log " 端口 $Port 仍占用,放弃自动重启" Red
return $false
}
function Start-NewBackend {
Write-Log "启动新 uvicorn 进程..." Yellow
$ts = Get-Date -Format "yyyyMMdd_HHmmss"
$beTmp = Join-Path $env:TEMP "neozqyy_watchdog_be_${ts}.ps1"
$q = [char]39
# CHANGE 2026-05-05 v3 | 调用 start_uvicorn.py,wildcard 字符串封装在 Python 内部
$beLines = @(
"`$env:NEOZQYY_ROOT = ${q}${ProjectRoot}${q}"
""
"if (`$PSVersionTable.PSVersion.Major -lt 7 -and -not `$env:WT_SESSION) {"
" chcp 65001 | Out-Null"
" `$env:NO_COLOR = ${q}1${q}"
" [Console]::OutputEncoding = [System.Text.Encoding]::UTF8"
" Write-Host ${q}[提示] PS 5.1 传统控制台,已启用 UTF-8 + 禁用 ANSI 颜色${q} -ForegroundColor Yellow"
"}"
""
"Set-Location -LiteralPath ${q}${backendDir}${q}"
"Write-Host ${q}=== 后端 FastAPI (watchdog 自动重启) ===${q} -ForegroundColor Magenta"
"Write-Host `"NEOZQYY_ROOT=`$env:NEOZQYY_ROOT`""
""
"& ${q}${venvPython}${q} ${q}${backendDir}\start_uvicorn.py${q} --port ${Port}"
"Write-Host ${q}后端已退出,按任意键关闭...${q} -ForegroundColor Red"
"`$null = `$Host.UI.RawUI.ReadKey(${q}NoEcho,IncludeKeyDown${q})"
)
$beLines | Set-Content -Path $beTmp -Encoding UTF8
Start-Process $psExe -ArgumentList "-NoExit", "-ExecutionPolicy", "Bypass", "-File", $beTmp | Out-Null
Write-Log "新 uvicorn 进程已 spawn,等待 ${StartupGrace}s 启动" Green
}
function Force-RestartBackend {
Write-Log "===========================================" Red
Write-Log "检测到后端卡死,执行强制重启" Red
Write-Log "===========================================" Red
if (-not (Stop-StuckBackend)) { return $false }
Start-Sleep -Seconds 1
Start-NewBackend
return $true
}
# ── 主循环 ─────────────────────────────────────────
Write-Host ""
Write-Host "========================================" -ForegroundColor Cyan
Write-Host " 后端看门狗 backend-watchdog" -ForegroundColor Cyan
Write-Host "========================================" -ForegroundColor Cyan
Write-Log "ProjectRoot = $ProjectRoot" Gray
Write-Log "HealthURL = $healthUrl" Gray
Write-Log "CheckInterval = ${CheckInterval}s" Gray
Write-Log "FailThreshold = $FailThreshold (累计 $($CheckInterval * $FailThreshold)s 不响应触发重启)" Gray
Write-Log "StartupGrace = ${StartupGrace}s (重启后宽限期)" Gray
Write-Host ""
$failCount = 0
$lastRestart = [DateTime]::MinValue
$startupSeen = $false # 首次见到健康才启用探针
$restartTimes = @() # 重启时间窗(滑动 1 小时,PS array 避免 List null 异常)
while ($true) {
# grace 期间不探针
$sinceRestart = (Get-Date) - $lastRestart
if ($sinceRestart.TotalSeconds -lt $StartupGrace) {
Start-Sleep -Seconds $CheckInterval
continue
}
if (Test-BackendAlive) {
if (-not $startupSeen) {
Write-Log "首次健康探针 200,看门狗激活" Green
$startupSeen = $true
} elseif ($failCount -gt 0) {
Write-Log "后端恢复 (此前连续失败 $failCount 次)" Green
}
$failCount = 0
} else {
if (-not $startupSeen) {
# 启动期还没见过 200,先静默等待
Start-Sleep -Seconds $CheckInterval
continue
}
$failCount++
Write-Log "健康探针失败 ($failCount/$FailThreshold)" Yellow
if ($failCount -ge $FailThreshold) {
# CHANGE 2026-05-05 | 重启频率限制:每小时最多 N 次,超出停止看门狗自我保护
$cutoff = (Get-Date).AddHours(-1)
$restartTimes = @($restartTimes | Where-Object { $_ -gt $cutoff })
if ($restartTimes.Count -ge $MaxRestartsPerHour) {
Write-Log "===========================================" Red
Write-Log "已达每小时最大重启次数 $MaxRestartsPerHour,看门狗停止" Red
Write-Log "可能是探针误报或后端持续异常,请人工排查" Red
Write-Log "===========================================" Red
exit 1
}
if (Force-RestartBackend) {
$lastRestart = Get-Date
$restartTimes += $lastRestart
Write-Log "本小时已重启 $($restartTimes.Count)/$MaxRestartsPerHour" DarkYellow
$startupSeen = $false
}
$failCount = 0
}
}
Start-Sleep -Seconds $CheckInterval
}