EdgeClaw Memory scheduled maintenance 偶发 SQLite database is locked

标注

AI 提交（由用户授权的 AI 助手根据本机日志与源码检查整理）。

问题描述

PilotDeck 在启用 EdgeClaw Memory 后，运行过程中会偶发出现 SQLite 报错：database is locked。

该问题不是单个 workspace 的数据库损坏；从日志看，多个 memory workspace 都出现过同类锁错误。发生时会导致 memory scheduler / maintenance 任务失败，并可能在用户会话执行期间伴随 gateway retry 或错误事件。

观察到的报错

日志文件：~/.pilotdeck/server-3001.log

出现过的典型日志：

[server] [memory-scheduler] scheduled maintenance failed for /home/ye/.pilotdeck/memory/workspaces/8547594726: Error: database is locked
  code: 'ERR_SQLITE_ERROR',
  errcode: 5,
  errstr: 'database is locked'

[server] [memory-scheduler] scheduled maintenance failed for /home/ye/.pilotdeck/memory/workspaces/da975b3393: Error: database is locked
    at MemoryRepository.deletePipelineState (.../src/context/memory/edgeclaw-memory-core/lib/core/storage/sqlite.js:709:75)
    at EdgeClawMemoryService.runDueScheduledMaintenance (.../src/context/memory/edgeclaw-memory-core/lib/service.js:687:36)
  code: 'ERR_SQLITE_ERROR',
  errcode: 5,
  errstr: 'database is locked'

[server] [memory-scheduler] scheduled maintenance failed for /home/ye/.pilotdeck/memory/workspaces/80be3f1ce8: Error: database is locked
    at MemoryRepository.deletePipelineState (.../src/context/memory/edgeclaw-memory-core/lib/core/storage/sqlite.js:709:75)
    at EdgeClawMemoryService.runDueScheduledMaintenance (.../src/context/memory/edgeclaw-memory-core/lib/service.js:669:36)
  code: 'ERR_SQLITE_ERROR',
  errcode: 5,
  errstr: 'database is locked'

另有 gateway 侧日志：

[gateway] database is locked

涉及的 workspace 不止一个

本机日志中至少出现过以下 memory workspace：

/home/ye/.pilotdeck/memory/workspaces/8547594726
/home/ye/.pilotdeck/memory/workspaces/da975b3393
/home/ye/.pilotdeck/memory/workspaces/80be3f1ce8

读取 pipeline_state.workspaceDir 后，对应关系大致为：

8547594726  workspace=/home/ye/projects/PilotDeck
80be3f1ce8  workspace=/home/ye/.pilotdeck/projects/gongzuoqu2
da975b3393  workspace=/home/ye/.pilotdeck

进程与文件句柄证据

运行时观察到两个 PilotDeck 相关 Node 进程同时打开同一批 memory SQLite 文件：

PID 2528  node --import tsx /home/ye/projects/PilotDeck/ui/server/index.js
PID 2557  node ... src/cli/pilotdeck.ts server --port 18790

二者均持有类似文件句柄：

/home/ye/.pilotdeck/memory/workspaces/80be3f1ce8/control.sqlite
/home/ye/.pilotdeck/memory/workspaces/80be3f1ce8/control.sqlite-wal
/home/ye/.pilotdeck/memory/workspaces/80be3f1ce8/control.sqlite-shm

/home/ye/.pilotdeck/memory/workspaces/8547594726/control.sqlite
/home/ye/.pilotdeck/memory/workspaces/8547594726/control.sqlite-wal
/home/ye/.pilotdeck/memory/workspaces/8547594726/control.sqlite-shm

/home/ye/.pilotdeck/memory/workspaces/da975b3393/control.sqlite
/home/ye/.pilotdeck/memory/workspaces/da975b3393/control.sqlite-wal
/home/ye/.pilotdeck/memory/workspaces/da975b3393/control.sqlite-shm

相关源码位置

本地源码检查发现，Web server 和 Gateway 都会触发 memory maintenance。

Web server 侧：

// ui/server/services/memoryService.js
const MEMORY_SCHEDULER_INTERVAL_MS = 60_000;

async function executeScheduledMaintenanceForDataDir(dataDir) {
  const { service } = getOrCreateServiceForDataDir(dataDir);
  return enqueueMaintenanceTask(dataDir, async () =>
    service.runDueScheduledMaintenance('scheduled:server_scheduler')
  );
}

Gateway 侧：

// src/cli/createLocalGateway.ts
scheduleMemoryMaintenance(projectKey?: string): void {
  const runtime = this.resolve(projectKey);
  const service = runtime.memoryService;
  if (!service) return;
  ...
  await service.runDueScheduledMaintenance("scheduled");
}

报错堆栈对应的写入点：

// src/context/memory/edgeclaw-memory-core/lib/core/storage/sqlite.js
deletePipelineState(key) {
  this.db.prepare("DELETE FROM pipeline_state WHERE state_key = ?").run(key);
}

SQLite 初始化中观察到 WAL：

PRAGMA journal_mode = WAL;

影响

• memory scheduler / maintenance 偶发失败；
• 多个 workspace 都可能触发，不局限于单一项目；
• 发生在用户交互/任务执行期间时，可能导致日志中出现 retry 或错误事件；
• 当前观察中 pending sessions 没有长期积压，但该错误会反复污染日志，并可能导致维护任务未按预期完成。

环境信息

• OS: Linux 6.17.0-35-generic x64
• Node.js: v24.15.0
• PilotDeck 启动方式：systemd user service
• 相关服务：

node --import tsx /home/ye/projects/PilotDeck/ui/server/index.js
node ... src/cli/pilotdeck.ts server --port 18790

期望行为

EdgeClaw Memory 的 scheduled maintenance 在多进程/多 runtime 同时存在时，不应因为 SQLite 写锁直接失败；至少不应让常规维护任务反复产出 database is locked 错误。