feat(workflow): 恢复例程 + 心跳 + 启动/周期扫描
This commit is contained in:
@@ -156,9 +156,26 @@ func (d *execWorkflowDao) UpdateRetry(ctx context.Context, id int64, retryable i
|
||||
return err
|
||||
}
|
||||
|
||||
// ListRecoverable 返回可恢复执行:僵尸运行中(status=1 且心跳陈旧)或可重试失败(status=3 且 retryable=1 且未耗尽)
|
||||
// GetByIdNoTenant 跨租户按 id 读取执行记录(不追加 tenant_id 过滤)。
|
||||
// 供恢复例程抢锁后、尚未知晓租户时重读 exec 使用;用户路径请继续用租户隔离的 GetById。
|
||||
func (d *execWorkflowDao) GetByIdNoTenant(ctx context.Context, id int64) (res *entity.ExecWorkflow, err error) {
|
||||
r, err := gfdb.DB(ctx, public.DbNameBlackDeacon).Model(ctx, public.TableNameExecWorkflow).
|
||||
NoTenantId(ctx).
|
||||
Where(entity.ExecWorkflowCol.Id, id).
|
||||
One()
|
||||
if err != nil {
|
||||
return
|
||||
}
|
||||
err = r.Struct(&res)
|
||||
return
|
||||
}
|
||||
|
||||
// ListRecoverable 返回可恢复执行:僵尸运行中(status=1 且心跳陈旧)或可重试失败(status=3 且 retryable=1 且未耗尽)。
|
||||
// 调用方是跨租户的恢复扫描(无 HTTP 用户),显式 NoTenantId 走系统级扫描,避免僵尸执行因租户过滤漏检。
|
||||
// ctx 必须携带 OTel span(NoTenantId 依赖 traceID 作为 gcache 标记键),恢复扫描由 scanAndRecover 保证。
|
||||
func (d *execWorkflowDao) ListRecoverable(ctx context.Context, now int64, staleBefore int64, maxRetry int) (res []*entity.ExecWorkflow, err error) {
|
||||
r, err := gfdb.DB(ctx, public.DbNameBlackDeacon).Model(ctx, public.TableNameExecWorkflow).
|
||||
NoTenantId(ctx).
|
||||
Where(fmt.Sprintf("(%s = ? AND %s < ?) OR (%s = ? AND %s = 1 AND %s < ?)",
|
||||
entity.ExecWorkflowCol.Status, entity.ExecWorkflowCol.LastHeartbeat,
|
||||
entity.ExecWorkflowCol.Status, entity.ExecWorkflowCol.Retryable, entity.ExecWorkflowCol.RetryCount),
|
||||
|
||||
@@ -12,6 +12,7 @@ import (
|
||||
|
||||
"gitea.redpowerfuture.com/red-future/common/beans"
|
||||
_ "github.com/gogf/gf/contrib/drivers/pgsql/v2"
|
||||
"github.com/gogf/gf/v2/net/gtrace"
|
||||
)
|
||||
|
||||
// TestExecWorkflowRetryColsRoundTrip 验证三新列读写 + ListRecoverable 三类判定(需 AI_AGENT_TEST_DB=1 且本地 PG 可用)
|
||||
@@ -19,8 +20,11 @@ func TestExecWorkflowRetryColsRoundTrip(t *testing.T) {
|
||||
if !isTestDBEnabled(t) {
|
||||
t.Skip("skip: AI_AGENT_TEST_DB not set")
|
||||
}
|
||||
// gfdb 的 Hook 依赖 ctx 中注入用户信息(租户/创建人),否则 Insert 报 "token 数据为空"
|
||||
ctx := context.WithValue(context.Background(), "user", &beans.User{
|
||||
// gfdb 的 Hook 依赖 ctx 中注入用户信息(租户/创建人),否则 Insert 报 "token 数据为空"。
|
||||
// 先造 span:ListRecoverable 链 NoTenantId 依赖 traceID 作 gcache 标记键,无 span 时 getTraceID 返回空 → NoTenantId 返回 nil 会 panic
|
||||
ctx, span := gtrace.NewSpan(context.Background(), "test.recover")
|
||||
defer span.End()
|
||||
ctx = context.WithValue(ctx, "user", &beans.User{
|
||||
UserName: "test-retry", TenantId: 1,
|
||||
})
|
||||
id, err := ExecWorkflowDao.Insert(ctx, &sessionDto.CreateWorkflowReq{
|
||||
|
||||
@@ -0,0 +1,172 @@
|
||||
package flow
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"fmt"
|
||||
"time"
|
||||
|
||||
"ai-agent/workflow/consts/flow"
|
||||
flowDao "ai-agent/workflow/dao/flow"
|
||||
sessionDao "ai-agent/workflow/dao/session"
|
||||
"ai-agent/workflow/model/entity"
|
||||
|
||||
"gitea.redpowerfuture.com/red-future/common/beans"
|
||||
"github.com/gogf/gf/v2/frame/g"
|
||||
"github.com/gogf/gf/v2/net/gtrace"
|
||||
"github.com/google/uuid"
|
||||
)
|
||||
|
||||
// 恢复相关常量(可参数化调整)
|
||||
const (
|
||||
heartbeatInterval = 30 * time.Second // 心跳 touch 间隔
|
||||
heartbeatStaleAfter = 2 * heartbeatInterval // 心跳陈旧阈值(>2×间隔),小于此视为僵尸
|
||||
recoverScanInterval = 30 * time.Second // 周期扫描间隔
|
||||
recoverLockTTL = 15 * time.Minute // 恢复锁 TTL(覆盖长时间续跑)
|
||||
execMaxRetryCount = 2 // 整次执行最多自动重试 2 次(共 3 次尝试)
|
||||
)
|
||||
|
||||
// errExecAlreadyRunning 用户触发时该执行已在运行(本节点/其它节点后台恢复),不新建执行
|
||||
var errExecAlreadyRunning = errors.New("工作流正在执行中,不重复执行")
|
||||
|
||||
// shouldRetry 错误分类(spec §3):用户取消不重试;其余(模型/DB/网络/超时/panic)程序报错重试
|
||||
func shouldRetry(err error) bool {
|
||||
return err != nil && !errors.Is(err, context.Canceled) && !errors.Is(err, errExecAlreadyRunning)
|
||||
}
|
||||
|
||||
// StartRecoveryLoop 启动恢复扫描:先立即扫一次,再周期扫描。
|
||||
// 多节点各自扫描,靠 Redis 锁对同一 exec 抢占去重(spec §8.2/8.3)
|
||||
func StartRecoveryLoop(ctx context.Context) {
|
||||
go func() {
|
||||
scanAndRecover(ctx)
|
||||
ticker := time.NewTicker(recoverScanInterval)
|
||||
defer ticker.Stop()
|
||||
for {
|
||||
select {
|
||||
case <-ticker.C:
|
||||
scanAndRecover(ctx)
|
||||
case <-ctx.Done():
|
||||
return
|
||||
}
|
||||
}
|
||||
}()
|
||||
}
|
||||
|
||||
// scanAndRecover 扫描可恢复执行并逐个异步恢复(不阻塞扫描循环)。
|
||||
// 恢复运行在无 HTTP 用户的后台:ListRecoverable 走跨租户 NoTenantId,需要 ctx 携带 OTel span
|
||||
// (NoTenantId 以 traceID 为 gcache 标记键,无 span 时 getTraceID 返回空 → NoTenantId 返回 nil 会 panic)
|
||||
func scanAndRecover(ctx context.Context) {
|
||||
scanCtx, span := gtrace.NewSpan(ctx, "workflow.recover.scan")
|
||||
defer span.End()
|
||||
now := time.Now().UnixMilli()
|
||||
rows, err := sessionDao.ExecWorkflowDao.ListRecoverable(scanCtx, now, now-int64(heartbeatStaleAfter/time.Millisecond), execMaxRetryCount)
|
||||
if err != nil {
|
||||
g.Log().Errorf(scanCtx, "扫描可恢复执行失败: %v", err)
|
||||
return
|
||||
}
|
||||
for _, r := range rows {
|
||||
go recoverExecution(context.WithoutCancel(scanCtx), r.Id)
|
||||
}
|
||||
}
|
||||
|
||||
// recoverExecution 统一恢复例程(spec §7):抢锁 → 判定 → 置运行中续跑。
|
||||
// 两个触发源共用:启动/周期扫描、executeOrResume 对僵尸行的用户触发。
|
||||
func recoverExecution(parentCtx context.Context, execId int64) {
|
||||
lock := NewRedisLock(fmt.Sprintf("workflow:exec:recover:%d", execId), int64(recoverLockTTL/time.Second))
|
||||
ctx, cancel := context.WithTimeout(context.WithoutCancel(parentCtx), 2*time.Hour)
|
||||
defer cancel()
|
||||
|
||||
ok, err := lock.Acquire(ctx)
|
||||
if err != nil {
|
||||
g.Log().Errorf(ctx, "恢复抢锁失败 execId=%d: %v", execId, err)
|
||||
return
|
||||
}
|
||||
if !ok {
|
||||
g.Log().Infof(ctx, "恢复锁被其它节点持有,跳过 execId=%d", execId)
|
||||
return
|
||||
}
|
||||
defer lock.Release(context.Background())
|
||||
|
||||
// 抢锁后重读:此刻租户未知(该 exec 可能属于任意租户),必须跨租户读
|
||||
exec, err := sessionDao.ExecWorkflowDao.GetByIdNoTenant(ctx, execId)
|
||||
if err != nil || exec == nil {
|
||||
return
|
||||
}
|
||||
if !isRecoverable(exec, time.Now().UnixMilli()) {
|
||||
return // 锁等待期间状态已变化(其它节点已恢复/用户取消/已耗尽)
|
||||
}
|
||||
|
||||
stop := startHeartbeat(ctx, execId)
|
||||
defer stop()
|
||||
|
||||
// 置运行中并续跑(BuildExecution false → 断点续跑:checkpoint 有则从中继续,无则从头 + 段/memo 复用)
|
||||
saveCtx := context.WithoutCancel(parentCtx)
|
||||
// 恢复无 HTTP 用户,但图节点 lambda 的 INSERT(node_execution/flow_async_task/segment_result)走
|
||||
// insertHook 硬性要求 user;用 exec 所属租户合成系统用户 ctx(保留 span),供后续落库使用
|
||||
userCtx := context.WithValue(saveCtx, "user", &beans.User{UserName: "workflow-recover", TenantId: exec.TenantId})
|
||||
nodeGroupId := uuid.NewString() // 置运行中与图执行的节点组标识须一致
|
||||
if err := sessionDao.ExecWorkflowDao.ResetRunning(userCtx, execId, nodeGroupId); err != nil {
|
||||
g.Log().Errorf(ctx, "恢复置运行中失败 execId=%d: %v", execId, err)
|
||||
return
|
||||
}
|
||||
err = BuildExecution(userCtx, false, exec.FlowId, execId, nodeGroupId, exec.SessionId, exec.RequestParams)
|
||||
if err != nil {
|
||||
// 续跑失败:错误分类落库(retryable/retry_count),留给下一轮扫描决定是否再恢复
|
||||
if errors.Is(err, context.Canceled) {
|
||||
_ = sessionDao.ExecWorkflowDao.UpdateRetry(userCtx, execId, 0, exec.RetryCount)
|
||||
} else {
|
||||
_ = sessionDao.ExecWorkflowDao.UpdateRetry(userCtx, execId, 1, exec.RetryCount+1)
|
||||
// 终局清理(Task 5 用户裁定):非用户取消且重试耗尽 → 执行永久失败,
|
||||
// 清理该 exec 残留的 flow_async_task 孤儿缓存,避免未来复用同一 execId 的运行误取到过期 done 结果
|
||||
if exec.RetryCount+1 >= execMaxRetryCount {
|
||||
_ = flowDao.FlowAsyncTaskDao.DeleteByExecution(userCtx, execId)
|
||||
}
|
||||
}
|
||||
recordWorkflow(userCtx, execId, 0, err)
|
||||
return
|
||||
}
|
||||
// 续跑成功:recordWorkflow 落 status=2;BuildExecution 已清理 checkpoint/segment_result/flow_async_task
|
||||
recordWorkflow(userCtx, execId, 0, nil)
|
||||
}
|
||||
|
||||
// isRecoverable 判定可恢复(spec §3):僵尸运行中(status=1 且心跳陈旧)或可重试失败(status=3 且 retryable=1 且未耗尽)。
|
||||
// 与 ListRecoverable SQL 判定一致;last_heartbeat=0(老数据/默认)视为陈旧。
|
||||
func isRecoverable(exec *entity.ExecWorkflow, nowMs int64) bool {
|
||||
if exec == nil || exec.Status == nil {
|
||||
return false
|
||||
}
|
||||
status := *exec.Status
|
||||
switch status {
|
||||
case *flow.FlowExecutionStatusRunning.Code():
|
||||
return exec.LastHeartbeat < nowMs-int64(heartbeatStaleAfter/time.Millisecond)
|
||||
case *flow.FlowExecutionStatusFailed.Code():
|
||||
return exec.Retryable == 1 && exec.RetryCount < execMaxRetryCount
|
||||
default:
|
||||
return false
|
||||
}
|
||||
}
|
||||
|
||||
// startHeartbeat 后台心跳 goroutine:每 30s touch last_heartbeat,返回 stop 函数。
|
||||
// 覆盖正常执行与恢复执行,崩溃前最后一次心跳即崩溃近似时间戳(spec §4)
|
||||
func startHeartbeat(ctx context.Context, execId int64) func() {
|
||||
stopCh := make(chan struct{})
|
||||
done := make(chan struct{})
|
||||
go func() {
|
||||
defer close(done)
|
||||
ticker := time.NewTicker(heartbeatInterval)
|
||||
defer ticker.Stop()
|
||||
for {
|
||||
select {
|
||||
case <-ticker.C:
|
||||
if err := sessionDao.ExecWorkflowDao.TouchHeartbeat(ctx, execId); err != nil {
|
||||
g.Log().Warningf(ctx, "心跳落库失败 execId=%d: %v", execId, err)
|
||||
}
|
||||
case <-stopCh:
|
||||
return
|
||||
case <-ctx.Done():
|
||||
return
|
||||
}
|
||||
}
|
||||
}()
|
||||
return func() { close(stopCh); <-done }
|
||||
}
|
||||
@@ -0,0 +1,60 @@
|
||||
package flow
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
"ai-agent/workflow/consts/flow"
|
||||
"ai-agent/workflow/model/entity"
|
||||
)
|
||||
|
||||
func TestIsRecoverable(t *testing.T) {
|
||||
now := time.Now().UnixMilli()
|
||||
staleBefore := now - int64(heartbeatStaleAfter/time.Millisecond) - 1
|
||||
fresh := now // 心跳新鲜
|
||||
|
||||
running := func(hb int64) *entity.ExecWorkflow {
|
||||
return &entity.ExecWorkflow{Status: flow.FlowExecutionStatusRunning.Code(), LastHeartbeat: hb}
|
||||
}
|
||||
failed := func(retryable, retryCount int) *entity.ExecWorkflow {
|
||||
return &entity.ExecWorkflow{Status: flow.FlowExecutionStatusFailed.Code(), Retryable: retryable, RetryCount: retryCount}
|
||||
}
|
||||
|
||||
cases := []struct {
|
||||
name string
|
||||
exec *entity.ExecWorkflow
|
||||
want bool
|
||||
}{
|
||||
{"nil", nil, false},
|
||||
{"status=1 心跳陈旧 → 恢复", running(staleBefore), true},
|
||||
{"status=1 心跳新鲜 → 不恢复", running(fresh), false},
|
||||
{"status=1 心跳为0(老数据)→ 恢复", running(0), true},
|
||||
{"status=3 retryable=1 未耗尽 → 恢复", failed(1, 0), true},
|
||||
{"status=3 retryable=1 已耗尽 → 不恢复", failed(1, execMaxRetryCount), false},
|
||||
{"status=3 retryable=0(用户取消)→ 不恢复", failed(0, 0), false},
|
||||
{"status=2 成功 → 不恢复", &entity.ExecWorkflow{Status: flow.FlowExecutionStatusSuccess.Code()}, false},
|
||||
}
|
||||
for _, c := range cases {
|
||||
if got := isRecoverable(c.exec, now); got != c.want {
|
||||
t.Fatalf("%s: got %v want %v", c.name, got, c.want)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func TestShouldRetry(t *testing.T) {
|
||||
canceled := context.Canceled
|
||||
if shouldRetry(canceled) {
|
||||
t.Fatalf("用户取消不应重试")
|
||||
}
|
||||
if shouldRetry(nil) {
|
||||
t.Fatalf("nil 不应重试")
|
||||
}
|
||||
if shouldRetry(errExecAlreadyRunning) {
|
||||
t.Fatalf("运行中标记不应重试")
|
||||
}
|
||||
if !shouldRetry(errors.New("模型调用失败")) {
|
||||
t.Fatalf("程序报错应重试")
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user