dxf: add manual recovery mode (#59641)

ref #59640

Author: D3Hunter

pkg/ddl/backfilling_dist_scheduler_test.go

@@ -158,7 +158,7 @@ func TestBackfillingSchedulerGlobalSortMode(t *testing.T) {
 	ext.(*ddl.LitBackfillScheduler).GlobalSort = true
 	sch.Extension = ext
 
-	taskID, err := mgr.CreateTask(ctx, task.Key, proto.Backfill, 1, "", 0, task.Meta)
+	taskID, err := mgr.CreateTask(ctx, task.Key, proto.Backfill, 1, "", 0, proto.ExtraParams{}, task.Meta)
 	require.NoError(t, err)
 	task.ID = taskID
 	execIDs := []string{":4000"}

pkg/disttask/framework/doc.go

@@ -125,6 +125,14 @@
 //	   └─────────►│cancelling├────┘
 //	              └──────────┘
 //
+// Note: if ManualRecovery is enabled, when some subtask failed, the task will
+// move to `awaiting-resolution` state, and manual operation is needed for the
+// task to continue. This mechanism is used for debugging, some bug such as those
+// on global-sort are harder to investigate without the intermediate files, or to
+// manually recover from some error when importing large mount of data using
+// global-sort where one round of import takes a lot of time, it might be more
+// flexible and efficient than retrying the whole task.
+//
 // pause/resume state transition:
 // as we don't know the state of the task before `paused`, so the state after
 // `resuming` is always `running`.

pkg/disttask/framework/handle/handle.go

@@ -69,7 +69,7 @@ func SubmitTask(ctx context.Context, taskKey string, taskType proto.TaskType, co
 		return nil, storage.ErrTaskAlreadyExists
 	}
 
-	taskID, err := taskManager.CreateTask(ctx, taskKey, taskType, concurrency, targetScope, maxNodeCnt, taskMeta)
+	taskID, err := taskManager.CreateTask(ctx, taskKey, taskType, concurrency, targetScope, maxNodeCnt, proto.ExtraParams{}, taskMeta)
 	if err != nil {
 		return nil, err
 	}

pkg/disttask/framework/integrationtests/BUILD.bazel

@@ -17,7 +17,7 @@ go_test(
     ],
     flaky = True,
     race = "off",
-    shard_count = 23,
+    shard_count = 22,
     deps = [
         "//pkg/config",
         "//pkg/ddl",

pkg/disttask/framework/integrationtests/framework_err_handling_test.go

@@ -15,26 +15,84 @@
 package integrationtests
 
 import (
+	"context"
+	"fmt"
+	"sync/atomic"
 	"testing"
+	"time"
 
+	"github.com/pingcap/errors"
 	"github.com/pingcap/tidb/pkg/disttask/framework/proto"
+	"github.com/pingcap/tidb/pkg/disttask/framework/storage"
+	"github.com/pingcap/tidb/pkg/disttask/framework/taskexecutor/execute"
 	"github.com/pingcap/tidb/pkg/disttask/framework/testutil"
+	"github.com/pingcap/tidb/pkg/testkit"
 	"github.com/stretchr/testify/require"
 )
 
-func TestRetryErrOnNextSubtasksBatch(t *testing.T) {
+func TestOnTaskError(t *testing.T) {
 	c := testutil.NewTestDXFContext(t, 2, 16, true)
-	registerExampleTask(t, c.MockCtrl, testutil.GetPlanErrSchedulerExt(c.MockCtrl, c.TestContext), c.TestContext, nil)
-	submitTaskAndCheckSuccessForBasic(c.Ctx, t, "key1", c.TestContext)
-}
 
-func TestPlanNotRetryableOnNextSubtasksBatchErr(t *testing.T) {
-	c := testutil.NewTestDXFContext(t, 2, 16, true)
+	t.Run("retryable error on OnNextSubtasksBatch", func(t *testing.T) {
+		registerExampleTask(t, c.MockCtrl, testutil.GetPlanErrSchedulerExt(c.MockCtrl, c.TestContext), c.TestContext, nil)
+		submitTaskAndCheckSuccessForBasic(c.Ctx, t, "key1", c.TestContext)
+	})
+
+	t.Run("non retryable error on OnNextSubtasksBatch", func(t *testing.T) {
+		registerExampleTask(t, c.MockCtrl, testutil.GetPlanNotRetryableErrSchedulerExt(c.MockCtrl), c.TestContext, nil)
+		task := testutil.SubmitAndWaitTask(c.Ctx, t, "key2-1", "", 1)
+		require.Equal(t, proto.TaskStateReverted, task.State)
+		registerExampleTask(t, c.MockCtrl, testutil.GetStepTwoPlanNotRetryableErrSchedulerExt(c.MockCtrl), c.TestContext, nil)
+		task = testutil.SubmitAndWaitTask(c.Ctx, t, "key2-2", "", 1)
+		require.Equal(t, proto.TaskStateReverted, task.State)
+	})
+
+	prepareForAwaitingResolutionTestFn := func(t *testing.T, taskKey string) int64 {
+		subtaskErrRetryable := atomic.Bool{}
+		executorExt := testutil.GetTaskExecutorExt(c.MockCtrl,
+			func(task *proto.Task) (execute.StepExecutor, error) {
+				return testutil.GetCommonStepExecutor(c.MockCtrl, task.Step, func(ctx context.Context, subtask *proto.Subtask) error {
+					if !subtaskErrRetryable.Load() {
+						return errors.New("non retryable subtask error")
+					}
+					return nil
+				}), nil
+			},
+			func(error) bool {
+				return subtaskErrRetryable.Load()
+			},
+		)
+		testutil.RegisterExampleTask(t, testutil.GetPlanErrSchedulerExt(c.MockCtrl, c.TestContext),
+			executorExt, testutil.GetCommonCleanUpRoutine(c.MockCtrl))
+		tm, err := storage.GetTaskManager()
+		require.NoError(t, err)
+		taskID, err := tm.CreateTask(c.Ctx, taskKey, proto.TaskTypeExample, 1, "",
+			2, proto.ExtraParams{ManualRecovery: true}, nil)
+		require.NoError(t, err)
+		require.Eventually(t, func() bool {
+			task, err := tm.GetTaskByID(c.Ctx, taskID)
+			require.NoError(t, err)
+			return task.State == proto.TaskStateAwaitingResolution
+		}, 10*time.Second, 100*time.Millisecond)
+		subtaskErrRetryable.Store(true)
+		return taskID
+	}
+
+	t.Run("task enter awaiting-resolution state if ManualRecovery set, success after manual recover", func(t *testing.T) {
+		taskKey := "key3-1"
+		taskID := prepareForAwaitingResolutionTestFn(t, taskKey)
+		tk := testkit.NewTestKit(t, c.Store)
+		tk.MustExec(fmt.Sprintf("update mysql.tidb_background_subtask set state='pending' where state='failed' and task_key= %d", taskID))
+		tk.MustExec(fmt.Sprintf("update mysql.tidb_global_task set state='running' where id = %d", taskID))
+		task := testutil.WaitTaskDone(c.Ctx, t, taskKey)
+		require.Equal(t, proto.TaskStateSucceed, task.State)
+	})
 
-	registerExampleTask(t, c.MockCtrl, testutil.GetPlanNotRetryableErrSchedulerExt(c.MockCtrl), c.TestContext, nil)
-	task := testutil.SubmitAndWaitTask(c.Ctx, t, "key1", "", 1)
-	require.Equal(t, proto.TaskStateReverted, task.State)
-	registerExampleTask(t, c.MockCtrl, testutil.GetStepTwoPlanNotRetryableErrSchedulerExt(c.MockCtrl), c.TestContext, nil)
-	task = testutil.SubmitAndWaitTask(c.Ctx, t, "key2", "", 1)
-	require.Equal(t, proto.TaskStateReverted, task.State)
+	t.Run("task enter awaiting-resolution state if ManualRecovery set, cancel also works", func(t *testing.T) {
+		taskKey := "key4-1"
+		taskID := prepareForAwaitingResolutionTestFn(t, taskKey)
+		require.NoError(t, c.TaskMgr.CancelTask(c.Ctx, taskID))
+		task := testutil.WaitTaskDone(c.Ctx, t, taskKey)
+		require.Equal(t, proto.TaskStateReverted, task.State)
+	})
 }

pkg/disttask/framework/mock/scheduler_mock.go

@@ -16,6 +16,7 @@ package planner
 
 import (
 	"github.com/pingcap/tidb/pkg/config"
+	"github.com/pingcap/tidb/pkg/disttask/framework/proto"
 	"github.com/pingcap/tidb/pkg/disttask/framework/storage"
 )
 
@@ -47,6 +48,7 @@ func (*Planner) Run(planCtx PlanCtx, plan LogicalPlan) (int64, error) {
 		planCtx.ThreadCnt,
 		config.GetGlobalConfig().Instance.TiDBServiceScope,
 		planCtx.MaxNodeCnt,
+		proto.ExtraParams{},
 		taskMeta,
 	)
 }

pkg/disttask/framework/planner/planner.go

@@ -22,17 +22,18 @@ import (
 
 // see doc.go for more details.
 const (
-	TaskStatePending    TaskState = "pending"
-	TaskStateRunning    TaskState = "running"
-	TaskStateSucceed    TaskState = "succeed"
-	TaskStateFailed     TaskState = "failed"
-	TaskStateReverting  TaskState = "reverting"
-	TaskStateReverted   TaskState = "reverted"
-	TaskStateCancelling TaskState = "cancelling"
-	TaskStatePausing    TaskState = "pausing"
-	TaskStatePaused     TaskState = "paused"
-	TaskStateResuming   TaskState = "resuming"
-	TaskStateModifying  TaskState = "modifying"
+	TaskStatePending            TaskState = "pending"
+	TaskStateRunning            TaskState = "running"
+	TaskStateSucceed            TaskState = "succeed"
+	TaskStateFailed             TaskState = "failed"
+	TaskStateReverting          TaskState = "reverting"
+	TaskStateAwaitingResolution TaskState = "awaiting-resolution"
+	TaskStateReverted           TaskState = "reverted"
+	TaskStateCancelling         TaskState = "cancelling"
+	TaskStatePausing            TaskState = "pausing"
+	TaskStatePaused             TaskState = "paused"
+	TaskStateResuming           TaskState = "resuming"
+	TaskStateModifying          TaskState = "modifying"
 )
 
 type (
@@ -66,6 +67,15 @@ const (
 // TODO: remove this limit later.
 var MaxConcurrentTask = 16
 
+// ExtraParams is the extra params of task.
+// Note: only store params that's not used for filter or sort in this struct.
+type ExtraParams struct {
+	// ManualRecovery indicates whether the task can be recovered manually.
+	// if enabled, the task will enter 'awaiting-resolution' state when it failed,
+	// then the user can recover the task manually or fail it if it's not recoverable.
+	ManualRecovery bool `json:"manual_recovery"`
+}
+
 // TaskBase contains the basic information of a task.
 // we define this to avoid load task meta which might be very large into memory.
 type TaskBase struct {
@@ -87,6 +97,7 @@ type TaskBase struct {
 	TargetScope  string
 	CreateTime   time.Time
 	MaxNodeCount int
+	ExtraParams
 }
 
 // IsDone checks if the task is done.

pkg/disttask/framework/proto/task.go

@@ -45,6 +45,8 @@ type TaskManager interface {
 	FailTask(ctx context.Context, taskID int64, currentState proto.TaskState, taskErr error) error
 	// RevertTask updates task state to reverting, and task error.
 	RevertTask(ctx context.Context, taskID int64, taskState proto.TaskState, taskErr error) error
+	// AwaitingResolveTask updates task state to awaiting-resolve, also set task err.
+	AwaitingResolveTask(ctx context.Context, taskID int64, taskState proto.TaskState, taskErr error) error
 	// RevertedTask updates task state to reverted.
 	RevertedTask(ctx context.Context, taskID int64) error
 	// PauseTask updated task state to pausing.

pkg/disttask/framework/scheduler/interface.go

@@ -388,7 +388,7 @@ func (s *BaseScheduler) onRunning() error {
 		if len(subTaskErrs) > 0 {
 			s.logger.Warn("subtasks encounter errors", zap.Errors("subtask-errs", subTaskErrs))
 			// we only store the first error as task error.
-			return s.revertTask(subTaskErrs[0])
+			return s.revertTaskOrManualRecover(subTaskErrs[0])
 		}
 	} else if s.isStepSucceed(cntByStates) {
 		return s.switch2NextStep()
@@ -586,6 +586,20 @@ func (s *BaseScheduler) revertTask(taskErr error) error {
 	return nil
 }
 
+func (s *BaseScheduler) revertTaskOrManualRecover(taskErr error) error {
+	task := s.getTaskClone()
+	if task.ManualRecovery {
+		if err := s.taskMgr.AwaitingResolveTask(s.ctx, task.ID, task.State, taskErr); err != nil {
+			return err
+		}
+		task.State = proto.TaskStateAwaitingResolution
+		task.Error = taskErr
+		s.task.Store(task)
+		return nil
+	}
+	return s.revertTask(taskErr)
+}
+
 // MockServerInfo exported for scheduler_test.go
 var MockServerInfo atomic.Pointer[[]string]