MPP coordinator will leak when error happened before respIter is created

[yibin87]

When NewExecutorWithRetry returns error, the mpp coordinator may have been already created inside mppExec and registered in global MPPCoordinatorManager map:

tidb/pkg/executor/mpp_gather.go

Lines 104 to 108 in 907eddd

	if e.mppExec, err = mpp.NewExecutorWithRetry(ctx, e.Ctx(), e.memTracker, planIDs, e.originalPlan, e.startTS, e.mppQueryID, e.is); err != nil {
	return err
	}
	e.kvRanges = e.mppExec.KVRanges
	e.respIter = distsql.GenSelectResultFromMPPResponse(e.Ctx().GetDistSQLCtx(), e.RetFieldTypes(), planIDs, e.ID(), e.mppExec)

The mpp coordinator resource will not be released even when MPPGather is closed, because mppExec will be closed only when respIter is not nil:

tidb/pkg/executor/mpp_gather.go

Lines 133 to 135 in 907eddd

	if e.respIter != nil {
	return e.respIter.Close()
	}

[yibin87]

/type bug

[yibin87]

/severity major

[yibin87]

/sig execution

[yibin87]

The issue is detected by our random failpoint tests, occasionally. And we have clean-overtime mpp-coordinator mechanism to delete them, so set to minor severity.

[yibin87]

/severity minor