executor.go

// Core Nodes Executor implementation
// This module implements the core Nodes executor.
// This executor is the starting point for executing any node in the workflow. Since Nodes in a workflow are composable,
// i.e., one node may contain other nodes, the Node Handler is recursive in nature.
// This executor handles the core logic for all nodes, but specific logic for handling different kinds of nodes is delegated
// to the respective node handlers
//
// Available node handlers are
// - Task: Arguably the most important handler as it handles all tasks. These include all plugins. The goal of the workflow is
//         is to run tasks, thus every workflow will contain atleast one TaskNode (except for the case, where the workflow
//          is purely a meta-workflow and can run other workflows
// - SubWorkflow: This is one of the most important handlers. It can executes Workflows that are nested inside a workflow
// - DynamicTask Handler: This is just a decorator on the Task Handler. It handles cases, in which the Task returns a futures
//                        file. Every Task is actually executed through the DynamicTaskHandler
// - Branch Handler: This handler is used to execute branches
// - Start & End Node handler: these are nominal handlers for the start and end node and do no really carry a lot of logic
package nodes

import (
	"context"
	"fmt"
	"time"

	"github.com/flyteorg/flytepropeller/pkg/controller/nodes/common"

	"github.com/flyteorg/flyteplugins/go/tasks/pluginmachinery/ioutils"
	errors2 "github.com/flyteorg/flytestdlib/errors"

	"github.com/flyteorg/flyteidl/clients/go/events"
	eventsErr "github.com/flyteorg/flyteidl/clients/go/events/errors"
	"github.com/flyteorg/flyteidl/gen/pb-go/flyteidl/core"
	"github.com/flyteorg/flyteidl/gen/pb-go/flyteidl/event"
	"github.com/flyteorg/flytestdlib/contextutils"
	"github.com/flyteorg/flytestdlib/logger"
	"github.com/flyteorg/flytestdlib/promutils"
	"github.com/flyteorg/flytestdlib/promutils/labeled"
	"github.com/flyteorg/flytestdlib/storage"
	"github.com/golang/protobuf/ptypes"
	v1 "k8s.io/apimachinery/pkg/apis/meta/v1"

	"github.com/flyteorg/flyteplugins/go/tasks/pluginmachinery/catalog"

	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"

	"github.com/flyteorg/flytepropeller/pkg/controller/config"

	"github.com/flyteorg/flytepropeller/pkg/apis/flyteworkflow/v1alpha1"
	"github.com/flyteorg/flytepropeller/pkg/controller/executors"
	"github.com/flyteorg/flytepropeller/pkg/controller/nodes/errors"
	"github.com/flyteorg/flytepropeller/pkg/controller/nodes/handler"
	"github.com/flyteorg/flytepropeller/pkg/controller/nodes/subworkflow/launchplan"
)

type nodeMetrics struct {
	Scope                         promutils.Scope
	FailureDuration               labeled.StopWatch
	SuccessDuration               labeled.StopWatch
	UserErrorDuration             labeled.StopWatch
	SystemErrorDuration           labeled.StopWatch
	UnknownErrorDuration          labeled.StopWatch
	PermanentUserErrorDuration    labeled.StopWatch
	PermanentSystemErrorDuration  labeled.StopWatch
	PermanentUnknownErrorDuration labeled.StopWatch
	ResolutionFailure             labeled.Counter
	InputsWriteFailure            labeled.Counter
	TimedOutFailure               labeled.Counter

	InterruptedThresholdHit      labeled.Counter
	InterruptibleNodesRunning    labeled.Counter
	InterruptibleNodesTerminated labeled.Counter

	// Measures the latency between the last parent node stoppedAt time and current node's queued time.
	TransitionLatency labeled.StopWatch
	// Measures the latency between the time a node's been queued to the time the handler reported the executable moved
	// to running state
	QueuingLatency         labeled.StopWatch
	NodeExecutionTime      labeled.StopWatch
	NodeInputGatherLatency labeled.StopWatch
}

// Implements the executors.Node interface
type nodeExecutor struct {
	nodeHandlerFactory              HandlerFactory
	enqueueWorkflow                 v1alpha1.EnqueueWorkflow
	store                           *storage.DataStore
	nodeRecorder                    events.NodeEventRecorder
	taskRecorder                    events.TaskEventRecorder
	metrics                         *nodeMetrics
	maxDatasetSizeBytes             int64
	outputResolver                  OutputResolver
	defaultExecutionDeadline        time.Duration
	defaultActiveDeadline           time.Duration
	maxNodeRetriesForSystemFailures uint32
	interruptibleFailureThreshold   uint32
	defaultDataSandbox              storage.DataReference
	shardSelector                   ioutils.ShardSelector
}

func (c *nodeExecutor) RecordTransitionLatency(ctx context.Context, dag executors.DAGStructure, nl executors.NodeLookup, node v1alpha1.ExecutableNode, nodeStatus v1alpha1.ExecutableNodeStatus) {
	if nodeStatus.GetPhase() == v1alpha1.NodePhaseNotYetStarted || nodeStatus.GetPhase() == v1alpha1.NodePhaseQueued {
		// Log transition latency (The most recently finished parent node endAt time to this node's queuedAt time -now-)
		t, err := GetParentNodeMaxEndTime(ctx, dag, nl, node)
		if err != nil {
			logger.Warnf(ctx, "Failed to record transition latency for node. Error: %s", err.Error())
			return
		}
		if !t.IsZero() {
			c.metrics.TransitionLatency.Observe(ctx, t.Time, time.Now())
		}
	} else if nodeStatus.GetPhase() == v1alpha1.NodePhaseRetryableFailure && nodeStatus.GetLastUpdatedAt() != nil {
		c.metrics.TransitionLatency.Observe(ctx, nodeStatus.GetLastUpdatedAt().Time, time.Now())
	}
}

func (c *nodeExecutor) IdempotentRecordEvent(ctx context.Context, nodeEvent *event.NodeExecutionEvent) error {
	if nodeEvent == nil {
		return fmt.Errorf("event recording attempt of Nil Node execution event")
	}

	if nodeEvent.Id == nil {
		return fmt.Errorf("event recording attempt of with nil node Event ID")
	}

	logger.Infof(ctx, "Recording event p[%+v]", nodeEvent)
	err := c.nodeRecorder.RecordNodeEvent(ctx, nodeEvent)
	if err != nil {
		if nodeEvent.GetId().NodeId == v1alpha1.EndNodeID {
			return nil
		}

		if eventsErr.IsAlreadyExists(err) {
			logger.Infof(ctx, "Node event phase: %s, nodeId %s already exist",
				nodeEvent.Phase.String(), nodeEvent.GetId().NodeId)
			return nil
		} else if eventsErr.IsEventAlreadyInTerminalStateError(err) {
			logger.Warningf(ctx, "Failed to record nodeEvent, error [%s]", err.Error())
			return errors.Wrapf(errors.IllegalStateError, nodeEvent.Id.NodeId, err, "phase mis-match mismatch between propeller and control plane; Trying to record Node p: %s", nodeEvent.Phase)
		}
	}
	return err
}

// In this method we check if the queue is ready to be processed and if so, we prime it in Admin as queued
// Before we start the node execution, we need to transition this Node status to Queued.
// This is because a node execution has to exist before task/wf executions can start.
func (c *nodeExecutor) preExecute(ctx context.Context, dag executors.DAGStructure, nCtx handler.NodeExecutionContext) (handler.PhaseInfo, error) {
	logger.Debugf(ctx, "Node not yet started")
	// Query the nodes information to figure out if it can be executed.
	predicatePhase, err := CanExecute(ctx, dag, nCtx.ContextualNodeLookup(), nCtx.Node())
	if err != nil {
		logger.Debugf(ctx, "Node failed in CanExecute. Error [%s]", err)
		return handler.PhaseInfoUndefined, err
	}

	if predicatePhase == PredicatePhaseReady {
		// TODO: Performance problem, we maybe in a retry loop and do not need to resolve the inputs again.
		// For now we will do this
		node := nCtx.Node()
		nodeStatus := nCtx.NodeStatus()
		dataDir := nodeStatus.GetDataDir()
		var nodeInputs *core.LiteralMap
		if !node.IsStartNode() {
			t := c.metrics.NodeInputGatherLatency.Start(ctx)
			defer t.Stop()
			// Can execute
			var err error
			nodeInputs, err = Resolve(ctx, c.outputResolver, nCtx.ContextualNodeLookup(), node.GetID(), node.GetInputBindings())
			// TODO we need to handle retryable, network errors here!!
			if err != nil {
				c.metrics.ResolutionFailure.Inc(ctx)
				logger.Warningf(ctx, "Failed to resolve inputs for Node. Error [%v]", err)
				return handler.PhaseInfoFailure(core.ExecutionError_SYSTEM, "BindingResolutionFailure", err.Error(), nil), nil
			}

			if nodeInputs != nil {
				inputsFile := v1alpha1.GetInputsFile(dataDir)
				if err := c.store.WriteProtobuf(ctx, inputsFile, storage.Options{}, nodeInputs); err != nil {
					c.metrics.InputsWriteFailure.Inc(ctx)
					logger.Errorf(ctx, "Failed to store inputs for Node. Error [%v]. InputsFile [%s]", err, inputsFile)
					return handler.PhaseInfoUndefined, errors.Wrapf(
						errors.StorageError, node.GetID(), err, "Failed to store inputs for Node. InputsFile [%s]", inputsFile)
				}
			}

			logger.Debugf(ctx, "Node Data Directory [%s].", nodeStatus.GetDataDir())
		}

		return handler.PhaseInfoQueued("node queued"), nil
	}

	// Now that we have resolved the inputs, we can record as a transition latency. This is because we have completed
	// all the overhead that we have to compute. Any failures after this will incur this penalty, but it could be due
	// to various external reasons - like queuing, overuse of quota, plugin overhead etc.
	logger.Debugf(ctx, "preExecute completed in phase [%s]", predicatePhase.String())
	if predicatePhase == PredicatePhaseSkip {
		return handler.PhaseInfoSkip(nil, "Node Skipped as parent node was skipped"), nil
	}

	return handler.PhaseInfoNotReady("predecessor node not yet complete"), nil
}

func isTimeoutExpired(queuedAt *metav1.Time, timeout time.Duration) bool {
	if !queuedAt.IsZero() && timeout != 0 {
		deadline := queuedAt.Add(timeout)
		if deadline.Before(time.Now()) {
			return true
		}
	}
	return false
}

func (c *nodeExecutor) isEligibleForRetry(nCtx *nodeExecContext, nodeStatus v1alpha1.ExecutableNodeStatus, err *core.ExecutionError) (currentAttempt, maxAttempts uint32, isEligible bool) {
	if err.Kind == core.ExecutionError_SYSTEM {
		currentAttempt = nodeStatus.GetSystemFailures()
		maxAttempts = c.maxNodeRetriesForSystemFailures
		isEligible = currentAttempt < c.maxNodeRetriesForSystemFailures
		return
	}

	currentAttempt = (nodeStatus.GetAttempts() + 1) - nodeStatus.GetSystemFailures()
	if nCtx.Node().GetRetryStrategy() != nil && nCtx.Node().GetRetryStrategy().MinAttempts != nil {
		maxAttempts = uint32(*nCtx.Node().GetRetryStrategy().MinAttempts)
	}
	isEligible = currentAttempt < maxAttempts
	return
}

func (c *nodeExecutor) execute(ctx context.Context, h handler.Node, nCtx *nodeExecContext, nodeStatus v1alpha1.ExecutableNodeStatus) (handler.PhaseInfo, error) {
	logger.Debugf(ctx, "Executing node")
	defer logger.Debugf(ctx, "Node execution round complete")

	t, err := h.Handle(ctx, nCtx)
	if err != nil {
		return handler.PhaseInfoUndefined, err
	}

	phase := t.Info()
	// check for timeout for non-terminal phases
	if !phase.GetPhase().IsTerminal() {
		activeDeadline := c.defaultActiveDeadline
		if nCtx.Node().GetActiveDeadline() != nil && *nCtx.Node().GetActiveDeadline() > 0 {
			activeDeadline = *nCtx.Node().GetActiveDeadline()
		}
		if isTimeoutExpired(nodeStatus.GetQueuedAt(), activeDeadline) {
			logger.Errorf(ctx, "Node has timed out; timeout configured: %v", activeDeadline)
			return handler.PhaseInfoTimedOut(nil, fmt.Sprintf("task active timeout [%s] expired", activeDeadline.String())), nil
		}

		// Execution timeout is a retry-able error
		executionDeadline := c.defaultExecutionDeadline
		if nCtx.Node().GetExecutionDeadline() != nil && *nCtx.Node().GetExecutionDeadline() > 0 {
			executionDeadline = *nCtx.Node().GetExecutionDeadline()
		}
		if isTimeoutExpired(nodeStatus.GetLastAttemptStartedAt(), executionDeadline) {
			logger.Errorf(ctx, "Current execution for the node timed out; timeout configured: %v", executionDeadline)
			executionErr := &core.ExecutionError{Code: "TimeoutExpired", Message: fmt.Sprintf("task execution timeout [%s] expired", executionDeadline.String()), Kind: core.ExecutionError_USER}
			phase = handler.PhaseInfoRetryableFailureErr(executionErr, nil)
		}
	}

	if phase.GetPhase() == handler.EPhaseRetryableFailure {
		currentAttempt, maxAttempts, isEligible := c.isEligibleForRetry(nCtx, nodeStatus, phase.GetErr())
		if !isEligible {
			return handler.PhaseInfoFailure(
				core.ExecutionError_USER,
				fmt.Sprintf("RetriesExhausted|%s", phase.GetErr().Code),
				fmt.Sprintf("[%d/%d] currentAttempt done. Last Error: %s::%s", currentAttempt, maxAttempts, phase.GetErr().Kind.String(), phase.GetErr().Message),
				phase.GetInfo(),
			), nil
		}

		// Retrying to clearing all status
		nCtx.nsm.clearNodeStatus()
	}

	return phase, nil
}

func (c *nodeExecutor) abort(ctx context.Context, h handler.Node, nCtx handler.NodeExecutionContext, reason string) error {
	logger.Debugf(ctx, "Calling aborting & finalize")
	if err := h.Abort(ctx, nCtx, reason); err != nil {
		finalizeErr := h.Finalize(ctx, nCtx)
		if finalizeErr != nil {
			return errors.ErrorCollection{Errors: []error{err, finalizeErr}}
		}
		return err
	}

	return h.Finalize(ctx, nCtx)
}

func (c *nodeExecutor) finalize(ctx context.Context, h handler.Node, nCtx handler.NodeExecutionContext) error {
	return h.Finalize(ctx, nCtx)
}

func (c *nodeExecutor) handleNotYetStartedNode(ctx context.Context, dag executors.DAGStructure, nCtx *nodeExecContext, _ handler.Node) (executors.NodeStatus, error) {
	logger.Debugf(ctx, "Node not yet started, running pre-execute")
	defer logger.Debugf(ctx, "Node pre-execute completed")
	p, err := c.preExecute(ctx, dag, nCtx)
	if err != nil {
		logger.Errorf(ctx, "failed preExecute for node. Error: %s", err.Error())
		return executors.NodeStatusUndefined, err
	}

	if p.GetPhase() == handler.EPhaseUndefined {
		return executors.NodeStatusUndefined, errors.Errorf(errors.IllegalStateError, nCtx.NodeID(), "received undefined phase.")
	}

	if p.GetPhase() == handler.EPhaseNotReady {
		return executors.NodeStatusPending, nil
	}

	np, err := ToNodePhase(p.GetPhase())
	if err != nil {
		return executors.NodeStatusUndefined, errors.Wrapf(errors.IllegalStateError, nCtx.NodeID(), err, "failed to move from queued")
	}

	nodeStatus := nCtx.NodeStatus()
	if np != nodeStatus.GetPhase() {
		// assert np == Queued!
		logger.Infof(ctx, "Change in node state detected from [%s] -> [%s]", nodeStatus.GetPhase().String(), np.String())
		nev, err := ToNodeExecutionEvent(nCtx.NodeExecutionMetadata().GetNodeExecutionID(),
			p, nCtx.InputReader(), nodeStatus, nCtx.ExecutionContext().GetEventVersion(),
			nCtx.ExecutionContext().GetParentInfo(), nCtx.node)
		if err != nil {
			return executors.NodeStatusUndefined, errors.Wrapf(errors.IllegalStateError, nCtx.NodeID(), err, "could not convert phase info to event")
		}
		err = c.IdempotentRecordEvent(ctx, nev)
		if err != nil {
			logger.Warningf(ctx, "Failed to record nodeEvent, error [%s]", err.Error())
			return executors.NodeStatusUndefined, errors.Wrapf(errors.EventRecordingFailed, nCtx.NodeID(), err, "failed to record node event")
		}
		UpdateNodeStatus(np, p, nCtx.nsm, nodeStatus)
		c.RecordTransitionLatency(ctx, dag, nCtx.ContextualNodeLookup(), nCtx.Node(), nodeStatus)
	}

	if np == v1alpha1.NodePhaseQueued {
		if nCtx.md.IsInterruptible() {
			c.metrics.InterruptibleNodesRunning.Inc(ctx)
		}
		return executors.NodeStatusQueued, nil
	} else if np == v1alpha1.NodePhaseSkipped {
		return executors.NodeStatusSuccess, nil
	}

	return executors.NodeStatusPending, nil
}

func (c *nodeExecutor) handleQueuedOrRunningNode(ctx context.Context, nCtx *nodeExecContext, h handler.Node) (executors.NodeStatus, error) {
	nodeStatus := nCtx.NodeStatus()
	currentPhase := nodeStatus.GetPhase()

	// case v1alpha1.NodePhaseQueued, v1alpha1.NodePhaseRunning:
	logger.Debugf(ctx, "node executing, current phase [%s]", currentPhase)
	defer logger.Debugf(ctx, "node execution completed")

	// Since we reset node status inside execute for retryable failure, we use lastAttemptStartTime to carry that information
	// across execute which is used to emit metrics
	lastAttemptStartTime := nodeStatus.GetLastAttemptStartedAt()

	p, err := c.execute(ctx, h, nCtx, nodeStatus)
	if err != nil {
		logger.Errorf(ctx, "failed Execute for node. Error: %s", err.Error())
		return executors.NodeStatusUndefined, err
	}

	if p.GetPhase() == handler.EPhaseUndefined {
		return executors.NodeStatusUndefined, errors.Errorf(errors.IllegalStateError, nCtx.NodeID(), "received undefined phase.")
	}

	np, err := ToNodePhase(p.GetPhase())
	if err != nil {
		return executors.NodeStatusUndefined, errors.Wrapf(errors.IllegalStateError, nCtx.NodeID(), err, "failed to move from queued")
	}

	// execErr in phase-info 'p' is only available if node has failed to execute, and the current phase at that time
	// will be v1alpha1.NodePhaseRunning
	execErr := p.GetErr()
	if execErr != nil && (currentPhase == v1alpha1.NodePhaseRunning || currentPhase == v1alpha1.NodePhaseQueued ||
		currentPhase == v1alpha1.NodePhaseDynamicRunning) {
		endTime := time.Now()
		startTime := endTime
		if lastAttemptStartTime != nil {
			startTime = lastAttemptStartTime.Time
		}

		if execErr.GetKind() == core.ExecutionError_SYSTEM {
			nodeStatus.IncrementSystemFailures()
			c.metrics.SystemErrorDuration.Observe(ctx, startTime, endTime)
		} else if execErr.GetKind() == core.ExecutionError_USER {
			c.metrics.UserErrorDuration.Observe(ctx, startTime, endTime)
		} else {
			c.metrics.UnknownErrorDuration.Observe(ctx, startTime, endTime)
		}
		// When a node fails, we fail the workflow. Independent of number of nodes succeeding/failing, whenever a first node fails,
		// the entire workflow is failed.
		if np == v1alpha1.NodePhaseFailing {
			if execErr.GetKind() == core.ExecutionError_SYSTEM {
				nodeStatus.IncrementSystemFailures()
				c.metrics.PermanentSystemErrorDuration.Observe(ctx, startTime, endTime)
			} else if execErr.GetKind() == core.ExecutionError_USER {
				c.metrics.PermanentUserErrorDuration.Observe(ctx, startTime, endTime)
			} else {
				c.metrics.PermanentUnknownErrorDuration.Observe(ctx, startTime, endTime)
			}
		}
	}
	finalStatus := executors.NodeStatusRunning
	if np == v1alpha1.NodePhaseFailing && !h.FinalizeRequired() {
		logger.Infof(ctx, "Finalize not required, moving node to Failed")
		np = v1alpha1.NodePhaseFailed
		finalStatus = executors.NodeStatusFailed(p.GetErr())
	}

	if np == v1alpha1.NodePhaseTimingOut && !h.FinalizeRequired() {
		logger.Infof(ctx, "Finalize not required, moving node to TimedOut")
		np = v1alpha1.NodePhaseTimedOut
		finalStatus = executors.NodeStatusTimedOut
	}

	if np == v1alpha1.NodePhaseSucceeding && !h.FinalizeRequired() {
		logger.Infof(ctx, "Finalize not required, moving node to Succeeded")
		np = v1alpha1.NodePhaseSucceeded
		finalStatus = executors.NodeStatusSuccess
	}

	// If it is retryable failure, we do no want to send any events, as the node is essentially still running
	// Similarly if the phase has not changed from the last time, events do not need to be sent
	if np != nodeStatus.GetPhase() && np != v1alpha1.NodePhaseRetryableFailure {
		// assert np == skipped, succeeding or failing
		logger.Infof(ctx, "Change in node state detected from [%s] -> [%s], (handler phase [%s])", nodeStatus.GetPhase().String(), np.String(), p.GetPhase().String())
		nev, err := ToNodeExecutionEvent(nCtx.NodeExecutionMetadata().GetNodeExecutionID(),
			p, nCtx.InputReader(), nCtx.NodeStatus(), nCtx.ExecutionContext().GetEventVersion(),
			nCtx.ExecutionContext().GetParentInfo(), nCtx.node)
		if err != nil {
			return executors.NodeStatusUndefined, errors.Wrapf(errors.IllegalStateError, nCtx.NodeID(), err, "could not convert phase info to event")
		}

		err = c.IdempotentRecordEvent(ctx, nev)
		if err != nil {
			logger.Warningf(ctx, "Failed to record nodeEvent, error [%s]", err.Error())
			return executors.NodeStatusUndefined, errors.Wrapf(errors.EventRecordingFailed, nCtx.NodeID(), err, "failed to record node event")
		}

		// We reach here only when transitioning from Queued to Running. In this case, the startedAt is not set.
		if np == v1alpha1.NodePhaseRunning {
			if nodeStatus.GetQueuedAt() != nil {
				c.metrics.QueuingLatency.Observe(ctx, nodeStatus.GetQueuedAt().Time, time.Now())
			}
		}
	}

	UpdateNodeStatus(np, p, nCtx.nsm, nodeStatus)
	return finalStatus, nil
}

func (c *nodeExecutor) handleRetryableFailure(ctx context.Context, nCtx *nodeExecContext, h handler.Node) (executors.NodeStatus, error) {
	nodeStatus := nCtx.NodeStatus()
	logger.Debugf(ctx, "node failed with retryable failure, aborting and finalizing, message: %s", nodeStatus.GetMessage())
	if err := c.abort(ctx, h, nCtx, nodeStatus.GetMessage()); err != nil {
		return executors.NodeStatusUndefined, err
	}

	// NOTE: It is important to increment attempts only after abort has been called. Increment attempt mutates the state
	// Attempt is used throughout the system to determine the idempotent resource version.
	nodeStatus.IncrementAttempts()
	nodeStatus.UpdatePhase(v1alpha1.NodePhaseRunning, v1.Now(), "retrying", nil)
	// We are going to retry in the next round, so we should clear all current state
	nodeStatus.ClearSubNodeStatus()
	nodeStatus.ClearTaskStatus()
	nodeStatus.ClearWorkflowStatus()
	nodeStatus.ClearDynamicNodeStatus()
	return executors.NodeStatusPending, nil
}

func (c *nodeExecutor) handleNode(ctx context.Context, dag executors.DAGStructure, nCtx *nodeExecContext, h handler.Node) (executors.NodeStatus, error) {
	logger.Debugf(ctx, "Handling Node [%s]", nCtx.NodeID())
	defer logger.Debugf(ctx, "Completed node [%s]", nCtx.NodeID())

	nodeStatus := nCtx.NodeStatus()
	currentPhase := nodeStatus.GetPhase()

	// Optimization!
	// If it is start node we directly move it to Queued without needing to run preExecute
	if currentPhase == v1alpha1.NodePhaseNotYetStarted && !nCtx.Node().IsStartNode() {
		return c.handleNotYetStartedNode(ctx, dag, nCtx, h)
	}

	if currentPhase == v1alpha1.NodePhaseFailing {
		logger.Debugf(ctx, "node failing")
		if err := c.finalize(ctx, h, nCtx); err != nil {
			return executors.NodeStatusUndefined, err
		}
		nodeStatus.UpdatePhase(v1alpha1.NodePhaseFailed, v1.Now(), nodeStatus.GetMessage(), nodeStatus.GetExecutionError())
		c.metrics.FailureDuration.Observe(ctx, nodeStatus.GetStartedAt().Time, nodeStatus.GetStoppedAt().Time)
		if nCtx.md.IsInterruptible() {
			c.metrics.InterruptibleNodesTerminated.Inc(ctx)
		}
		return executors.NodeStatusFailed(nodeStatus.GetExecutionError()), nil
	}

	if currentPhase == v1alpha1.NodePhaseTimingOut {
		logger.Debugf(ctx, "node timing out")
		if err := c.abort(ctx, h, nCtx, "node timed out"); err != nil {
			return executors.NodeStatusUndefined, err
		}

		nodeStatus.ClearSubNodeStatus()
		nodeStatus.UpdatePhase(v1alpha1.NodePhaseTimedOut, v1.Now(), nodeStatus.GetMessage(), nodeStatus.GetExecutionError())
		c.metrics.TimedOutFailure.Inc(ctx)
		if nCtx.md.IsInterruptible() {
			c.metrics.InterruptibleNodesTerminated.Inc(ctx)
		}
		return executors.NodeStatusTimedOut, nil
	}

	if currentPhase == v1alpha1.NodePhaseSucceeding {
		logger.Debugf(ctx, "node succeeding")
		if err := c.finalize(ctx, h, nCtx); err != nil {
			return executors.NodeStatusUndefined, err
		}

		nodeStatus.ClearSubNodeStatus()
		nodeStatus.UpdatePhase(v1alpha1.NodePhaseSucceeded, v1.Now(), "completed successfully", nil)
		c.metrics.SuccessDuration.Observe(ctx, nodeStatus.GetStartedAt().Time, nodeStatus.GetStoppedAt().Time)
		if nCtx.md.IsInterruptible() {
			c.metrics.InterruptibleNodesTerminated.Inc(ctx)
		}
		return executors.NodeStatusSuccess, nil
	}

	if currentPhase == v1alpha1.NodePhaseRetryableFailure {
		return c.handleRetryableFailure(ctx, nCtx, h)
	}

	if currentPhase == v1alpha1.NodePhaseFailed {
		// This should never happen
		return executors.NodeStatusFailed(nodeStatus.GetExecutionError()), nil
	}

	return c.handleQueuedOrRunningNode(ctx, nCtx, h)
}

// The space search for the next node to execute is implemented like a DFS algorithm. handleDownstream visits all the nodes downstream from
// the currentNode. Visit a node is the RecursiveNodeHandler. A visit may be partial, complete or may result in a failure.
func (c *nodeExecutor) handleDownstream(ctx context.Context, execContext executors.ExecutionContext, dag executors.DAGStructure, nl executors.NodeLookup, currentNode v1alpha1.ExecutableNode) (executors.NodeStatus, error) {
	logger.Debugf(ctx, "Handling downstream Nodes")
	// This node is success. Handle all downstream nodes
	downstreamNodes, err := dag.FromNode(currentNode.GetID())
	if err != nil {
		logger.Debugf(ctx, "Error when retrieving downstream nodes, [%s]", err)
		return executors.NodeStatusFailed(&core.ExecutionError{
			Code:    errors.BadSpecificationError,
			Message: fmt.Sprintf("failed to retrieve downstream nodes for [%s]", currentNode.GetID()),
			Kind:    core.ExecutionError_SYSTEM,
		}), nil
	}
	if len(downstreamNodes) == 0 {
		logger.Debugf(ctx, "No downstream nodes found. Complete.")
		return executors.NodeStatusComplete, nil
	}
	// If any downstream node is failed, fail, all
	// Else if all are success then success
	// Else if any one is running then Downstream is still running
	allCompleted := true
	partialNodeCompletion := false
	onFailurePolicy := execContext.GetOnFailurePolicy()
	stateOnComplete := executors.NodeStatusComplete
	for _, downstreamNodeName := range downstreamNodes {
		downstreamNode, ok := nl.GetNode(downstreamNodeName)
		if !ok {
			return executors.NodeStatusFailed(&core.ExecutionError{
				Code:    errors.BadSpecificationError,
				Message: fmt.Sprintf("failed to retrieve downstream node [%s] for [%s]", downstreamNodeName, currentNode.GetID()),
				Kind:    core.ExecutionError_SYSTEM,
			}), nil
		}

		state, err := c.RecursiveNodeHandler(ctx, execContext, dag, nl, downstreamNode)
		if err != nil {
			return executors.NodeStatusUndefined, err
		}

		if state.HasFailed() || state.HasTimedOut() {
			logger.Debugf(ctx, "Some downstream node has failed. Failed: [%v]. TimedOut: [%v]. Error: [%s]", state.HasFailed(), state.HasTimedOut(), state.Err)
			if onFailurePolicy == v1alpha1.WorkflowOnFailurePolicy(core.WorkflowMetadata_FAIL_AFTER_EXECUTABLE_NODES_COMPLETE) {
				// If the failure policy allows other nodes to continue running, do not exit the loop,
				// Keep track of the last failed state in the loop since it'll be the one to return.
				// TODO: If multiple nodes fail (which this mode allows), consolidate/summarize failure states in one.
				stateOnComplete = state
			} else {
				return state, nil
			}
		} else if !state.IsComplete() {
			// A Failed/Timedout node is implicitly considered "complete" this means none of the downstream nodes from
			// that node will ever be allowed to run.
			// This else block, therefore, deals with all other states. IsComplete will return true if and only if this
			// node as well as all of its downstream nodes have finished executing with success statuses. Otherwise we
			// mark this node's state as not completed to ensure we will visit it again later.
			allCompleted = false
		}

		if state.PartiallyComplete() {
			// This implies that one of the downstream nodes has just succeeded and workflow is ready for propagation
			// We do not propagate in current cycle to make it possible to store the state between transitions
			partialNodeCompletion = true
		}
	}

	if allCompleted {
		logger.Debugf(ctx, "All downstream nodes completed")
		return stateOnComplete, nil
	}

	if partialNodeCompletion {
		return executors.NodeStatusSuccess, nil
	}

	return executors.NodeStatusPending, nil
}

func (c *nodeExecutor) SetInputsForStartNode(ctx context.Context, execContext executors.ExecutionContext, dag executors.DAGStructureWithStartNode, nl executors.NodeLookup, inputs *core.LiteralMap) (executors.NodeStatus, error) {
	startNode := dag.StartNode()
	ctx = contextutils.WithNodeID(ctx, startNode.GetID())
	if inputs == nil {
		logger.Infof(ctx, "No inputs for the workflow. Skipping storing inputs")
		return executors.NodeStatusComplete, nil
	}

	// StartNode is special. It does not have any processing step. It just takes the workflow (or subworkflow) inputs and converts to its own outputs
	nodeStatus := nl.GetNodeExecutionStatus(ctx, startNode.GetID())

	if len(nodeStatus.GetDataDir()) == 0 {
		return executors.NodeStatusUndefined, errors.Errorf(errors.IllegalStateError, startNode.GetID(), "no data-dir set, cannot store inputs")
	}
	outputFile := v1alpha1.GetOutputsFile(nodeStatus.GetOutputDir())

	so := storage.Options{}
	if err := c.store.WriteProtobuf(ctx, outputFile, so, inputs); err != nil {
		logger.Errorf(ctx, "Failed to write protobuf (metadata). Error [%v]", err)
		return executors.NodeStatusUndefined, errors.Wrapf(errors.CausedByError, startNode.GetID(), err, "Failed to store workflow inputs (as start node)")
	}

	return executors.NodeStatusComplete, nil
}

func canHandleNode(phase v1alpha1.NodePhase) bool {
	return phase == v1alpha1.NodePhaseNotYetStarted ||
		phase == v1alpha1.NodePhaseQueued ||
		phase == v1alpha1.NodePhaseRunning ||
		phase == v1alpha1.NodePhaseFailing ||
		phase == v1alpha1.NodePhaseTimingOut ||
		phase == v1alpha1.NodePhaseRetryableFailure ||
		phase == v1alpha1.NodePhaseSucceeding ||
		phase == v1alpha1.NodePhaseDynamicRunning
}

func (c *nodeExecutor) RecursiveNodeHandler(ctx context.Context, execContext executors.ExecutionContext,
	dag executors.DAGStructure, nl executors.NodeLookup, currentNode v1alpha1.ExecutableNode) (
	executors.NodeStatus, error) {

	currentNodeCtx := contextutils.WithNodeID(ctx, currentNode.GetID())
	nodeStatus := nl.GetNodeExecutionStatus(ctx, currentNode.GetID())
	nodePhase := nodeStatus.GetPhase()

	if canHandleNode(nodePhase) {
		// TODO Follow up Pull Request,
		// 1. Rename this method to DAGTraversalHandleNode (accepts a DAGStructure along-with) the remaining arguments
		// 2. Create a new method called HandleNode (part of the interface) (remaining all args as the previous method, but no DAGStructure
		// 3. Additional both methods will receive inputs reader
		// 4. The Downstream nodes handler will Resolve the Inputs
		// 5. the method will delegate all other node handling to HandleNode.
		// 6. Thus we can get rid of SetInputs for StartNode as well
		logger.Debugf(currentNodeCtx, "Handling node Status [%v]", nodeStatus.GetPhase().String())

		t := c.metrics.NodeExecutionTime.Start(ctx)
		defer t.Stop()

		// This is an optimization to avoid creating the nodeContext object in case the node has already been looked at.
		// If the overhead was zero, we would just do the isDirtyCheck after the nodeContext is created
		if nodeStatus.IsDirty() {
			return executors.NodeStatusRunning, nil
		}

		// Now if the node is of type task, then let us check if we are within the parallelism limit, only if the node
		// has been queued already
		if currentNode.GetKind() == v1alpha1.NodeKindTask && nodeStatus.GetPhase() == v1alpha1.NodePhaseQueued {
			maxParallelism := execContext.GetExecutionConfig().MaxParallelism
			if maxParallelism > 0 {
				// If we are queued, let us see if we can proceed within the node parallelism bounds
				if execContext.CurrentParallelism() >= maxParallelism {
					logger.Infof(ctx, "Maximum Parallelism for task nodes achieved [%d] >= Max [%d], Round will be short-circuited.", execContext.CurrentParallelism(), maxParallelism)
					return executors.NodeStatusRunning, nil
				}
				// We know that Propeller goes through each workflow in a single thread, thus every node is really processed
				// sequentially. So, we can continue - now that we know we are under the parallelism limits and increment the
				// parallelism if the node, enters a running state
				logger.Debugf(ctx, "Parallelism criteria not met, Current [%d], Max [%d]", execContext.CurrentParallelism(), maxParallelism)
			} else {
				logger.Debugf(ctx, "Parallelism control disabled")
			}
		} else {
			logger.Debugf(ctx, "NodeKind: %s in status [%s]. Parallelism control is not applicable. Current Parallelism [%d]",
				currentNode.GetKind().String(), nodeStatus.GetPhase().String(), execContext.CurrentParallelism())
		}

		nCtx, err := c.newNodeExecContextDefault(ctx, currentNode.GetID(), execContext, nl)
		if err != nil {
			// NodeExecution creation failure is a permanent fail / system error.
			// Should a system failure always return an err?
			return executors.NodeStatusFailed(&core.ExecutionError{
				Code:    "InternalError",
				Message: err.Error(),
				Kind:    core.ExecutionError_SYSTEM,
			}), nil
		}

		// Now depending on the node type decide
		h, err := c.nodeHandlerFactory.GetHandler(nCtx.Node().GetKind())
		if err != nil {
			return executors.NodeStatusUndefined, err
		}

		return c.handleNode(currentNodeCtx, dag, nCtx, h)

		// TODO we can optimize skip state handling by iterating down the graph and marking all as skipped
		// Currently we treat either Skip or Success the same way. In this approach only one node will be skipped
		// at a time. As we iterate down, further nodes will be skipped
	} else if nodePhase == v1alpha1.NodePhaseSucceeded || nodePhase == v1alpha1.NodePhaseSkipped {
		logger.Debugf(currentNodeCtx, "Node has [%v], traversing downstream.", nodePhase)
		return c.handleDownstream(ctx, execContext, dag, nl, currentNode)
	} else if nodePhase == v1alpha1.NodePhaseFailed {
		logger.Debugf(currentNodeCtx, "Node has failed, traversing downstream.")
		_, err := c.handleDownstream(ctx, execContext, dag, nl, currentNode)
		if err != nil {
			return executors.NodeStatusUndefined, err
		}

		return executors.NodeStatusFailed(nodeStatus.GetExecutionError()), nil
	} else if nodePhase == v1alpha1.NodePhaseTimedOut {
		logger.Debugf(currentNodeCtx, "Node has timed out, traversing downstream.")
		_, err := c.handleDownstream(ctx, execContext, dag, nl, currentNode)
		if err != nil {
			return executors.NodeStatusUndefined, err
		}

		return executors.NodeStatusTimedOut, nil
	}

	return executors.NodeStatusUndefined, errors.Errorf(errors.IllegalStateError, currentNode.GetID(),
		"Should never reach here. Current Phase: %v", nodePhase)
}

func (c *nodeExecutor) FinalizeHandler(ctx context.Context, execContext executors.ExecutionContext, dag executors.DAGStructure, nl executors.NodeLookup, currentNode v1alpha1.ExecutableNode) error {
	nodeStatus := nl.GetNodeExecutionStatus(ctx, currentNode.GetID())
	nodePhase := nodeStatus.GetPhase()

	if nodePhase == v1alpha1.NodePhaseNotYetStarted {
		logger.Infof(ctx, "Node not yet started, will not finalize")
		// Nothing to be aborted
		return nil
	}

	if canHandleNode(nodePhase) {
		ctx = contextutils.WithNodeID(ctx, currentNode.GetID())

		// Now depending on the node type decide
		h, err := c.nodeHandlerFactory.GetHandler(currentNode.GetKind())
		if err != nil {
			return err
		}

		nCtx, err := c.newNodeExecContextDefault(ctx, currentNode.GetID(), execContext, nl)
		if err != nil {
			return err
		}
		// Abort this node
		err = c.finalize(ctx, h, nCtx)
		if err != nil {
			return err
		}
	} else {
		// Abort downstream nodes
		downstreamNodes, err := dag.FromNode(currentNode.GetID())
		if err != nil {
			logger.Debugf(ctx, "Error when retrieving downstream nodes. Error [%v]", err)
			return nil
		}

		errs := make([]error, 0, len(downstreamNodes))
		for _, d := range downstreamNodes {
			downstreamNode, ok := nl.GetNode(d)
			if !ok {
				return errors.Errorf(errors.BadSpecificationError, currentNode.GetID(), "Unable to find Downstream Node [%v]", d)
			}

			if err := c.FinalizeHandler(ctx, execContext, dag, nl, downstreamNode); err != nil {
				logger.Infof(ctx, "Failed to abort node [%v]. Error: %v", d, err)
				errs = append(errs, err)
			}
		}

		if len(errs) > 0 {
			return errors.ErrorCollection{Errors: errs}
		}

		return nil
	}

	return nil
}

func (c *nodeExecutor) AbortHandler(ctx context.Context, execContext executors.ExecutionContext, dag executors.DAGStructure, nl executors.NodeLookup, currentNode v1alpha1.ExecutableNode, reason string) error {
	nodeStatus := nl.GetNodeExecutionStatus(ctx, currentNode.GetID())
	nodePhase := nodeStatus.GetPhase()

	if nodePhase == v1alpha1.NodePhaseNotYetStarted {
		logger.Infof(ctx, "Node not yet started, will not finalize")
		// Nothing to be aborted
		return nil
	}

	if canHandleNode(nodePhase) {
		ctx = contextutils.WithNodeID(ctx, currentNode.GetID())

		// Now depending on the node type decide
		h, err := c.nodeHandlerFactory.GetHandler(currentNode.GetKind())
		if err != nil {
			return err
		}

		nCtx, err := c.newNodeExecContextDefault(ctx, currentNode.GetID(), execContext, nl)
		if err != nil {
			return err
		}
		// Abort this node
		err = c.abort(ctx, h, nCtx, reason)
		if err != nil {
			return err
		}
		nodeExecutionID := &core.NodeExecutionIdentifier{
			ExecutionId: nCtx.NodeExecutionMetadata().GetNodeExecutionID().ExecutionId,
			NodeId:      nCtx.NodeExecutionMetadata().GetNodeExecutionID().NodeId,
		}
		if nCtx.ExecutionContext().GetEventVersion() != v1alpha1.EventVersion0 {
			currentNodeUniqueID, err := common.GenerateUniqueID(nCtx.ExecutionContext().GetParentInfo(), nodeExecutionID.NodeId)
			if err != nil {
				return err
			}
			nodeExecutionID.NodeId = currentNodeUniqueID
		}

		err = c.IdempotentRecordEvent(ctx, &event.NodeExecutionEvent{
			Id:         nodeExecutionID,
			Phase:      core.NodeExecution_ABORTED,
			OccurredAt: ptypes.TimestampNow(),
			OutputResult: &event.NodeExecutionEvent_Error{
				Error: &core.ExecutionError{
					Code:    "NodeAborted",
					Message: reason,
				},
			},
		})
		if err != nil {
			if errors2.IsCausedBy(err, errors.IllegalStateError) {
				logger.Debugf(ctx, "Failed to record abort event due to illegal state transition. Ignoring the error. Error: %v", err)
			} else {
				logger.Warningf(ctx, "Failed to record nodeEvent, error [%s]", err.Error())
				return errors.Wrapf(errors.EventRecordingFailed, nCtx.NodeID(), err, "failed to record node event")
			}
		}
	} else if nodePhase == v1alpha1.NodePhaseSucceeded || nodePhase == v1alpha1.NodePhaseSkipped {
		// Abort downstream nodes
		downstreamNodes, err := dag.FromNode(currentNode.GetID())
		if err != nil {
			logger.Debugf(ctx, "Error when retrieving downstream nodes. Error [%v]", err)
			return nil
		}

		errs := make([]error, 0, len(downstreamNodes))
		for _, d := range downstreamNodes {
			downstreamNode, ok := nl.GetNode(d)
			if !ok {
				return errors.Errorf(errors.BadSpecificationError, currentNode.GetID(), "Unable to find Downstream Node [%v]", d)
			}

			if err := c.AbortHandler(ctx, execContext, dag, nl, downstreamNode, reason); err != nil {
				logger.Infof(ctx, "Failed to abort node [%v]. Error: %v", d, err)
				errs = append(errs, err)
			}
		}

		if len(errs) > 0 {
			return errors.ErrorCollection{Errors: errs}
		}

		return nil
	} else {
		ctx = contextutils.WithNodeID(ctx, currentNode.GetID())
		logger.Warnf(ctx, "Trying to abort a node in state [%s]", nodeStatus.GetPhase().String())
	}

	return nil
}

func (c *nodeExecutor) Initialize(ctx context.Context) error {
	logger.Infof(ctx, "Initializing Core Node Executor")
	s := c.newSetupContext(ctx)
	return c.nodeHandlerFactory.Setup(ctx, s)
}

func NewExecutor(ctx context.Context, nodeConfig config.NodeConfig, store *storage.DataStore, enQWorkflow v1alpha1.EnqueueWorkflow, eventSink events.EventSink,
	workflowLauncher launchplan.Executor, launchPlanReader launchplan.Reader, maxDatasetSize int64,
	defaultRawOutputPrefix storage.DataReference, kubeClient executors.Client,
	catalogClient catalog.Client, scope promutils.Scope) (executors.Node, error) {

	// TODO we may want to make this configurable.
	shardSelector, err := ioutils.NewBase36PrefixShardSelector(ctx)
	if err != nil {
		return nil, err
	}

	nodeScope := scope.NewSubScope("node")
	exec := &nodeExecutor{
		store:               store,
		enqueueWorkflow:     enQWorkflow,
		nodeRecorder:        events.NewNodeEventRecorder(eventSink, nodeScope),
		taskRecorder:        events.NewTaskEventRecorder(eventSink, scope.NewSubScope("task")),
		maxDatasetSizeBytes: maxDatasetSize,
		metrics: &nodeMetrics{
			Scope:                         nodeScope,
			FailureDuration:               labeled.NewStopWatch("failure_duration", "Indicates the total execution time of a failed workflow.", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			SuccessDuration:               labeled.NewStopWatch("success_duration", "Indicates the total execution time of a successful workflow.", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			UserErrorDuration:             labeled.NewStopWatch("user_error_duration", "Indicates the total execution time before user error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			SystemErrorDuration:           labeled.NewStopWatch("system_error_duration", "Indicates the total execution time before system error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			UnknownErrorDuration:          labeled.NewStopWatch("unknown_error_duration", "Indicates the total execution time before unknown error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			PermanentUserErrorDuration:    labeled.NewStopWatch("perma_user_error_duration", "Indicates the total execution time before non recoverable user error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			PermanentSystemErrorDuration:  labeled.NewStopWatch("perma_system_error_duration", "Indicates the total execution time before non recoverable system error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			PermanentUnknownErrorDuration: labeled.NewStopWatch("perma_unknown_error_duration", "Indicates the total execution time before non recoverable unknown error", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			InputsWriteFailure:            labeled.NewCounter("inputs_write_fail", "Indicates failure in writing node inputs to metastore", nodeScope),
			TimedOutFailure:               labeled.NewCounter("timeout_fail", "Indicates failure due to timeout", nodeScope),
			InterruptedThresholdHit:       labeled.NewCounter("interrupted_threshold", "Indicates the node interruptible disabled because it hit max failure count", nodeScope),
			InterruptibleNodesRunning:     labeled.NewCounter("interruptible_nodes_running", "number of interruptible nodes running", nodeScope),
			InterruptibleNodesTerminated:  labeled.NewCounter("interruptible_nodes_terminated", "number of interruptible nodes finished running", nodeScope),
			ResolutionFailure:             labeled.NewCounter("input_resolve_fail", "Indicates failure in resolving node inputs", nodeScope),
			TransitionLatency:             labeled.NewStopWatch("transition_latency", "Measures the latency between the last parent node stoppedAt time and current node's queued time.", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			QueuingLatency:                labeled.NewStopWatch("queueing_latency", "Measures the latency between the time a node's been queued to the time the handler reported the executable moved to running state", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
			NodeExecutionTime:             labeled.NewStopWatch("node_exec_latency", "Measures the time taken to execute one node, a node can be complex so it may encompass sub-node latency.", time.Microsecond, nodeScope, labeled.EmitUnlabeledMetric),
			NodeInputGatherLatency:        labeled.NewStopWatch("node_input_latency", "Measures the latency to aggregate inputs and check readiness of a node", time.Millisecond, nodeScope, labeled.EmitUnlabeledMetric),
		},
		outputResolver:                  NewRemoteFileOutputResolver(store),
		defaultExecutionDeadline:        nodeConfig.DefaultDeadlines.DefaultNodeExecutionDeadline.Duration,
		defaultActiveDeadline:           nodeConfig.DefaultDeadlines.DefaultNodeActiveDeadline.Duration,
		maxNodeRetriesForSystemFailures: uint32(nodeConfig.MaxNodeRetriesOnSystemFailures),
		interruptibleFailureThreshold:   uint32(nodeConfig.InterruptibleFailureThreshold),
		defaultDataSandbox:              defaultRawOutputPrefix,
		shardSelector:                   shardSelector,
	}
	nodeHandlerFactory, err := NewHandlerFactory(ctx, exec, workflowLauncher, launchPlanReader, kubeClient, catalogClient, nodeScope)
	exec.nodeHandlerFactory = nodeHandlerFactory
	return exec, err
}

func init() {
	labeled.SetMetricKeys(contextutils.ProjectKey, contextutils.DomainKey, contextutils.WorkflowIDKey,
		contextutils.TaskIDKey)
}