open-nomad/nomad/deploymentwatcher/deployment_watcher.go

// Copyright (c) HashiCorp, Inc.
// SPDX-License-Identifier: MPL-2.0

package deploymentwatcher

import (
	"context"
	"fmt"
	"sync"
	"time"

	log "github.com/hashicorp/go-hclog"
	memdb "github.com/hashicorp/go-memdb"
	"github.com/hashicorp/nomad/helper/pointer"
	"github.com/hashicorp/nomad/helper/uuid"
	"github.com/hashicorp/nomad/nomad/state"
	"github.com/hashicorp/nomad/nomad/structs"
	"golang.org/x/time/rate"
)

const (
	// perJobEvalBatchPeriod is the batching length before creating an evaluation to
	// trigger the scheduler when allocations are marked as healthy.
	perJobEvalBatchPeriod = 1 * time.Second
)

var (
	// allowRescheduleTransition is the transition that allows failed
	// allocations part of a deployment to be rescheduled. We create a one off
	// variable to avoid creating a new object for every request.
	allowRescheduleTransition = &structs.DesiredTransition{
		Reschedule: pointer.Of(true),
	}
)

// deploymentTriggers are the set of functions required to trigger changes on
// behalf of a deployment
type deploymentTriggers interface {
	// createUpdate is used to create allocation desired transition updates and
	// an evaluation.
	createUpdate(allocs map[string]*structs.DesiredTransition, eval *structs.Evaluation) (uint64, error)

	// upsertJob is used to roll back a job when autoreverting for a deployment
	upsertJob(job *structs.Job) (uint64, error)

	// upsertDeploymentStatusUpdate is used to upsert a deployment status update
	// and an optional evaluation and job to upsert
	upsertDeploymentStatusUpdate(u *structs.DeploymentStatusUpdate, eval *structs.Evaluation, job *structs.Job) (uint64, error)

	// upsertDeploymentPromotion is used to promote canaries in a deployment
	upsertDeploymentPromotion(req *structs.ApplyDeploymentPromoteRequest) (uint64, error)

	// upsertDeploymentAllocHealth is used to set the health of allocations in a
	// deployment
	upsertDeploymentAllocHealth(req *structs.ApplyDeploymentAllocHealthRequest) (uint64, error)
}

// deploymentWatcher is used to watch a single deployment and trigger the
// scheduler when allocation health transitions.
type deploymentWatcher struct {
	// queryLimiter is used to limit the rate of blocking queries
	queryLimiter *rate.Limiter

	// deploymentTriggers holds the methods required to trigger changes on behalf of the
	// deployment
	deploymentTriggers

	// DeploymentRPC holds methods for interacting with peer regions
	// in enterprise edition
	DeploymentRPC

	// JobRPC holds methods for interacting with peer regions
	// in enterprise edition
	JobRPC

	// state is the state that is watched for state changes.
	state *state.StateStore

	// deploymentID is the deployment's ID being watched
	deploymentID string

	// deploymentUpdateCh is triggered when there is an updated deployment
	deploymentUpdateCh chan struct{}

	// d is the deployment being watched
	d *structs.Deployment

	// j is the job the deployment is for
	j *structs.Job

	// outstandingBatch marks whether an outstanding function exists to create
	// the evaluation. Access should be done through the lock.
	outstandingBatch bool

	// outstandingAllowReplacements is the map of allocations that will be
	// marked as allowing a replacement. Access should be done through the lock.
	outstandingAllowReplacements map[string]*structs.DesiredTransition

	// latestEval is the latest eval for the job. It is updated by the watch
	// loop and any time an evaluation is created. The field should be accessed
	// by holding the lock or using the setter and getter methods.
	latestEval uint64

	logger log.Logger
	ctx    context.Context
	exitFn context.CancelFunc
	l      sync.RWMutex
}

// newDeploymentWatcher returns a deployment watcher that is used to watch
// deployments and trigger the scheduler as needed.
func newDeploymentWatcher(parent context.Context, queryLimiter *rate.Limiter,
	logger log.Logger, state *state.StateStore, d *structs.Deployment,
	j *structs.Job, triggers deploymentTriggers,
	deploymentRPC DeploymentRPC, jobRPC JobRPC) *deploymentWatcher {

	ctx, exitFn := context.WithCancel(parent)
	w := &deploymentWatcher{
		queryLimiter:       queryLimiter,
		deploymentID:       d.ID,
		deploymentUpdateCh: make(chan struct{}, 1),
		d:                  d,
		j:                  j,
		state:              state,
		deploymentTriggers: triggers,
		DeploymentRPC:      deploymentRPC,
		JobRPC:             jobRPC,
		logger:             logger.With("deployment_id", d.ID, "job", j.NamespacedID()),
		ctx:                ctx,
		exitFn:             exitFn,
	}

	// Start the long lived watcher that scans for allocation updates
	go w.watch()

	return w
}

// updateDeployment is used to update the tracked deployment.
func (w *deploymentWatcher) updateDeployment(d *structs.Deployment) {
	w.l.Lock()
	defer w.l.Unlock()

	// Update and trigger
	w.d = d
	select {
	case w.deploymentUpdateCh <- struct{}{}:
	default:
	}
}

// getDeployment returns the tracked deployment.
func (w *deploymentWatcher) getDeployment() *structs.Deployment {
	w.l.RLock()
	defer w.l.RUnlock()
	return w.d
}

func (w *deploymentWatcher) SetAllocHealth(
	req *structs.DeploymentAllocHealthRequest,
	resp *structs.DeploymentUpdateResponse) error {

	// If we are failing the deployment, update the status and potentially
	// rollback
	var j *structs.Job
	var u *structs.DeploymentStatusUpdate

	// If there are unhealthy allocations we need to mark the deployment as
	// failed and check if we should roll back to a stable job.
	if l := len(req.UnhealthyAllocationIDs); l != 0 {
		unhealthy := make(map[string]struct{}, l)
		for _, alloc := range req.UnhealthyAllocationIDs {
			unhealthy[alloc] = struct{}{}
		}

		// Get the allocations for the deployment
		snap, err := w.state.Snapshot()
		if err != nil {
			return err
		}

		allocs, err := snap.AllocsByDeployment(nil, req.DeploymentID)
		if err != nil {
			return err
		}

		// Determine if we should autorevert to an older job
		desc := structs.DeploymentStatusDescriptionFailedAllocations
		for _, alloc := range allocs {
			// Check that the alloc has been marked unhealthy
			if _, ok := unhealthy[alloc.ID]; !ok {
				continue
			}

			// Check if the group has autorevert set
			dstate, ok := w.getDeployment().TaskGroups[alloc.TaskGroup]
			if !ok || !dstate.AutoRevert {
				continue
			}

			var err error
			j, err = w.latestStableJob()
			if err != nil {
				return err
			}

			if j != nil {
				j, desc = w.handleRollbackValidity(j, desc)
			}
			break
		}

		u = w.getDeploymentStatusUpdate(structs.DeploymentStatusFailed, desc)
	}

	// Canonicalize the job in case it doesn't have namespace set
	j.Canonicalize()

	// Create the request
	areq := &structs.ApplyDeploymentAllocHealthRequest{
		DeploymentAllocHealthRequest: *req,
		Timestamp:                    time.Now(),
		Eval:                         w.getEval(),
		DeploymentUpdate:             u,
		Job:                          j,
	}

	index, err := w.upsertDeploymentAllocHealth(areq)
	if err != nil {
		return err
	}

	// Build the response
	resp.EvalID = areq.Eval.ID
	resp.EvalCreateIndex = index
	resp.DeploymentModifyIndex = index
	resp.Index = index
	if j != nil {
		resp.RevertedJobVersion = pointer.Of(j.Version)
	}
	return nil
}

// handleRollbackValidity checks if the job being rolled back to has the same spec as the existing job
// Returns a modified description and job accordingly.
func (w *deploymentWatcher) handleRollbackValidity(rollbackJob *structs.Job, desc string) (*structs.Job, string) {
	// Only rollback if job being changed has a different spec.
	// This prevents an infinite revert cycle when a previously stable version of the job fails to start up during a rollback
	// If the job we are trying to rollback to is identical to the current job, we stop because the rollback will not succeed.
	if w.j.SpecChanged(rollbackJob) {
		desc = structs.DeploymentStatusDescriptionRollback(desc, rollbackJob.Version)
	} else {
		desc = structs.DeploymentStatusDescriptionRollbackNoop(desc, rollbackJob.Version)
		rollbackJob = nil
	}
	return rollbackJob, desc
}

func (w *deploymentWatcher) PromoteDeployment(
	req *structs.DeploymentPromoteRequest,
	resp *structs.DeploymentUpdateResponse) error {

	// Create the request
	areq := &structs.ApplyDeploymentPromoteRequest{
		DeploymentPromoteRequest: *req,
		Eval:                     w.getEval(),
	}

	index, err := w.upsertDeploymentPromotion(areq)
	if err != nil {
		return err
	}

	// Build the response
	resp.EvalID = areq.Eval.ID
	resp.EvalCreateIndex = index
	resp.DeploymentModifyIndex = index
	resp.Index = index
	return nil
}

// autoPromoteDeployment creates a synthetic promotion request, and upserts it for processing
func (w *deploymentWatcher) autoPromoteDeployment(allocs []*structs.AllocListStub) error {
	d := w.getDeployment()
	if !d.HasPlacedCanaries() || !d.RequiresPromotion() {
		return nil
	}

	// AutoPromote iff every task group with canaries is marked auto_promote and is healthy. The whole
	// job version has been incremented, so we promote together. See also AutoRevert
	for _, dstate := range d.TaskGroups {

		// skip auto promote canary validation if the task group has no canaries
		// to prevent auto promote hanging on mixed canary/non-canary taskgroup deploys
		if dstate.DesiredCanaries < 1 {
			continue
		}

		if !dstate.AutoPromote || len(dstate.PlacedCanaries) < dstate.DesiredCanaries {
			return nil
		}

		healthyCanaries := 0
		// Find the health status of each canary
		for _, c := range dstate.PlacedCanaries {
			for _, a := range allocs {
				if c == a.ID && a.DeploymentStatus.IsHealthy() {
					healthyCanaries += 1
				}
			}
		}
		if healthyCanaries != dstate.DesiredCanaries {
			return nil
		}
	}

	// Send the request
	_, err := w.upsertDeploymentPromotion(&structs.ApplyDeploymentPromoteRequest{
		DeploymentPromoteRequest: structs.DeploymentPromoteRequest{DeploymentID: d.GetID(), All: true},
		Eval:                     w.getEval(),
	})
	return err
}

func (w *deploymentWatcher) PauseDeployment(
	req *structs.DeploymentPauseRequest,
	resp *structs.DeploymentUpdateResponse) error {
	// Determine the status we should transition to and if we need to create an
	// evaluation
	status, desc := structs.DeploymentStatusPaused, structs.DeploymentStatusDescriptionPaused
	var eval *structs.Evaluation
	evalID := ""
	if !req.Pause {
		status, desc = structs.DeploymentStatusRunning, structs.DeploymentStatusDescriptionRunning
		eval = w.getEval()
		evalID = eval.ID
	}
	update := w.getDeploymentStatusUpdate(status, desc)

	// Commit the change
	i, err := w.upsertDeploymentStatusUpdate(update, eval, nil)
	if err != nil {
		return err
	}

	// Build the response
	if evalID != "" {
		resp.EvalID = evalID
		resp.EvalCreateIndex = i
	}
	resp.DeploymentModifyIndex = i
	resp.Index = i
	return nil
}

func (w *deploymentWatcher) FailDeployment(
	req *structs.DeploymentFailRequest,
	resp *structs.DeploymentUpdateResponse) error {

	status, desc := structs.DeploymentStatusFailed, structs.DeploymentStatusDescriptionFailedByUser

	// Determine if we should rollback
	rollback := false
	for _, dstate := range w.getDeployment().TaskGroups {
		if dstate.AutoRevert {
			rollback = true
			break
		}
	}

	var rollbackJob *structs.Job
	if rollback {
		var err error
		rollbackJob, err = w.latestStableJob()
		if err != nil {
			return err
		}

		if rollbackJob != nil {
			rollbackJob, desc = w.handleRollbackValidity(rollbackJob, desc)
		} else {
			desc = structs.DeploymentStatusDescriptionNoRollbackTarget(desc)
		}
	}

	// Commit the change
	update := w.getDeploymentStatusUpdate(status, desc)
	eval := w.getEval()
	i, err := w.upsertDeploymentStatusUpdate(update, eval, rollbackJob)
	if err != nil {
		return err
	}

	// Build the response
	resp.EvalID = eval.ID
	resp.EvalCreateIndex = i
	resp.DeploymentModifyIndex = i
	resp.Index = i
	if rollbackJob != nil {
		resp.RevertedJobVersion = pointer.Of(rollbackJob.Version)
	}
	return nil
}

// StopWatch stops watching the deployment. This should be called whenever a
// deployment is completed or the watcher is no longer needed.
func (w *deploymentWatcher) StopWatch() {
	w.exitFn()
}

// watch is the long running watcher that watches for both allocation and
// deployment changes. Its function is to create evaluations to trigger the
// scheduler when more progress can be made, to fail the deployment if it has
// failed and potentially rolling back the job. Progress can be made when an
// allocation transitions to healthy, so we create an eval.
func (w *deploymentWatcher) watch() {
	// Get the deadline. This is likely a zero time to begin with but we need to
	// handle the case that the deployment has already progressed and we are now
	// just starting to watch it. This must likely would occur if there was a
	// leader transition and we are now starting our watcher.
	currentDeadline := w.getDeploymentProgressCutoff(w.getDeployment())
	var deadlineTimer *time.Timer
	if currentDeadline.IsZero() {
		deadlineTimer = time.NewTimer(0)
		if !deadlineTimer.Stop() {
			<-deadlineTimer.C
		}
	} else {
		deadlineTimer = time.NewTimer(time.Until(currentDeadline))
	}

	allocIndex := uint64(1)
	allocsCh := w.getAllocsCh(allocIndex)
	var updates *allocUpdates

	rollback, deadlineHit := false, false

FAIL:
	for {
		select {
		case <-w.ctx.Done():
			// This is the successful case, and we stop the loop
			return
		case <-deadlineTimer.C:
			// We have hit the progress deadline, so fail the deployment
			// unless we're waiting for manual promotion. We need to determine
			// whether we should roll back the job by inspecting which allocs
			// as part of the deployment are healthy and which aren't. The
			// deadlineHit flag is never reset, so even in the case of a
			// manual promotion, we'll describe any failure as a progress
			// deadline failure at this point.
			deadlineHit = true
			fail, rback, err := w.shouldFail()
			if err != nil {
				w.logger.Error("failed to determine whether to rollback job", "error", err)
			}
			if !fail {
				w.logger.Debug("skipping deadline")
				continue
			}

			w.logger.Debug("deadline hit", "rollback", rback)
			rollback = rback
			err = w.nextRegion(structs.DeploymentStatusFailed)
			if err != nil {
				w.logger.Error("multiregion deployment error", "error", err)
			}
			break FAIL
		case <-w.deploymentUpdateCh:
			// Get the updated deployment and check if we should change the
			// deadline timer
			next := w.getDeploymentProgressCutoff(w.getDeployment())
			if !next.Equal(currentDeadline) {
				prevDeadlineZero := currentDeadline.IsZero()
				currentDeadline = next
				// The most recent deadline can be zero if no allocs were created for this deployment.
				// The deadline timer would have already been stopped once in that case. To prevent
				// deadlocking on the already stopped deadline timer, we only drain the channel if
				// the previous deadline was not zero.
				if !prevDeadlineZero && !deadlineTimer.Stop() {
					select {
					case <-deadlineTimer.C:
					default:
					}
				}

				// If the next deadline is zero, we should not reset the timer
				// as we aren't tracking towards a progress deadline yet. This
				// can happen if you have multiple task groups with progress
				// deadlines and one of the task groups hasn't made any
				// placements. As soon as the other task group finishes its
				// rollout, the next progress deadline becomes zero, so we want
				// to avoid resetting, causing a deployment failure.
				if !next.IsZero() {
					deadlineTimer.Reset(time.Until(next))
					w.logger.Trace("resetting deadline")
				}
			}

			err := w.nextRegion(w.getStatus())
			if err != nil {
				break FAIL
			}

		case updates = <-allocsCh:
			if err := updates.err; err != nil {
				if err == context.Canceled || w.ctx.Err() == context.Canceled {
					return
				}

				w.logger.Error("failed to retrieve allocations", "error", err)
				return
			}
			allocIndex = updates.index

			// We have allocation changes for this deployment so determine the
			// steps to take.
			res, err := w.handleAllocUpdate(updates.allocs)
			if err != nil {
				if err == context.Canceled || w.ctx.Err() == context.Canceled {
					return
				}

				w.logger.Error("failed handling allocation updates", "error", err)
				return
			}

			// The deployment has failed, so break out of the watch loop and
			// handle the failure
			if res.failDeployment {
				rollback = res.rollback
				err := w.nextRegion(structs.DeploymentStatusFailed)
				if err != nil {
					w.logger.Error("multiregion deployment error", "error", err)
				}
				break FAIL
			}

			// If permitted, automatically promote this canary deployment
			err = w.autoPromoteDeployment(updates.allocs)
			if err != nil {
				w.logger.Error("failed to auto promote deployment", "error", err)
			}

			// Create an eval to push the deployment along
			if res.createEval || len(res.allowReplacements) != 0 {
				w.createBatchedUpdate(res.allowReplacements, allocIndex)
			}

			// only start a new blocking query if we haven't returned early
			allocsCh = w.getAllocsCh(allocIndex)
		}
	}

	// Change the deployments status to failed
	desc := structs.DeploymentStatusDescriptionFailedAllocations
	if deadlineHit {
		desc = structs.DeploymentStatusDescriptionProgressDeadline
	}

	// Rollback to the old job if necessary
	var j *structs.Job
	if rollback {
		var err error
		j, err = w.latestStableJob()
		if err != nil {
			w.logger.Error("failed to lookup latest stable job", "error", err)
		}

		// Description should include that the job is being rolled back to
		// version N
		if j != nil {
			j, desc = w.handleRollbackValidity(j, desc)
		} else {
			desc = structs.DeploymentStatusDescriptionNoRollbackTarget(desc)
		}
	}

	// Update the status of the deployment to failed and create an evaluation.
	e := w.getEval()
	u := w.getDeploymentStatusUpdate(structs.DeploymentStatusFailed, desc)
	if _, err := w.upsertDeploymentStatusUpdate(u, e, j); err != nil {
		w.logger.Error("failed to update deployment status", "error", err)
	}
}

// allocUpdateResult is used to return the desired actions given the newest set
// of allocations for the deployment.
type allocUpdateResult struct {
	createEval        bool
	failDeployment    bool
	rollback          bool
	allowReplacements []string
}

// handleAllocUpdate is used to compute the set of actions to take based on the
// updated allocations for the deployment.
func (w *deploymentWatcher) handleAllocUpdate(allocs []*structs.AllocListStub) (allocUpdateResult, error) {
	var res allocUpdateResult

	// Get the latest evaluation index
	latestEval, err := w.jobEvalStatus()
	if err != nil {
		if err == context.Canceled || w.ctx.Err() == context.Canceled {
			return res, err
		}

		return res, fmt.Errorf("failed to determine last evaluation index for job %q: %v", w.j.ID, err)
	}

	deployment := w.getDeployment()
	for _, alloc := range allocs {
		dstate, ok := deployment.TaskGroups[alloc.TaskGroup]
		if !ok {
			continue
		}

		// Determine if the update block for this group is progress based
		progressBased := dstate.ProgressDeadline != 0

		// Check if the allocation has failed and we need to mark it for allow
		// replacements
		if progressBased && alloc.DeploymentStatus.IsUnhealthy() &&
			deployment.Active() && !alloc.DesiredTransition.ShouldReschedule() {
			res.allowReplacements = append(res.allowReplacements, alloc.ID)
			continue
		}

		// We need to create an eval so the job can progress.
		if alloc.DeploymentStatus.IsHealthy() && alloc.DeploymentStatus.ModifyIndex > latestEval {
			res.createEval = true
		}

		// If the group is using a progress deadline, we don't have to do anything.
		if progressBased {
			continue
		}

		// Fail on the first bad allocation
		if alloc.DeploymentStatus.IsUnhealthy() {
			// Check if the group has autorevert set
			if dstate.AutoRevert {
				res.rollback = true
			}

			// Since we have an unhealthy allocation, fail the deployment
			res.failDeployment = true
		}

		// All conditions have been hit so we can break
		if res.createEval && res.failDeployment && res.rollback {
			break
		}
	}

	return res, nil
}

// shouldFail returns whether the job should be failed and whether it should
// rolled back to an earlier stable version by examining the allocations in the
// deployment.
func (w *deploymentWatcher) shouldFail() (fail, rollback bool, err error) {
	snap, err := w.state.Snapshot()
	if err != nil {
		return false, false, err
	}

	d, err := snap.DeploymentByID(nil, w.deploymentID)
	if err != nil {
		return false, false, err
	}
	if d == nil {
		// The deployment wasn't in the state store, possibly due to a system gc
		return false, false, fmt.Errorf("deployment id not found: %q", w.deploymentID)
	}

	fail = false
	for tg, dstate := range d.TaskGroups {
		// If we are in a canary state we fail if there aren't enough healthy
		// allocs to satisfy DesiredCanaries
		if dstate.DesiredCanaries > 0 && !dstate.Promoted {
			if dstate.HealthyAllocs >= dstate.DesiredCanaries {
				continue
			}
		} else if dstate.HealthyAllocs >= dstate.DesiredTotal {
			continue
		}

		// We have failed this TG
		fail = true

		// We don't need to autorevert this group
		upd := w.j.LookupTaskGroup(tg).Update
		if upd == nil || !upd.AutoRevert {
			continue
		}

		// Unhealthy allocs and we need to autorevert
		return fail, true, nil
	}

	return fail, false, nil
}

// getDeploymentProgressCutoff returns the progress cutoff for the given
// deployment
func (w *deploymentWatcher) getDeploymentProgressCutoff(d *structs.Deployment) time.Time {
	var next time.Time
	doneTGs := w.doneGroups(d)
	for name, dstate := range d.TaskGroups {
		// This task group is done so we don't have to concern ourselves with
		// its progress deadline.
		if done, ok := doneTGs[name]; ok && done {
			continue
		}

		if dstate.RequireProgressBy.IsZero() {
			continue
		}

		if next.IsZero() || dstate.RequireProgressBy.Before(next) {
			next = dstate.RequireProgressBy
		}
	}
	return next
}

// doneGroups returns a map of task group to whether the deployment appears to
// be done for the group. A true value doesn't mean no more action will be taken
// in the life time of the deployment because there could always be node
// failures, or rescheduling events.
func (w *deploymentWatcher) doneGroups(d *structs.Deployment) map[string]bool {
	if d == nil {
		return nil
	}

	// Collect the allocations by the task group
	snap, err := w.state.Snapshot()
	if err != nil {
		return nil
	}

	allocs, err := snap.AllocsByDeployment(nil, d.ID)
	if err != nil {
		return nil
	}

	// Go through the allocs and count up how many healthy allocs we have
	healthy := make(map[string]int, len(d.TaskGroups))
	for _, a := range allocs {
		if a.TerminalStatus() || !a.DeploymentStatus.IsHealthy() {
			continue
		}
		healthy[a.TaskGroup]++
	}

	// Go through each group and check if it done
	groups := make(map[string]bool, len(d.TaskGroups))
	for name, dstate := range d.TaskGroups {
		// Requires promotion
		if dstate.DesiredCanaries != 0 && !dstate.Promoted {
			groups[name] = false
			continue
		}

		// Check we have enough healthy currently running allocations
		groups[name] = healthy[name] >= dstate.DesiredTotal
	}

	return groups
}

// latestStableJob returns the latest stable job. It may be nil if none exist
func (w *deploymentWatcher) latestStableJob() (*structs.Job, error) {
	snap, err := w.state.Snapshot()
	if err != nil {
		return nil, err
	}

	versions, err := snap.JobVersionsByID(nil, w.j.Namespace, w.j.ID)
	if err != nil {
		return nil, err
	}

	var stable *structs.Job
	for _, job := range versions {
		if job.Stable {
			stable = job
			break
		}
	}

	return stable, nil
}

// createBatchedUpdate creates an eval for the given index as well as updating
// the given allocations to allow them to reschedule.
func (w *deploymentWatcher) createBatchedUpdate(allowReplacements []string, forIndex uint64) {
	w.l.Lock()
	defer w.l.Unlock()

	// Store the allocations that can be replaced
	for _, allocID := range allowReplacements {
		if w.outstandingAllowReplacements == nil {
			w.outstandingAllowReplacements = make(map[string]*structs.DesiredTransition, len(allowReplacements))
		}
		w.outstandingAllowReplacements[allocID] = allowRescheduleTransition
	}

	if w.outstandingBatch || (forIndex < w.latestEval && len(allowReplacements) == 0) {
		return
	}

	w.outstandingBatch = true

	time.AfterFunc(perJobEvalBatchPeriod, func() {
		// If the timer has been created and then we shutdown, we need to no-op
		// the evaluation creation.
		select {
		case <-w.ctx.Done():
			return
		default:
		}

		w.l.Lock()
		replacements := w.outstandingAllowReplacements
		w.outstandingAllowReplacements = nil
		w.outstandingBatch = false
		w.l.Unlock()

		// Create the eval
		if _, err := w.createUpdate(replacements, w.getEval()); err != nil {
			w.logger.Error("failed to create evaluation for deployment", "deployment_id", w.deploymentID, "error", err)
		}
	})
}

// getEval returns an evaluation suitable for the deployment
func (w *deploymentWatcher) getEval() *structs.Evaluation {
	now := time.Now().UTC().UnixNano()

	// During a server upgrade it's possible we end up with deployments created
	// on the previous version that are then "watched" on a leader that's on
	// the new version. This would result in an eval with its priority set to
	// zero which would be bad. This therefore protects against that.
	w.l.Lock()
	priority := w.d.EvalPriority
	if priority == 0 {
		priority = w.j.Priority
	}
	w.l.Unlock()

	return &structs.Evaluation{
		ID:           uuid.Generate(),
		Namespace:    w.j.Namespace,
		Priority:     priority,
		Type:         w.j.Type,
		TriggeredBy:  structs.EvalTriggerDeploymentWatcher,
		JobID:        w.j.ID,
		DeploymentID: w.deploymentID,
		Status:       structs.EvalStatusPending,
		CreateTime:   now,
		ModifyTime:   now,
	}
}

// getDeploymentStatusUpdate returns a deployment status update
func (w *deploymentWatcher) getDeploymentStatusUpdate(status, desc string) *structs.DeploymentStatusUpdate {
	return &structs.DeploymentStatusUpdate{
		DeploymentID:      w.deploymentID,
		Status:            status,
		StatusDescription: desc,
	}
}

// getStatus returns the current status of the deployment
func (w *deploymentWatcher) getStatus() string {
	w.l.RLock()
	defer w.l.RUnlock()
	return w.d.Status
}

type allocUpdates struct {
	allocs []*structs.AllocListStub
	index  uint64
	err    error
}

// getAllocsCh creates a channel and starts a goroutine that
// 1. parks a blocking query for allocations on the state
// 2. reads those and drops them on the channel
// This query runs once here, but watch calls it in a loop
func (w *deploymentWatcher) getAllocsCh(index uint64) <-chan *allocUpdates {
	out := make(chan *allocUpdates, 1)
	go func() {
		allocs, index, err := w.getAllocs(index)
		out <- &allocUpdates{
			allocs: allocs,
			index:  index,
			err:    err,
		}
	}()

	return out
}

// getAllocs retrieves the allocations that are part of the deployment blocking
// at the given index.
func (w *deploymentWatcher) getAllocs(index uint64) ([]*structs.AllocListStub, uint64, error) {
	resp, index, err := w.state.BlockingQuery(w.getAllocsImpl, index, w.ctx)
	if err != nil {
		return nil, 0, err
	}
	if err := w.ctx.Err(); err != nil {
		return nil, 0, err
	}

	return resp.([]*structs.AllocListStub), index, nil
}

// getDeploysImpl retrieves all deployments from the passed state store.
func (w *deploymentWatcher) getAllocsImpl(ws memdb.WatchSet, state *state.StateStore) (interface{}, uint64, error) {
	if err := w.queryLimiter.Wait(w.ctx); err != nil {
		return nil, 0, err
	}

	// Capture all the allocations
	allocs, err := state.AllocsByDeployment(ws, w.deploymentID)
	if err != nil {
		return nil, 0, err
	}

	maxIndex := uint64(0)
	stubs := make([]*structs.AllocListStub, 0, len(allocs))
	for _, alloc := range allocs {
		stubs = append(stubs, alloc.Stub(nil))

		if maxIndex < alloc.ModifyIndex {
			maxIndex = alloc.ModifyIndex
		}
	}

	// Use the last index that affected the allocs table
	if len(stubs) == 0 {
		index, err := state.Index("allocs")
		if err != nil {
			return nil, index, err
		}
		maxIndex = index
	}

	return stubs, maxIndex, nil
}

// jobEvalStatus returns the latest eval index for a job. The index is used to
// determine if an allocation update requires an evaluation to be triggered.
func (w *deploymentWatcher) jobEvalStatus() (latestIndex uint64, err error) {
	if err := w.queryLimiter.Wait(w.ctx); err != nil {
		return 0, err
	}

	snap, err := w.state.Snapshot()
	if err != nil {
		return 0, err
	}

	evals, err := snap.EvalsByJob(nil, w.j.Namespace, w.j.ID)
	if err != nil {
		return 0, err
	}

	// If there are no evals for the job, return zero, since we want any
	// allocation change to trigger an evaluation.
	if len(evals) == 0 {
		return 0, nil
	}

	var max uint64
	for _, eval := range evals {
		// A cancelled eval never impacts what the scheduler has saw, so do not
		// use it's indexes.
		if eval.Status == structs.EvalStatusCancelled {
			continue
		}

		// Prefer using the snapshot index. Otherwise use the create index
		if eval.SnapshotIndex != 0 && max < eval.SnapshotIndex {
			max = eval.SnapshotIndex
		} else if max < eval.CreateIndex {
			max = eval.CreateIndex
		}
	}

	return max, nil
}