// evaluatePlan is used to determine what portions of a plan
// can be applied if any. Returns if there should be a plan application
// which may be partial or if there was an error
func evaluatePlan(snap *state.StateSnapshot, plan *structs.Plan) (*structs.PlanResult, error) {
defer metrics.MeasureSince([]string{"nomad", "plan", "evaluate"}, time.Now())
// Create a result holder for the plan
result := &structs.PlanResult{
NodeUpdate: make(map[string][]*structs.Allocation),
NodeAllocation: make(map[string][]*structs.Allocation),
FailedAllocs: plan.FailedAllocs,
}
// Collect all the nodeIDs
nodeIDs := make(map[string]struct{})
for nodeID := range plan.NodeUpdate {
nodeIDs[nodeID] = struct{}{}
}
for nodeID := range plan.NodeAllocation {
nodeIDs[nodeID] = struct{}{}
}
// Check each allocation to see if it should be allowed
for nodeID := range nodeIDs {
// Evaluate the plan for this node
fit, err := evaluateNodePlan(snap, plan, nodeID)
if err != nil {
return nil, err
}
if !fit {
// Scheduler must have stale data, RefreshIndex should force
// the latest view of allocations and nodes
allocIndex, err := snap.Index("allocs")
if err != nil {
return nil, err
}
nodeIndex, err := snap.Index("nodes")
if err != nil {
return nil, err
}
result.RefreshIndex = maxUint64(nodeIndex, allocIndex)
// If we require all-at-once scheduling, there is no point
// to continue the evaluation, as we've already failed.
if plan.AllAtOnce {
result.NodeUpdate = nil
result.NodeAllocation = nil
return result, nil
}
// Skip this node, since it cannot be used.
continue
}
// Add this to the plan result
if nodeUpdate := plan.NodeUpdate[nodeID]; len(nodeUpdate) > 0 {
result.NodeUpdate[nodeID] = nodeUpdate
}
if nodeAlloc := plan.NodeAllocation[nodeID]; len(nodeAlloc) > 0 {
result.NodeAllocation[nodeID] = nodeAlloc
}
}
return result, nil
}
// evaluatePlan is used to determine what portions of a plan
// can be applied if any. Returns if there should be a plan application
// which may be partial or if there was an error
func evaluatePlan(pool *EvaluatePool, snap *state.StateSnapshot, plan *structs.Plan) (*structs.PlanResult, error) {
defer metrics.MeasureSince([]string{"nomad", "plan", "evaluate"}, time.Now())
// Create a result holder for the plan
result := &structs.PlanResult{
NodeUpdate: make(map[string][]*structs.Allocation),
NodeAllocation: make(map[string][]*structs.Allocation),
}
// Collect all the nodeIDs
nodeIDs := make(map[string]struct{})
nodeIDList := make([]string, 0, len(plan.NodeUpdate)+len(plan.NodeAllocation))
for nodeID := range plan.NodeUpdate {
if _, ok := nodeIDs[nodeID]; !ok {
nodeIDs[nodeID] = struct{}{}
nodeIDList = append(nodeIDList, nodeID)
}
}
for nodeID := range plan.NodeAllocation {
if _, ok := nodeIDs[nodeID]; !ok {
nodeIDs[nodeID] = struct{}{}
nodeIDList = append(nodeIDList, nodeID)
}
}
// Setup a multierror to handle potentially getting many
// errors since we are processing in parallel.
var mErr multierror.Error
partialCommit := false
// handleResult is used to process the result of evaluateNodePlan
handleResult := func(nodeID string, fit bool, err error) (cancel bool) {
// Evaluate the plan for this node
if err != nil {
mErr.Errors = append(mErr.Errors, err)
return true
}
if !fit {
// Set that this is a partial commit
partialCommit = true
// If we require all-at-once scheduling, there is no point
// to continue the evaluation, as we've already failed.
if plan.AllAtOnce {
result.NodeUpdate = nil
result.NodeAllocation = nil
return true
}
// Skip this node, since it cannot be used.
return
}
// Add this to the plan result
if nodeUpdate := plan.NodeUpdate[nodeID]; len(nodeUpdate) > 0 {
result.NodeUpdate[nodeID] = nodeUpdate
}
if nodeAlloc := plan.NodeAllocation[nodeID]; len(nodeAlloc) > 0 {
result.NodeAllocation[nodeID] = nodeAlloc
}
return
}
// Get the pool channels
req := pool.RequestCh()
resp := pool.ResultCh()
outstanding := 0
didCancel := false
// Evalute each node in the plan, handling results as they are ready to
// avoid blocking.
for len(nodeIDList) > 0 {
nodeID := nodeIDList[0]
select {
case req <- evaluateRequest{snap, plan, nodeID}:
outstanding++
nodeIDList = nodeIDList[1:]
case r := <-resp:
outstanding--
// Handle a result that allows us to cancel evaluation,
// which may save time processing additional entries.
if cancel := handleResult(r.nodeID, r.fit, r.err); cancel {
didCancel = true
break
}
}
}
// Drain the remaining results
for outstanding > 0 {
r := <-resp
if !didCancel {
if cancel := handleResult(r.nodeID, r.fit, r.err); cancel {
didCancel = true
}
}
outstanding--
}
// If the plan resulted in a partial commit, we need to determine
// a minimum refresh index to force the scheduler to work on a more
// up-to-date state to avoid the failures.
if partialCommit {
allocIndex, err := snap.Index("allocs")
if err != nil {
mErr.Errors = append(mErr.Errors, err)
}
nodeIndex, err := snap.Index("nodes")
if err != nil {
mErr.Errors = append(mErr.Errors, err)
}
result.RefreshIndex = maxUint64(nodeIndex, allocIndex)
if result.RefreshIndex == 0 {
err := fmt.Errorf("partialCommit with RefreshIndex of 0 (%d node, %d alloc)", nodeIndex, allocIndex)
mErr.Errors = append(mErr.Errors, err)
}
}
return result, mErr.ErrorOrNil()
}
