// applyPlan is used to apply the plan result and to return the alloc index
func (s *Server) applyPlan(result *structs.PlanResult, snap *state.StateSnapshot) (raft.ApplyFuture, error) {
req := structs.AllocUpdateRequest{}
for _, updateList := range result.NodeUpdate {
req.Alloc = append(req.Alloc, updateList...)
}
for _, allocList := range result.NodeAllocation {
req.Alloc = append(req.Alloc, allocList...)
}
req.Alloc = append(req.Alloc, result.FailedAllocs...)
// Set the time the alloc was applied for the first time. This can be used
// to approximate the scheduling time.
now := time.Now().UTC().UnixNano()
for _, alloc := range req.Alloc {
if alloc.CreateTime == 0 {
alloc.CreateTime = now
}
}
// Dispatch the Raft transaction
future, err := s.raftApplyFuture(structs.AllocUpdateRequestType, &req)
if err != nil {
return nil, err
}
// Optimistically apply to our state view
if snap != nil {
nextIdx := s.raft.AppliedIndex() + 1
if err := snap.UpsertAllocs(nextIdx, req.Alloc); err != nil {
return future, err
}
}
return future, nil
}
// updateNodeUpdateResponse assumes the n.srv.peerLock is held for reading.
func (n *Node) constructNodeServerInfoResponse(snap *state.StateSnapshot, reply *structs.NodeUpdateResponse) error {
reply.LeaderRPCAddr = n.srv.raft.Leader()
// Reply with config information required for future RPC requests
reply.Servers = make([]*structs.NodeServerInfo, 0, len(n.srv.localPeers))
for k, v := range n.srv.localPeers {
reply.Servers = append(reply.Servers,
&structs.NodeServerInfo{
RPCAdvertiseAddr: k,
RPCMajorVersion: int32(v.MajorVersion),
RPCMinorVersion: int32(v.MinorVersion),
Datacenter: v.Datacenter,
})
}
// TODO(sean@): Use an indexed node count instead
//
// Snapshot is used only to iterate over all nodes to create a node
// count to send back to Nomad Clients in their heartbeat so Clients
// can estimate the size of the cluster.
iter, err := snap.Nodes()
if err == nil {
for {
raw := iter.Next()
if raw == nil {
break
}
reply.NumNodes++
}
}
return nil
}
// applyPlan is used to apply the plan result and to return the alloc index
func (s *Server) applyPlan(result *structs.PlanResult, snap *state.StateSnapshot) (raft.ApplyFuture, error) {
req := structs.AllocUpdateRequest{}
for _, updateList := range result.NodeUpdate {
req.Alloc = append(req.Alloc, updateList...)
}
for _, allocList := range result.NodeAllocation {
req.Alloc = append(req.Alloc, allocList...)
}
req.Alloc = append(req.Alloc, result.FailedAllocs...)
// Dispatch the Raft transaction
future, err := s.raftApplyFuture(structs.AllocUpdateRequestType, &req)
if err != nil {
return nil, err
}
// Optimistically apply to our state view
if snap != nil {
nextIdx := s.raft.AppliedIndex() + 1
if err := snap.UpsertAllocs(nextIdx, req.Alloc); err != nil {
return future, err
}
}
return future, nil
}
// evaluateNodePlan is used to evalute the plan for a single node,
// returning if the plan is valid or if an error is encountered
func evaluateNodePlan(snap *state.StateSnapshot, plan *structs.Plan, nodeID string) (bool, error) {
// If this is an evict-only plan, it always 'fits' since we are removing things.
if len(plan.NodeAllocation[nodeID]) == 0 {
return true, nil
}
// Get the node itself
node, err := snap.NodeByID(nodeID)
if err != nil {
return false, fmt.Errorf("failed to get node '%s': %v", nodeID, err)
}
// If the node does not exist or is not ready for schduling it is not fit
// XXX: There is a potential race between when we do this check and when
// the Raft commit happens.
if node == nil || node.Status != structs.NodeStatusReady || node.Drain {
return false, nil
}
// Get the existing allocations
existingAlloc, err := snap.AllocsByNode(nodeID)
if err != nil {
return false, fmt.Errorf("failed to get existing allocations for '%s': %v", nodeID, err)
}
// Filter on alloc state
existingAlloc = structs.FilterTerminalAllocs(existingAlloc)
// Determine the proposed allocation by first removing allocations
// that are planned evictions and adding the new allocations.
proposed := existingAlloc
var remove []*structs.Allocation
if update := plan.NodeUpdate[nodeID]; len(update) > 0 {
remove = append(remove, update...)
}
if updated := plan.NodeAllocation[nodeID]; len(updated) > 0 {
for _, alloc := range updated {
remove = append(remove, alloc)
}
}
proposed = structs.RemoveAllocs(existingAlloc, remove)
proposed = append(proposed, plan.NodeAllocation[nodeID]...)
// Check if these allocations fit
fit, _, _, err := structs.AllocsFit(node, proposed, nil)
return fit, err
}
// applyPlan is used to apply the plan result and to return the alloc index
func (s *Server) applyPlan(job *structs.Job, result *structs.PlanResult, snap *state.StateSnapshot) (raft.ApplyFuture, error) {
// Determine the miniumum number of updates, could be more if there
// are multiple updates per node
minUpdates := len(result.NodeUpdate)
minUpdates += len(result.NodeAllocation)
minUpdates += len(result.FailedAllocs)
// Setup the update request
req := structs.AllocUpdateRequest{
Job: job,
Alloc: make([]*structs.Allocation, 0, minUpdates),
}
for _, updateList := range result.NodeUpdate {
req.Alloc = append(req.Alloc, updateList...)
}
for _, allocList := range result.NodeAllocation {
req.Alloc = append(req.Alloc, allocList...)
}
req.Alloc = append(req.Alloc, result.FailedAllocs...)
// Set the time the alloc was applied for the first time. This can be used
// to approximate the scheduling time.
now := time.Now().UTC().UnixNano()
for _, alloc := range req.Alloc {
if alloc.CreateTime == 0 {
alloc.CreateTime = now
}
}
// Dispatch the Raft transaction
future, err := s.raftApplyFuture(structs.AllocUpdateRequestType, &req)
if err != nil {
return nil, err
}
// Optimistically apply to our state view
if snap != nil {
nextIdx := s.raft.AppliedIndex() + 1
if err := snap.UpsertAllocs(nextIdx, req.Alloc); err != nil {
return future, err
}
}
return future, 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(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()
}
