forked from hashicorp/nomad
/
plan_apply.go
199 lines (175 loc) · 6.03 KB
/
plan_apply.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
package nomad
import (
"fmt"
"time"
"github.com/armon/go-metrics"
"github.com/hashicorp/nomad/nomad/state"
"github.com/hashicorp/nomad/nomad/structs"
)
// planApply is a long lived goroutine that reads plan allocations from
// the plan queue, determines if they can be applied safely and applies
// them via Raft.
func (s *Server) planApply() {
for {
// Pull the next pending plan, exit if we are no longer leader
pending, err := s.planQueue.Dequeue(0)
if err != nil {
return
}
// Verify the evaluation is outstanding, and that the tokens match.
token, ok := s.evalBroker.Outstanding(pending.plan.EvalID)
if !ok {
s.logger.Printf("[ERR] nomad: plan received for non-outstanding evaluation %s",
pending.plan.EvalID)
pending.respond(nil, fmt.Errorf("evaluation is not outstanding"))
continue
}
if pending.plan.EvalToken != token {
s.logger.Printf("[ERR] nomad: plan received for evaluation %s with wrong token",
pending.plan.EvalID)
pending.respond(nil, fmt.Errorf("evaluation token does not match"))
continue
}
// Snapshot the state so that we have a consistent view of the world
snap, err := s.fsm.State().Snapshot()
if err != nil {
s.logger.Printf("[ERR] nomad: failed to snapshot state: %v", err)
pending.respond(nil, err)
continue
}
// Evaluate the plan
result, err := evaluatePlan(snap, pending.plan)
if err != nil {
s.logger.Printf("[ERR] nomad: failed to evaluate plan: %v", err)
pending.respond(nil, err)
continue
}
// Apply the plan if there is anything to do
if !result.IsNoOp() {
allocIndex, err := s.applyPlan(result)
if err != nil {
s.logger.Printf("[ERR] nomad: failed to apply plan: %v", err)
pending.respond(nil, err)
continue
}
result.AllocIndex = allocIndex
}
// Respond to the plan
pending.respond(result, nil)
}
}
// applyPlan is used to apply the plan result and to return the alloc index
func (s *Server) applyPlan(result *structs.PlanResult) (uint64, error) {
defer metrics.MeasureSince([]string{"nomad", "plan", "apply"}, time.Now())
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...)
_, index, err := s.raftApply(structs.AllocUpdateRequestType, &req)
return index, err
}
// 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
}
// 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", node, 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", node, 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
}