lib/dispatchcloud/worker/pool_test.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package worker
   6
   7 import (
   8         "sort"
   9         "strings"
  10         "time"
  11
  12         "git.arvados.org/arvados.git/lib/cloud"
  13         "git.arvados.org/arvados.git/lib/dispatchcloud/test"
  14         "git.arvados.org/arvados.git/sdk/go/arvados"
  15         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  16         "github.com/prometheus/client_golang/prometheus"
  17         check "gopkg.in/check.v1"
  18 )
  19
  20 const GiB arvados.ByteSize = 1 << 30
  21
  22 var _ = check.Suite(&PoolSuite{})
  23
  24 type lessChecker struct {
  25         *check.CheckerInfo
  26 }
  27
  28 func (*lessChecker) Check(params []interface{}, names []string) (result bool, error string) {
  29         return params[0].(int) < params[1].(int), ""
  30 }
  31
  32 var less = &lessChecker{&check.CheckerInfo{Name: "less", Params: []string{"obtained", "expected"}}}
  33
  34 type PoolSuite struct{}
  35
  36 func (suite *PoolSuite) TestResumeAfterRestart(c *check.C) {
  37         type1 := test.InstanceType(1)
  38         type2 := test.InstanceType(2)
  39         type3 := test.InstanceType(3)
  40         waitForIdle := func(pool *Pool, notify <-chan struct{}) {
  41                 timeout := time.NewTimer(time.Second)
  42                 for {
  43                         instances := pool.Instances()
  44                         sort.Slice(instances, func(i, j int) bool {
  45                                 return strings.Compare(instances[i].ArvadosInstanceType, instances[j].ArvadosInstanceType) < 0
  46                         })
  47                         if len(instances) == 3 &&
  48                                 instances[0].ArvadosInstanceType == type1.Name &&
  49                                 instances[0].WorkerState == StateIdle.String() &&
  50                                 instances[1].ArvadosInstanceType == type1.Name &&
  51                                 instances[1].WorkerState == StateIdle.String() &&
  52                                 instances[2].ArvadosInstanceType == type2.Name &&
  53                                 instances[2].WorkerState == StateIdle.String() {
  54                                 return
  55                         }
  56                         select {
  57                         case <-timeout.C:
  58                                 c.Logf("pool.Instances() == %#v", instances)
  59                                 c.Error("timed out")
  60                                 return
  61                         case <-notify:
  62                         }
  63                 }
  64         }
  65
  66         logger := ctxlog.TestLogger(c)
  67         driver := &test.StubDriver{}
  68         instanceSetID := cloud.InstanceSetID("test-instance-set-id")
  69         is, err := driver.InstanceSet(nil, instanceSetID, nil, logger)
  70         c.Assert(err, check.IsNil)
  71
  72         newExecutor := func(cloud.Instance) Executor {
  73                 return &stubExecutor{
  74                         response: map[string]stubResp{
  75                                 "crunch-run --list": {},
  76                                 "true":              {},
  77                         },
  78                 }
  79         }
  80
  81         cluster := &arvados.Cluster{
  82                 Containers: arvados.ContainersConfig{
  83                         CloudVMs: arvados.CloudVMsConfig{
  84                                 BootProbeCommand:   "true",
  85                                 MaxProbesPerSecond: 1000,
  86                                 ProbeInterval:      arvados.Duration(time.Millisecond * 10),
  87                                 SyncInterval:       arvados.Duration(time.Millisecond * 10),
  88                                 TagKeyPrefix:       "testprefix:",
  89                         },
  90                 },
  91                 InstanceTypes: arvados.InstanceTypeMap{
  92                         type1.Name: type1,
  93                         type2.Name: type2,
  94                         type3.Name: type3,
  95                 },
  96         }
  97
  98         pool := NewPool(logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, cluster)
  99         notify := pool.Subscribe()
 100         defer pool.Unsubscribe(notify)
 101         pool.Create(type1)
 102         pool.Create(type1)
 103         pool.Create(type2)
 104         waitForIdle(pool, notify)
 105         var heldInstanceID cloud.InstanceID
 106         for _, inst := range pool.Instances() {
 107                 if inst.ArvadosInstanceType == type2.Name {
 108                         heldInstanceID = cloud.InstanceID(inst.Instance)
 109                         pool.SetIdleBehavior(heldInstanceID, IdleBehaviorHold)
 110                 }
 111         }
 112         // Wait for the tags to save to the cloud provider
 113         tagKey := cluster.Containers.CloudVMs.TagKeyPrefix + tagKeyIdleBehavior
 114         deadline := time.Now().Add(time.Second)
 115         for !func() bool {
 116                 pool.mtx.RLock()
 117                 defer pool.mtx.RUnlock()
 118                 for _, wkr := range pool.workers {
 119                         if wkr.instType == type2 {
 120                                 return wkr.instance.Tags()[tagKey] == string(IdleBehaviorHold)
 121                         }
 122                 }
 123                 return false
 124         }() {
 125                 if time.Now().After(deadline) {
 126                         c.Fatal("timeout")
 127                 }
 128                 time.Sleep(time.Millisecond * 10)
 129         }
 130         pool.Stop()
 131
 132         c.Log("------- starting new pool, waiting to recover state")
 133
 134         pool2 := NewPool(logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, cluster)
 135         notify2 := pool2.Subscribe()
 136         defer pool2.Unsubscribe(notify2)
 137         waitForIdle(pool2, notify2)
 138         for _, inst := range pool2.Instances() {
 139                 if inst.ArvadosInstanceType == type2.Name {
 140                         c.Check(inst.Instance, check.Equals, heldInstanceID)
 141                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorHold)
 142                 } else {
 143                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorRun)
 144                 }
 145         }
 146         pool2.Stop()
 147 }
 148
 149 func (suite *PoolSuite) TestDrain(c *check.C) {
 150         logger := ctxlog.TestLogger(c)
 151         driver := test.StubDriver{}
 152         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, logger)
 153         c.Assert(err, check.IsNil)
 154
 155         ac := arvados.NewClientFromEnv()
 156
 157         type1 := test.InstanceType(1)
 158         pool := &Pool{
 159                 arvClient:   ac,
 160                 logger:      logger,
 161                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 162                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 163                 instanceTypes: arvados.InstanceTypeMap{
 164                         type1.Name: type1,
 165                 },
 166         }
 167         notify := pool.Subscribe()
 168         defer pool.Unsubscribe(notify)
 169
 170         pool.Create(type1)
 171
 172         // Wait for the instance to either return from its Create
 173         // call, or show up in a poll.
 174         suite.wait(c, pool, notify, func() bool {
 175                 pool.mtx.RLock()
 176                 defer pool.mtx.RUnlock()
 177                 return len(pool.workers) == 1
 178         })
 179
 180         tests := []struct {
 181                 state        State
 182                 idleBehavior IdleBehavior
 183                 result       bool
 184         }{
 185                 {StateIdle, IdleBehaviorHold, false},
 186                 {StateIdle, IdleBehaviorDrain, false},
 187                 {StateIdle, IdleBehaviorRun, true},
 188         }
 189
 190         for _, test := range tests {
 191                 for _, wkr := range pool.workers {
 192                         wkr.state = test.state
 193                         wkr.idleBehavior = test.idleBehavior
 194                 }
 195
 196                 // Try to start a container
 197                 started := pool.StartContainer(type1, arvados.Container{UUID: "testcontainer"})
 198                 c.Check(started, check.Equals, test.result)
 199         }
 200 }
 201
 202 func (suite *PoolSuite) TestCreateUnallocShutdown(c *check.C) {
 203         logger := ctxlog.TestLogger(c)
 204         driver := test.StubDriver{HoldCloudOps: true}
 205         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, logger)
 206         c.Assert(err, check.IsNil)
 207
 208         type1 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 209         type2 := arvados.InstanceType{Name: "a2m", ProviderType: "a2.medium", VCPUs: 2, RAM: 2 * GiB, Price: .02}
 210         type3 := arvados.InstanceType{Name: "a2l", ProviderType: "a2.large", VCPUs: 4, RAM: 4 * GiB, Price: .04}
 211         pool := &Pool{
 212                 logger:      logger,
 213                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 214                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 215                 instanceTypes: arvados.InstanceTypeMap{
 216                         type1.Name: type1,
 217                         type2.Name: type2,
 218                         type3.Name: type3,
 219                 },
 220         }
 221         notify := pool.Subscribe()
 222         defer pool.Unsubscribe(notify)
 223         notify2 := pool.Subscribe()
 224         defer pool.Unsubscribe(notify2)
 225
 226         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 227         c.Check(pool.Unallocated()[type2], check.Equals, 0)
 228         c.Check(pool.Unallocated()[type3], check.Equals, 0)
 229         pool.Create(type2)
 230         pool.Create(type1)
 231         pool.Create(type2)
 232         pool.Create(type3)
 233         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 234         c.Check(pool.Unallocated()[type2], check.Equals, 2)
 235         c.Check(pool.Unallocated()[type3], check.Equals, 1)
 236
 237         // Unblock the pending Create calls.
 238         go driver.ReleaseCloudOps(4)
 239
 240         // Wait for each instance to either return from its Create
 241         // call, or show up in a poll.
 242         suite.wait(c, pool, notify, func() bool {
 243                 pool.mtx.RLock()
 244                 defer pool.mtx.RUnlock()
 245                 return len(pool.workers) == 4
 246         })
 247
 248         // Place type3 node on admin-hold
 249         ivs := suite.instancesByType(pool, type3)
 250         c.Assert(ivs, check.HasLen, 1)
 251         type3instanceID := ivs[0].Instance
 252         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorHold)
 253         c.Check(err, check.IsNil)
 254
 255         // Check admin-hold behavior: refuse to shutdown, and don't
 256         // report as Unallocated ("available now or soon").
 257         c.Check(pool.Shutdown(type3), check.Equals, false)
 258         suite.wait(c, pool, notify, func() bool {
 259                 return pool.Unallocated()[type3] == 0
 260         })
 261         c.Check(suite.instancesByType(pool, type3), check.HasLen, 1)
 262
 263         // Shutdown both type2 nodes
 264         c.Check(pool.Shutdown(type2), check.Equals, true)
 265         suite.wait(c, pool, notify, func() bool {
 266                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 1
 267         })
 268         c.Check(pool.Shutdown(type2), check.Equals, true)
 269         suite.wait(c, pool, notify, func() bool {
 270                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 0
 271         })
 272         c.Check(pool.Shutdown(type2), check.Equals, false)
 273         for {
 274                 // Consume any waiting notifications to ensure the
 275                 // next one we get is from Shutdown.
 276                 select {
 277                 case <-notify:
 278                         continue
 279                 default:
 280                 }
 281                 break
 282         }
 283
 284         // Shutdown type1 node
 285         c.Check(pool.Shutdown(type1), check.Equals, true)
 286         suite.wait(c, pool, notify, func() bool {
 287                 return pool.Unallocated()[type1] == 0 && pool.Unallocated()[type2] == 0 && pool.Unallocated()[type3] == 0
 288         })
 289         select {
 290         case <-notify2:
 291         case <-time.After(time.Second):
 292                 c.Error("notify did not receive")
 293         }
 294
 295         // Put type3 node back in service.
 296         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorRun)
 297         c.Check(err, check.IsNil)
 298         suite.wait(c, pool, notify, func() bool {
 299                 return pool.Unallocated()[type3] == 1
 300         })
 301
 302         // Check admin-drain behavior: shut down right away, and don't
 303         // report as Unallocated.
 304         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorDrain)
 305         c.Check(err, check.IsNil)
 306         suite.wait(c, pool, notify, func() bool {
 307                 return pool.Unallocated()[type3] == 0
 308         })
 309         suite.wait(c, pool, notify, func() bool {
 310                 ivs := suite.instancesByType(pool, type3)
 311                 return len(ivs) == 1 && ivs[0].WorkerState == StateShutdown.String()
 312         })
 313
 314         // Unblock all pending Destroy calls. Pool calls Destroy again
 315         // if a node still appears in the provider list after a
 316         // previous attempt, so there might be more than 4 Destroy
 317         // calls to unblock.
 318         go driver.ReleaseCloudOps(4444)
 319
 320         // Sync until all instances disappear from the provider list.
 321         suite.wait(c, pool, notify, func() bool {
 322                 pool.getInstancesAndSync()
 323                 return len(pool.Instances()) == 0
 324         })
 325 }
 326
 327 func (suite *PoolSuite) instancesByType(pool *Pool, it arvados.InstanceType) []InstanceView {
 328         var ivs []InstanceView
 329         for _, iv := range pool.Instances() {
 330                 if iv.ArvadosInstanceType == it.Name {
 331                         ivs = append(ivs, iv)
 332                 }
 333         }
 334         return ivs
 335 }
 336
 337 func (suite *PoolSuite) wait(c *check.C, pool *Pool, notify <-chan struct{}, ready func() bool) {
 338         timeout := time.NewTimer(time.Second).C
 339         for !ready() {
 340                 select {
 341                 case <-notify:
 342                         continue
 343                 case <-timeout:
 344                 }
 345                 break
 346         }
 347         c.Check(ready(), check.Equals, true)
 348 }