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/config"
  14         "git.arvados.org/arvados.git/lib/dispatchcloud/test"
  15         "git.arvados.org/arvados.git/sdk/go/arvados"
  16         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  17         "github.com/prometheus/client_golang/prometheus"
  18         "github.com/sirupsen/logrus"
  19         check "gopkg.in/check.v1"
  20 )
  21
  22 const GiB arvados.ByteSize = 1 << 30
  23
  24 var _ = check.Suite(&PoolSuite{})
  25
  26 type lessChecker struct {
  27         *check.CheckerInfo
  28 }
  29
  30 func (*lessChecker) Check(params []interface{}, names []string) (result bool, error string) {
  31         return params[0].(int) < params[1].(int), ""
  32 }
  33
  34 var less = &lessChecker{&check.CheckerInfo{Name: "less", Params: []string{"obtained", "expected"}}}
  35
  36 type PoolSuite struct {
  37         logger      logrus.FieldLogger
  38         testCluster *arvados.Cluster
  39 }
  40
  41 func (suite *PoolSuite) SetUpTest(c *check.C) {
  42         suite.logger = ctxlog.TestLogger(c)
  43         cfg, err := config.NewLoader(nil, suite.logger).Load()
  44         c.Assert(err, check.IsNil)
  45         suite.testCluster, err = cfg.GetCluster("")
  46         c.Assert(err, check.IsNil)
  47 }
  48
  49 func (suite *PoolSuite) TestResumeAfterRestart(c *check.C) {
  50         type1 := test.InstanceType(1)
  51         type2 := test.InstanceType(2)
  52         type3 := test.InstanceType(3)
  53         waitForIdle := func(pool *Pool, notify <-chan struct{}) {
  54                 timeout := time.NewTimer(time.Second)
  55                 for {
  56                         instances := pool.Instances()
  57                         sort.Slice(instances, func(i, j int) bool {
  58                                 return strings.Compare(instances[i].ArvadosInstanceType, instances[j].ArvadosInstanceType) < 0
  59                         })
  60                         if len(instances) == 3 &&
  61                                 instances[0].ArvadosInstanceType == type1.Name &&
  62                                 instances[0].WorkerState == StateIdle.String() &&
  63                                 instances[1].ArvadosInstanceType == type1.Name &&
  64                                 instances[1].WorkerState == StateIdle.String() &&
  65                                 instances[2].ArvadosInstanceType == type2.Name &&
  66                                 instances[2].WorkerState == StateIdle.String() {
  67                                 return
  68                         }
  69                         select {
  70                         case <-timeout.C:
  71                                 c.Logf("pool.Instances() == %#v", instances)
  72                                 c.Error("timed out")
  73                                 return
  74                         case <-notify:
  75                         }
  76                 }
  77         }
  78
  79         driver := &test.StubDriver{}
  80         instanceSetID := cloud.InstanceSetID("test-instance-set-id")
  81         is, err := driver.InstanceSet(nil, instanceSetID, nil, suite.logger)
  82         c.Assert(err, check.IsNil)
  83
  84         newExecutor := func(cloud.Instance) Executor {
  85                 return &stubExecutor{
  86                         response: map[string]stubResp{
  87                                 "crunch-run-custom --list": {},
  88                                 "true":                     {},
  89                         },
  90                 }
  91         }
  92
  93         suite.testCluster.Containers.CloudVMs = arvados.CloudVMsConfig{
  94                 BootProbeCommand:   "true",
  95                 MaxProbesPerSecond: 1000,
  96                 ProbeInterval:      arvados.Duration(time.Millisecond * 10),
  97                 SyncInterval:       arvados.Duration(time.Millisecond * 10),
  98                 TagKeyPrefix:       "testprefix:",
  99         }
 100         suite.testCluster.Containers.CrunchRunCommand = "crunch-run-custom"
 101         suite.testCluster.InstanceTypes = arvados.InstanceTypeMap{
 102                 type1.Name: type1,
 103                 type2.Name: type2,
 104                 type3.Name: type3,
 105         }
 106
 107         pool := NewPool(suite.logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, suite.testCluster)
 108         notify := pool.Subscribe()
 109         defer pool.Unsubscribe(notify)
 110         pool.Create(type1)
 111         pool.Create(type1)
 112         pool.Create(type2)
 113         waitForIdle(pool, notify)
 114         var heldInstanceID cloud.InstanceID
 115         for _, inst := range pool.Instances() {
 116                 if inst.ArvadosInstanceType == type2.Name {
 117                         heldInstanceID = cloud.InstanceID(inst.Instance)
 118                         pool.SetIdleBehavior(heldInstanceID, IdleBehaviorHold)
 119                 }
 120         }
 121         // Wait for the tags to save to the cloud provider
 122         tagKey := suite.testCluster.Containers.CloudVMs.TagKeyPrefix + tagKeyIdleBehavior
 123         deadline := time.Now().Add(time.Second)
 124         for !func() bool {
 125                 pool.mtx.RLock()
 126                 defer pool.mtx.RUnlock()
 127                 for _, wkr := range pool.workers {
 128                         if wkr.instType == type2 {
 129                                 return wkr.instance.Tags()[tagKey] == string(IdleBehaviorHold)
 130                         }
 131                 }
 132                 return false
 133         }() {
 134                 if time.Now().After(deadline) {
 135                         c.Fatal("timeout")
 136                 }
 137                 time.Sleep(time.Millisecond * 10)
 138         }
 139         pool.Stop()
 140
 141         c.Log("------- starting new pool, waiting to recover state")
 142
 143         pool2 := NewPool(suite.logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, suite.testCluster)
 144         notify2 := pool2.Subscribe()
 145         defer pool2.Unsubscribe(notify2)
 146         waitForIdle(pool2, notify2)
 147         for _, inst := range pool2.Instances() {
 148                 if inst.ArvadosInstanceType == type2.Name {
 149                         c.Check(inst.Instance, check.Equals, heldInstanceID)
 150                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorHold)
 151                 } else {
 152                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorRun)
 153                 }
 154         }
 155         pool2.Stop()
 156 }
 157
 158 func (suite *PoolSuite) TestDrain(c *check.C) {
 159         driver := test.StubDriver{}
 160         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, suite.logger)
 161         c.Assert(err, check.IsNil)
 162
 163         ac := arvados.NewClientFromEnv()
 164
 165         type1 := test.InstanceType(1)
 166         pool := &Pool{
 167                 arvClient:   ac,
 168                 logger:      suite.logger,
 169                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 170                 cluster:     suite.testCluster,
 171                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 172                 instanceTypes: arvados.InstanceTypeMap{
 173                         type1.Name: type1,
 174                 },
 175         }
 176         notify := pool.Subscribe()
 177         defer pool.Unsubscribe(notify)
 178
 179         pool.Create(type1)
 180
 181         // Wait for the instance to either return from its Create
 182         // call, or show up in a poll.
 183         suite.wait(c, pool, notify, func() bool {
 184                 pool.mtx.RLock()
 185                 defer pool.mtx.RUnlock()
 186                 return len(pool.workers) == 1
 187         })
 188
 189         tests := []struct {
 190                 state        State
 191                 idleBehavior IdleBehavior
 192                 result       bool
 193         }{
 194                 {StateIdle, IdleBehaviorHold, false},
 195                 {StateIdle, IdleBehaviorDrain, false},
 196                 {StateIdle, IdleBehaviorRun, true},
 197         }
 198
 199         for _, test := range tests {
 200                 for _, wkr := range pool.workers {
 201                         wkr.state = test.state
 202                         wkr.idleBehavior = test.idleBehavior
 203                 }
 204
 205                 // Try to start a container
 206                 started := pool.StartContainer(type1, arvados.Container{UUID: "testcontainer"})
 207                 c.Check(started, check.Equals, test.result)
 208         }
 209 }
 210
 211 func (suite *PoolSuite) TestNodeCreateThrottle(c *check.C) {
 212         driver := test.StubDriver{HoldCloudOps: true}
 213         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, suite.logger)
 214         c.Assert(err, check.IsNil)
 215
 216         type1 := test.InstanceType(1)
 217         pool := &Pool{
 218                 logger:                         suite.logger,
 219                 instanceSet:                    &throttledInstanceSet{InstanceSet: instanceSet},
 220                 cluster:                        suite.testCluster,
 221                 maxConcurrentInstanceCreateOps: 1,
 222                 instanceTypes: arvados.InstanceTypeMap{
 223                         type1.Name: type1,
 224                 },
 225         }
 226
 227         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 228         res := pool.Create(type1)
 229         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 230         c.Check(res, check.Equals, true)
 231
 232         res = pool.Create(type1)
 233         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 234         c.Check(res, check.Equals, false)
 235
 236         pool.instanceSet.throttleCreate.err = nil
 237         pool.maxConcurrentInstanceCreateOps = 2
 238
 239         res = pool.Create(type1)
 240         c.Check(pool.Unallocated()[type1], check.Equals, 2)
 241         c.Check(res, check.Equals, true)
 242
 243         pool.instanceSet.throttleCreate.err = nil
 244         pool.maxConcurrentInstanceCreateOps = 0
 245
 246         res = pool.Create(type1)
 247         c.Check(pool.Unallocated()[type1], check.Equals, 3)
 248         c.Check(res, check.Equals, true)
 249 }
 250
 251 func (suite *PoolSuite) TestCreateUnallocShutdown(c *check.C) {
 252         driver := test.StubDriver{HoldCloudOps: true}
 253         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, suite.logger)
 254         c.Assert(err, check.IsNil)
 255
 256         type1 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 257         type2 := arvados.InstanceType{Name: "a2m", ProviderType: "a2.medium", VCPUs: 2, RAM: 2 * GiB, Price: .02}
 258         type3 := arvados.InstanceType{Name: "a2l", ProviderType: "a2.large", VCPUs: 4, RAM: 4 * GiB, Price: .04}
 259         pool := &Pool{
 260                 logger:      suite.logger,
 261                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 262                 cluster:     suite.testCluster,
 263                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 264                 instanceTypes: arvados.InstanceTypeMap{
 265                         type1.Name: type1,
 266                         type2.Name: type2,
 267                         type3.Name: type3,
 268                 },
 269         }
 270         notify := pool.Subscribe()
 271         defer pool.Unsubscribe(notify)
 272         notify2 := pool.Subscribe()
 273         defer pool.Unsubscribe(notify2)
 274
 275         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 276         c.Check(pool.Unallocated()[type2], check.Equals, 0)
 277         c.Check(pool.Unallocated()[type3], check.Equals, 0)
 278         pool.Create(type2)
 279         pool.Create(type1)
 280         pool.Create(type2)
 281         pool.Create(type3)
 282         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 283         c.Check(pool.Unallocated()[type2], check.Equals, 2)
 284         c.Check(pool.Unallocated()[type3], check.Equals, 1)
 285
 286         // Unblock the pending Create calls.
 287         go driver.ReleaseCloudOps(4)
 288
 289         // Wait for each instance to either return from its Create
 290         // call, or show up in a poll.
 291         suite.wait(c, pool, notify, func() bool {
 292                 pool.mtx.RLock()
 293                 defer pool.mtx.RUnlock()
 294                 return len(pool.workers) == 4
 295         })
 296
 297         // Place type3 node on admin-hold
 298         ivs := suite.instancesByType(pool, type3)
 299         c.Assert(ivs, check.HasLen, 1)
 300         type3instanceID := ivs[0].Instance
 301         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorHold)
 302         c.Check(err, check.IsNil)
 303
 304         // Check admin-hold behavior: refuse to shutdown, and don't
 305         // report as Unallocated ("available now or soon").
 306         c.Check(pool.Shutdown(type3), check.Equals, false)
 307         suite.wait(c, pool, notify, func() bool {
 308                 return pool.Unallocated()[type3] == 0
 309         })
 310         c.Check(suite.instancesByType(pool, type3), check.HasLen, 1)
 311
 312         // Shutdown both type2 nodes
 313         c.Check(pool.Shutdown(type2), check.Equals, true)
 314         suite.wait(c, pool, notify, func() bool {
 315                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 1
 316         })
 317         c.Check(pool.Shutdown(type2), check.Equals, true)
 318         suite.wait(c, pool, notify, func() bool {
 319                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 0
 320         })
 321         c.Check(pool.Shutdown(type2), check.Equals, false)
 322         for {
 323                 // Consume any waiting notifications to ensure the
 324                 // next one we get is from Shutdown.
 325                 select {
 326                 case <-notify:
 327                         continue
 328                 default:
 329                 }
 330                 break
 331         }
 332
 333         // Shutdown type1 node
 334         c.Check(pool.Shutdown(type1), check.Equals, true)
 335         suite.wait(c, pool, notify, func() bool {
 336                 return pool.Unallocated()[type1] == 0 && pool.Unallocated()[type2] == 0 && pool.Unallocated()[type3] == 0
 337         })
 338         select {
 339         case <-notify2:
 340         case <-time.After(time.Second):
 341                 c.Error("notify did not receive")
 342         }
 343
 344         // Put type3 node back in service.
 345         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorRun)
 346         c.Check(err, check.IsNil)
 347         suite.wait(c, pool, notify, func() bool {
 348                 return pool.Unallocated()[type3] == 1
 349         })
 350
 351         // Check admin-drain behavior: shut down right away, and don't
 352         // report as Unallocated.
 353         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorDrain)
 354         c.Check(err, check.IsNil)
 355         suite.wait(c, pool, notify, func() bool {
 356                 return pool.Unallocated()[type3] == 0
 357         })
 358         suite.wait(c, pool, notify, func() bool {
 359                 ivs := suite.instancesByType(pool, type3)
 360                 return len(ivs) == 1 && ivs[0].WorkerState == StateShutdown.String()
 361         })
 362
 363         // Unblock all pending Destroy calls. Pool calls Destroy again
 364         // if a node still appears in the provider list after a
 365         // previous attempt, so there might be more than 4 Destroy
 366         // calls to unblock.
 367         go driver.ReleaseCloudOps(4444)
 368
 369         // Sync until all instances disappear from the provider list.
 370         suite.wait(c, pool, notify, func() bool {
 371                 pool.getInstancesAndSync()
 372                 return len(pool.Instances()) == 0
 373         })
 374 }
 375
 376 func (suite *PoolSuite) instancesByType(pool *Pool, it arvados.InstanceType) []InstanceView {
 377         var ivs []InstanceView
 378         for _, iv := range pool.Instances() {
 379                 if iv.ArvadosInstanceType == it.Name {
 380                         ivs = append(ivs, iv)
 381                 }
 382         }
 383         return ivs
 384 }
 385
 386 func (suite *PoolSuite) wait(c *check.C, pool *Pool, notify <-chan struct{}, ready func() bool) {
 387         timeout := time.NewTimer(time.Second).C
 388         for !ready() {
 389                 select {
 390                 case <-notify:
 391                         continue
 392                 case <-timeout:
 393                 }
 394                 break
 395         }
 396         c.Check(ready(), check.Equals, true)
 397 }