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{HoldCloudOps: true}
 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 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 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         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 171         pool.Create(type1)
 172         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 173
 174         // Unblock the pending Create call.
 175         go driver.ReleaseCloudOps(1)
 176
 177         // Wait for the instance to either return from its Create
 178         // call, or show up in a poll.
 179         suite.wait(c, pool, notify, func() bool {
 180                 pool.mtx.RLock()
 181                 defer pool.mtx.RUnlock()
 182                 return len(pool.workers) == 1
 183         })
 184
 185         tests := []struct {
 186                 state        State
 187                 idleBehavior IdleBehavior
 188                 result       bool
 189         }{
 190                 {StateIdle, IdleBehaviorHold, false},
 191                 {StateIdle, IdleBehaviorDrain, false},
 192                 {StateIdle, IdleBehaviorRun, true},
 193         }
 194
 195         for _, test := range tests {
 196                 for _, wkr := range pool.workers {
 197                         if wkr.instType == type1 {
 198                                 wkr.state = test.state
 199                                 wkr.idleBehavior = test.idleBehavior
 200                         }
 201                 }
 202
 203                 // Try to start another container
 204                 started := pool.StartContainer(type1, arvados.Container{UUID: "testcontainer"})
 205                 c.Check(started, check.Equals, test.result)
 206         }
 207 }
 208
 209 func (suite *PoolSuite) TestCreateUnallocShutdown(c *check.C) {
 210         logger := ctxlog.TestLogger(c)
 211         driver := test.StubDriver{HoldCloudOps: true}
 212         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, logger)
 213         c.Assert(err, check.IsNil)
 214
 215         type1 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 216         type2 := arvados.InstanceType{Name: "a2m", ProviderType: "a2.medium", VCPUs: 2, RAM: 2 * GiB, Price: .02}
 217         type3 := arvados.InstanceType{Name: "a2l", ProviderType: "a2.large", VCPUs: 4, RAM: 4 * GiB, Price: .04}
 218         pool := &Pool{
 219                 logger:      logger,
 220                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 221                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 222                 instanceTypes: arvados.InstanceTypeMap{
 223                         type1.Name: type1,
 224                         type2.Name: type2,
 225                         type3.Name: type3,
 226                 },
 227         }
 228         notify := pool.Subscribe()
 229         defer pool.Unsubscribe(notify)
 230         notify2 := pool.Subscribe()
 231         defer pool.Unsubscribe(notify2)
 232
 233         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 234         c.Check(pool.Unallocated()[type2], check.Equals, 0)
 235         c.Check(pool.Unallocated()[type3], check.Equals, 0)
 236         pool.Create(type2)
 237         pool.Create(type1)
 238         pool.Create(type2)
 239         pool.Create(type3)
 240         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 241         c.Check(pool.Unallocated()[type2], check.Equals, 2)
 242         c.Check(pool.Unallocated()[type3], check.Equals, 1)
 243
 244         // Unblock the pending Create calls.
 245         go driver.ReleaseCloudOps(4)
 246
 247         // Wait for each instance to either return from its Create
 248         // call, or show up in a poll.
 249         suite.wait(c, pool, notify, func() bool {
 250                 pool.mtx.RLock()
 251                 defer pool.mtx.RUnlock()
 252                 return len(pool.workers) == 4
 253         })
 254
 255         // Place type3 node on admin-hold
 256         ivs := suite.instancesByType(pool, type3)
 257         c.Assert(ivs, check.HasLen, 1)
 258         type3instanceID := ivs[0].Instance
 259         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorHold)
 260         c.Check(err, check.IsNil)
 261
 262         // Check admin-hold behavior: refuse to shutdown, and don't
 263         // report as Unallocated ("available now or soon").
 264         c.Check(pool.Shutdown(type3), check.Equals, false)
 265         suite.wait(c, pool, notify, func() bool {
 266                 return pool.Unallocated()[type3] == 0
 267         })
 268         c.Check(suite.instancesByType(pool, type3), check.HasLen, 1)
 269
 270         // Shutdown both type2 nodes
 271         c.Check(pool.Shutdown(type2), check.Equals, true)
 272         suite.wait(c, pool, notify, func() bool {
 273                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 1
 274         })
 275         c.Check(pool.Shutdown(type2), check.Equals, true)
 276         suite.wait(c, pool, notify, func() bool {
 277                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 0
 278         })
 279         c.Check(pool.Shutdown(type2), check.Equals, false)
 280         for {
 281                 // Consume any waiting notifications to ensure the
 282                 // next one we get is from Shutdown.
 283                 select {
 284                 case <-notify:
 285                         continue
 286                 default:
 287                 }
 288                 break
 289         }
 290
 291         // Shutdown type1 node
 292         c.Check(pool.Shutdown(type1), check.Equals, true)
 293         suite.wait(c, pool, notify, func() bool {
 294                 return pool.Unallocated()[type1] == 0 && pool.Unallocated()[type2] == 0 && pool.Unallocated()[type3] == 0
 295         })
 296         select {
 297         case <-notify2:
 298         case <-time.After(time.Second):
 299                 c.Error("notify did not receive")
 300         }
 301
 302         // Put type3 node back in service.
 303         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorRun)
 304         c.Check(err, check.IsNil)
 305         suite.wait(c, pool, notify, func() bool {
 306                 return pool.Unallocated()[type3] == 1
 307         })
 308
 309         // Check admin-drain behavior: shut down right away, and don't
 310         // report as Unallocated.
 311         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorDrain)
 312         c.Check(err, check.IsNil)
 313         suite.wait(c, pool, notify, func() bool {
 314                 return pool.Unallocated()[type3] == 0
 315         })
 316         suite.wait(c, pool, notify, func() bool {
 317                 ivs := suite.instancesByType(pool, type3)
 318                 return len(ivs) == 1 && ivs[0].WorkerState == StateShutdown.String()
 319         })
 320
 321         // Unblock all pending Destroy calls. Pool calls Destroy again
 322         // if a node still appears in the provider list after a
 323         // previous attempt, so there might be more than 4 Destroy
 324         // calls to unblock.
 325         go driver.ReleaseCloudOps(4444)
 326
 327         // Sync until all instances disappear from the provider list.
 328         suite.wait(c, pool, notify, func() bool {
 329                 pool.getInstancesAndSync()
 330                 return len(pool.Instances()) == 0
 331         })
 332 }
 333
 334 func (suite *PoolSuite) instancesByType(pool *Pool, it arvados.InstanceType) []InstanceView {
 335         var ivs []InstanceView
 336         for _, iv := range pool.Instances() {
 337                 if iv.ArvadosInstanceType == it.Name {
 338                         ivs = append(ivs, iv)
 339                 }
 340         }
 341         return ivs
 342 }
 343
 344 func (suite *PoolSuite) wait(c *check.C, pool *Pool, notify <-chan struct{}, ready func() bool) {
 345         timeout := time.NewTimer(time.Second).C
 346         for !ready() {
 347                 select {
 348                 case <-notify:
 349                         continue
 350                 case <-timeout:
 351                 }
 352                 break
 353         }
 354         c.Check(ready(), check.Equals, true)
 355 }