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": stubResp{},
  76                                 "true":              stubResp{},
  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) TestCreateUnallocShutdown(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         type1 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 156         type2 := arvados.InstanceType{Name: "a2m", ProviderType: "a2.medium", VCPUs: 2, RAM: 2 * GiB, Price: .02}
 157         type3 := arvados.InstanceType{Name: "a2l", ProviderType: "a2.large", VCPUs: 4, RAM: 4 * GiB, Price: .04}
 158         pool := &Pool{
 159                 logger:      logger,
 160                 newExecutor: func(cloud.Instance) Executor { return &stubExecutor{} },
 161                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 162                 instanceTypes: arvados.InstanceTypeMap{
 163                         type1.Name: type1,
 164                         type2.Name: type2,
 165                         type3.Name: type3,
 166                 },
 167         }
 168         notify := pool.Subscribe()
 169         defer pool.Unsubscribe(notify)
 170         notify2 := pool.Subscribe()
 171         defer pool.Unsubscribe(notify2)
 172
 173         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 174         c.Check(pool.Unallocated()[type2], check.Equals, 0)
 175         c.Check(pool.Unallocated()[type3], check.Equals, 0)
 176         pool.Create(type2)
 177         pool.Create(type1)
 178         pool.Create(type2)
 179         pool.Create(type3)
 180         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 181         c.Check(pool.Unallocated()[type2], check.Equals, 2)
 182         c.Check(pool.Unallocated()[type3], check.Equals, 1)
 183
 184         // Unblock the pending Create calls.
 185         go driver.ReleaseCloudOps(4)
 186
 187         // Wait for each instance to either return from its Create
 188         // call, or show up in a poll.
 189         suite.wait(c, pool, notify, func() bool {
 190                 pool.mtx.RLock()
 191                 defer pool.mtx.RUnlock()
 192                 return len(pool.workers) == 4
 193         })
 194
 195         // Place type3 node on admin-hold
 196         ivs := suite.instancesByType(pool, type3)
 197         c.Assert(ivs, check.HasLen, 1)
 198         type3instanceID := ivs[0].Instance
 199         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorHold)
 200         c.Check(err, check.IsNil)
 201
 202         // Check admin-hold behavior: refuse to shutdown, and don't
 203         // report as Unallocated ("available now or soon").
 204         c.Check(pool.Shutdown(type3), check.Equals, false)
 205         suite.wait(c, pool, notify, func() bool {
 206                 return pool.Unallocated()[type3] == 0
 207         })
 208         c.Check(suite.instancesByType(pool, type3), check.HasLen, 1)
 209
 210         // Shutdown both type2 nodes
 211         c.Check(pool.Shutdown(type2), check.Equals, true)
 212         suite.wait(c, pool, notify, func() bool {
 213                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 1
 214         })
 215         c.Check(pool.Shutdown(type2), check.Equals, true)
 216         suite.wait(c, pool, notify, func() bool {
 217                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 0
 218         })
 219         c.Check(pool.Shutdown(type2), check.Equals, false)
 220         for {
 221                 // Consume any waiting notifications to ensure the
 222                 // next one we get is from Shutdown.
 223                 select {
 224                 case <-notify:
 225                         continue
 226                 default:
 227                 }
 228                 break
 229         }
 230
 231         // Shutdown type1 node
 232         c.Check(pool.Shutdown(type1), check.Equals, true)
 233         suite.wait(c, pool, notify, func() bool {
 234                 return pool.Unallocated()[type1] == 0 && pool.Unallocated()[type2] == 0 && pool.Unallocated()[type3] == 0
 235         })
 236         select {
 237         case <-notify2:
 238         case <-time.After(time.Second):
 239                 c.Error("notify did not receive")
 240         }
 241
 242         // Put type3 node back in service.
 243         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorRun)
 244         c.Check(err, check.IsNil)
 245         suite.wait(c, pool, notify, func() bool {
 246                 return pool.Unallocated()[type3] == 1
 247         })
 248
 249         // Check admin-drain behavior: shut down right away, and don't
 250         // report as Unallocated.
 251         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorDrain)
 252         c.Check(err, check.IsNil)
 253         suite.wait(c, pool, notify, func() bool {
 254                 return pool.Unallocated()[type3] == 0
 255         })
 256         suite.wait(c, pool, notify, func() bool {
 257                 ivs := suite.instancesByType(pool, type3)
 258                 return len(ivs) == 1 && ivs[0].WorkerState == StateShutdown.String()
 259         })
 260
 261         // Unblock all pending Destroy calls. Pool calls Destroy again
 262         // if a node still appears in the provider list after a
 263         // previous attempt, so there might be more than 4 Destroy
 264         // calls to unblock.
 265         go driver.ReleaseCloudOps(4444)
 266
 267         // Sync until all instances disappear from the provider list.
 268         suite.wait(c, pool, notify, func() bool {
 269                 pool.getInstancesAndSync()
 270                 return len(pool.Instances()) == 0
 271         })
 272 }
 273
 274 func (suite *PoolSuite) instancesByType(pool *Pool, it arvados.InstanceType) []InstanceView {
 275         var ivs []InstanceView
 276         for _, iv := range pool.Instances() {
 277                 if iv.ArvadosInstanceType == it.Name {
 278                         ivs = append(ivs, iv)
 279                 }
 280         }
 281         return ivs
 282 }
 283
 284 func (suite *PoolSuite) wait(c *check.C, pool *Pool, notify <-chan struct{}, ready func() bool) {
 285         timeout := time.NewTimer(time.Second).C
 286         for !ready() {
 287                 select {
 288                 case <-notify:
 289                         continue
 290                 case <-timeout:
 291                 }
 292                 break
 293         }
 294         c.Check(ready(), check.Equals, true)
 295 }