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.curoverse.com/arvados.git/lib/cloud"
  13         "git.curoverse.com/arvados.git/lib/dispatchcloud/test"
  14         "git.curoverse.com/arvados.git/sdk/go/arvados"
  15         "git.curoverse.com/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                         "crunch-run --list": stubResp{},
  75                         "true":              stubResp{},
  76                 }
  77         }
  78
  79         cluster := &arvados.Cluster{
  80                 Containers: arvados.ContainersConfig{
  81                         CloudVMs: arvados.CloudVMsConfig{
  82                                 BootProbeCommand:   "true",
  83                                 MaxProbesPerSecond: 1000,
  84                                 ProbeInterval:      arvados.Duration(time.Millisecond * 10),
  85                                 SyncInterval:       arvados.Duration(time.Millisecond * 10),
  86                                 TagKeyPrefix:       "testprefix:",
  87                         },
  88                 },
  89                 InstanceTypes: arvados.InstanceTypeMap{
  90                         type1.Name: type1,
  91                         type2.Name: type2,
  92                         type3.Name: type3,
  93                 },
  94         }
  95
  96         pool := NewPool(logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, cluster)
  97         notify := pool.Subscribe()
  98         defer pool.Unsubscribe(notify)
  99         pool.Create(type1)
 100         pool.Create(type1)
 101         pool.Create(type2)
 102         waitForIdle(pool, notify)
 103         var heldInstanceID cloud.InstanceID
 104         for _, inst := range pool.Instances() {
 105                 if inst.ArvadosInstanceType == type2.Name {
 106                         heldInstanceID = cloud.InstanceID(inst.Instance)
 107                         pool.SetIdleBehavior(heldInstanceID, IdleBehaviorHold)
 108                 }
 109         }
 110         // Wait for the tags to save to the cloud provider
 111         tagKey := cluster.Containers.CloudVMs.TagKeyPrefix + tagKeyIdleBehavior
 112         deadline := time.Now().Add(time.Second)
 113         for !func() bool {
 114                 pool.mtx.RLock()
 115                 defer pool.mtx.RUnlock()
 116                 for _, wkr := range pool.workers {
 117                         if wkr.instType == type2 {
 118                                 return wkr.instance.Tags()[tagKey] == string(IdleBehaviorHold)
 119                         }
 120                 }
 121                 return false
 122         }() {
 123                 if time.Now().After(deadline) {
 124                         c.Fatal("timeout")
 125                 }
 126                 time.Sleep(time.Millisecond * 10)
 127         }
 128         pool.Stop()
 129
 130         c.Log("------- starting new pool, waiting to recover state")
 131
 132         pool2 := NewPool(logger, arvados.NewClientFromEnv(), prometheus.NewRegistry(), instanceSetID, is, newExecutor, nil, cluster)
 133         notify2 := pool2.Subscribe()
 134         defer pool2.Unsubscribe(notify2)
 135         waitForIdle(pool2, notify2)
 136         for _, inst := range pool2.Instances() {
 137                 if inst.ArvadosInstanceType == type2.Name {
 138                         c.Check(inst.Instance, check.Equals, heldInstanceID)
 139                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorHold)
 140                 } else {
 141                         c.Check(inst.IdleBehavior, check.Equals, IdleBehaviorRun)
 142                 }
 143         }
 144         pool2.Stop()
 145 }
 146
 147 func (suite *PoolSuite) TestCreateUnallocShutdown(c *check.C) {
 148         logger := ctxlog.TestLogger(c)
 149         driver := test.StubDriver{HoldCloudOps: true}
 150         instanceSet, err := driver.InstanceSet(nil, "test-instance-set-id", nil, logger)
 151         c.Assert(err, check.IsNil)
 152
 153         type1 := arvados.InstanceType{Name: "a1s", ProviderType: "a1.small", VCPUs: 1, RAM: 1 * GiB, Price: .01}
 154         type2 := arvados.InstanceType{Name: "a2m", ProviderType: "a2.medium", VCPUs: 2, RAM: 2 * GiB, Price: .02}
 155         type3 := arvados.InstanceType{Name: "a2l", ProviderType: "a2.large", VCPUs: 4, RAM: 4 * GiB, Price: .04}
 156         pool := &Pool{
 157                 logger:      logger,
 158                 newExecutor: func(cloud.Instance) Executor { return stubExecutor{} },
 159                 instanceSet: &throttledInstanceSet{InstanceSet: instanceSet},
 160                 instanceTypes: arvados.InstanceTypeMap{
 161                         type1.Name: type1,
 162                         type2.Name: type2,
 163                         type3.Name: type3,
 164                 },
 165         }
 166         notify := pool.Subscribe()
 167         defer pool.Unsubscribe(notify)
 168         notify2 := pool.Subscribe()
 169         defer pool.Unsubscribe(notify2)
 170
 171         c.Check(pool.Unallocated()[type1], check.Equals, 0)
 172         c.Check(pool.Unallocated()[type2], check.Equals, 0)
 173         c.Check(pool.Unallocated()[type3], check.Equals, 0)
 174         pool.Create(type2)
 175         pool.Create(type1)
 176         pool.Create(type2)
 177         pool.Create(type3)
 178         c.Check(pool.Unallocated()[type1], check.Equals, 1)
 179         c.Check(pool.Unallocated()[type2], check.Equals, 2)
 180         c.Check(pool.Unallocated()[type3], check.Equals, 1)
 181
 182         // Unblock the pending Create calls.
 183         go driver.ReleaseCloudOps(4)
 184
 185         // Wait for each instance to either return from its Create
 186         // call, or show up in a poll.
 187         suite.wait(c, pool, notify, func() bool {
 188                 pool.mtx.RLock()
 189                 defer pool.mtx.RUnlock()
 190                 return len(pool.workers) == 4
 191         })
 192
 193         // Place type3 node on admin-hold
 194         ivs := suite.instancesByType(pool, type3)
 195         c.Assert(ivs, check.HasLen, 1)
 196         type3instanceID := ivs[0].Instance
 197         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorHold)
 198         c.Check(err, check.IsNil)
 199
 200         // Check admin-hold behavior: refuse to shutdown, and don't
 201         // report as Unallocated ("available now or soon").
 202         c.Check(pool.Shutdown(type3), check.Equals, false)
 203         suite.wait(c, pool, notify, func() bool {
 204                 return pool.Unallocated()[type3] == 0
 205         })
 206         c.Check(suite.instancesByType(pool, type3), check.HasLen, 1)
 207
 208         // Shutdown both type2 nodes
 209         c.Check(pool.Shutdown(type2), check.Equals, true)
 210         suite.wait(c, pool, notify, func() bool {
 211                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 1
 212         })
 213         c.Check(pool.Shutdown(type2), check.Equals, true)
 214         suite.wait(c, pool, notify, func() bool {
 215                 return pool.Unallocated()[type1] == 1 && pool.Unallocated()[type2] == 0
 216         })
 217         c.Check(pool.Shutdown(type2), check.Equals, false)
 218         for {
 219                 // Consume any waiting notifications to ensure the
 220                 // next one we get is from Shutdown.
 221                 select {
 222                 case <-notify:
 223                         continue
 224                 default:
 225                 }
 226                 break
 227         }
 228
 229         // Shutdown type1 node
 230         c.Check(pool.Shutdown(type1), check.Equals, true)
 231         suite.wait(c, pool, notify, func() bool {
 232                 return pool.Unallocated()[type1] == 0 && pool.Unallocated()[type2] == 0 && pool.Unallocated()[type3] == 0
 233         })
 234         select {
 235         case <-notify2:
 236         case <-time.After(time.Second):
 237                 c.Error("notify did not receive")
 238         }
 239
 240         // Put type3 node back in service.
 241         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorRun)
 242         c.Check(err, check.IsNil)
 243         suite.wait(c, pool, notify, func() bool {
 244                 return pool.Unallocated()[type3] == 1
 245         })
 246
 247         // Check admin-drain behavior: shut down right away, and don't
 248         // report as Unallocated.
 249         err = pool.SetIdleBehavior(type3instanceID, IdleBehaviorDrain)
 250         c.Check(err, check.IsNil)
 251         suite.wait(c, pool, notify, func() bool {
 252                 return pool.Unallocated()[type3] == 0
 253         })
 254         suite.wait(c, pool, notify, func() bool {
 255                 ivs := suite.instancesByType(pool, type3)
 256                 return len(ivs) == 1 && ivs[0].WorkerState == StateShutdown.String()
 257         })
 258
 259         // Unblock all pending Destroy calls. Pool calls Destroy again
 260         // if a node still appears in the provider list after a
 261         // previous attempt, so there might be more than 4 Destroy
 262         // calls to unblock.
 263         go driver.ReleaseCloudOps(4444)
 264
 265         // Sync until all instances disappear from the provider list.
 266         suite.wait(c, pool, notify, func() bool {
 267                 pool.getInstancesAndSync()
 268                 return len(pool.Instances()) == 0
 269         })
 270 }
 271
 272 func (suite *PoolSuite) instancesByType(pool *Pool, it arvados.InstanceType) []InstanceView {
 273         var ivs []InstanceView
 274         for _, iv := range pool.Instances() {
 275                 if iv.ArvadosInstanceType == it.Name {
 276                         ivs = append(ivs, iv)
 277                 }
 278         }
 279         return ivs
 280 }
 281
 282 func (suite *PoolSuite) wait(c *check.C, pool *Pool, notify <-chan struct{}, ready func() bool) {
 283         timeout := time.NewTimer(time.Second).C
 284         for !ready() {
 285                 select {
 286                 case <-notify:
 287                         continue
 288                 case <-timeout:
 289                 }
 290                 break
 291         }
 292         c.Check(ready(), check.Equals, true)
 293 }