lib/dispatchcloud/dispatcher_test.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package dispatchcloud
   6
   7 import (
   8         "context"
   9         "encoding/json"
  10         "io/ioutil"
  11         "math/rand"
  12         "net/http"
  13         "net/http/httptest"
  14         "os"
  15         "sync"
  16         "time"
  17
  18         "git.curoverse.com/arvados.git/lib/dispatchcloud/test"
  19         "git.curoverse.com/arvados.git/sdk/go/arvados"
  20         "git.curoverse.com/arvados.git/sdk/go/ctxlog"
  21         "golang.org/x/crypto/ssh"
  22         check "gopkg.in/check.v1"
  23 )
  24
  25 var _ = check.Suite(&DispatcherSuite{})
  26
  27 type DispatcherSuite struct {
  28         ctx        context.Context
  29         cancel     context.CancelFunc
  30         cluster    *arvados.Cluster
  31         stubDriver *test.StubDriver
  32         disp       *dispatcher
  33 }
  34
  35 func (s *DispatcherSuite) SetUpTest(c *check.C) {
  36         s.ctx, s.cancel = context.WithCancel(context.Background())
  37         s.ctx = ctxlog.Context(s.ctx, ctxlog.TestLogger(c))
  38         dispatchpub, _ := test.LoadTestKey(c, "test/sshkey_dispatch")
  39         dispatchprivraw, err := ioutil.ReadFile("test/sshkey_dispatch")
  40         c.Assert(err, check.IsNil)
  41
  42         _, hostpriv := test.LoadTestKey(c, "test/sshkey_vm")
  43         s.stubDriver = &test.StubDriver{
  44                 HostKey:                   hostpriv,
  45                 AuthorizedKeys:            []ssh.PublicKey{dispatchpub},
  46                 ErrorRateDestroy:          0.1,
  47                 MinTimeBetweenCreateCalls: time.Millisecond,
  48         }
  49
  50         s.cluster = &arvados.Cluster{
  51                 CloudVMs: arvados.CloudVMs{
  52                         Driver:          "test",
  53                         SyncInterval:    arvados.Duration(10 * time.Millisecond),
  54                         TimeoutIdle:     arvados.Duration(150 * time.Millisecond),
  55                         TimeoutBooting:  arvados.Duration(150 * time.Millisecond),
  56                         TimeoutProbe:    arvados.Duration(15 * time.Millisecond),
  57                         TimeoutShutdown: arvados.Duration(5 * time.Millisecond),
  58                 },
  59                 Dispatch: arvados.Dispatch{
  60                         PrivateKey:         string(dispatchprivraw),
  61                         PollInterval:       arvados.Duration(5 * time.Millisecond),
  62                         ProbeInterval:      arvados.Duration(5 * time.Millisecond),
  63                         StaleLockTimeout:   arvados.Duration(5 * time.Millisecond),
  64                         MaxProbesPerSecond: 1000,
  65                 },
  66                 InstanceTypes: arvados.InstanceTypeMap{
  67                         test.InstanceType(1).Name:  test.InstanceType(1),
  68                         test.InstanceType(2).Name:  test.InstanceType(2),
  69                         test.InstanceType(3).Name:  test.InstanceType(3),
  70                         test.InstanceType(4).Name:  test.InstanceType(4),
  71                         test.InstanceType(6).Name:  test.InstanceType(6),
  72                         test.InstanceType(8).Name:  test.InstanceType(8),
  73                         test.InstanceType(16).Name: test.InstanceType(16),
  74                 },
  75                 NodeProfiles: map[string]arvados.NodeProfile{
  76                         "*": {
  77                                 Controller:    arvados.SystemServiceInstance{Listen: os.Getenv("ARVADOS_API_HOST")},
  78                                 DispatchCloud: arvados.SystemServiceInstance{Listen: ":"},
  79                         },
  80                 },
  81         }
  82         s.disp = &dispatcher{
  83                 Cluster: s.cluster,
  84                 Context: s.ctx,
  85         }
  86         // Test cases can modify s.cluster before calling
  87         // initialize(), and then modify private state before calling
  88         // go run().
  89 }
  90
  91 func (s *DispatcherSuite) TearDownTest(c *check.C) {
  92         s.cancel()
  93         s.disp.Close()
  94 }
  95
  96 // DispatchToStubDriver checks that the dispatcher wires everything
  97 // together effectively. It uses a real scheduler and worker pool with
  98 // a fake queue and cloud driver. The fake cloud driver injects
  99 // artificial errors in order to exercise a variety of code paths.
 100 func (s *DispatcherSuite) TestDispatchToStubDriver(c *check.C) {
 101         drivers["test"] = s.stubDriver
 102         s.disp.setupOnce.Do(s.disp.initialize)
 103         queue := &test.Queue{
 104                 ChooseType: func(ctr *arvados.Container) (arvados.InstanceType, error) {
 105                         return ChooseInstanceType(s.cluster, ctr)
 106                 },
 107         }
 108         for i := 0; i < 200; i++ {
 109                 queue.Containers = append(queue.Containers, arvados.Container{
 110                         UUID:     test.ContainerUUID(i + 1),
 111                         State:    arvados.ContainerStateQueued,
 112                         Priority: int64(i%20 + 1),
 113                         RuntimeConstraints: arvados.RuntimeConstraints{
 114                                 RAM:   int64(i%3+1) << 30,
 115                                 VCPUs: i%8 + 1,
 116                         },
 117                 })
 118         }
 119         s.disp.queue = queue
 120
 121         var mtx sync.Mutex
 122         done := make(chan struct{})
 123         waiting := map[string]struct{}{}
 124         for _, ctr := range queue.Containers {
 125                 waiting[ctr.UUID] = struct{}{}
 126         }
 127         executeContainer := func(ctr arvados.Container) int {
 128                 mtx.Lock()
 129                 defer mtx.Unlock()
 130                 if _, ok := waiting[ctr.UUID]; !ok {
 131                         c.Logf("container completed twice: %s -- perhaps completed after stub instance was killed?", ctr.UUID)
 132                         return 1
 133                 }
 134                 delete(waiting, ctr.UUID)
 135                 if len(waiting) == 0 {
 136                         close(done)
 137                 }
 138                 return int(rand.Uint32() & 0x3)
 139         }
 140         n := 0
 141         s.stubDriver.Queue = queue
 142         s.stubDriver.SetupVM = func(stubvm *test.StubVM) {
 143                 n++
 144                 stubvm.Boot = time.Now().Add(time.Duration(rand.Int63n(int64(5 * time.Millisecond))))
 145                 stubvm.CrunchRunDetachDelay = time.Duration(rand.Int63n(int64(10 * time.Millisecond)))
 146                 stubvm.ExecuteContainer = executeContainer
 147                 switch n % 7 {
 148                 case 0:
 149                         stubvm.Broken = time.Now().Add(time.Duration(rand.Int63n(90)) * time.Millisecond)
 150                 case 1:
 151                         stubvm.CrunchRunMissing = true
 152                 default:
 153                         stubvm.CrunchRunCrashRate = 0.1
 154                 }
 155         }
 156
 157         start := time.Now()
 158         go s.disp.run()
 159         err := s.disp.CheckHealth()
 160         c.Check(err, check.IsNil)
 161
 162         select {
 163         case <-done:
 164                 c.Logf("containers finished (%s), waiting for instances to shutdown and queue to clear", time.Since(start))
 165         case <-time.After(10 * time.Second):
 166                 c.Fatalf("timed out; still waiting for %d containers: %q", len(waiting), waiting)
 167         }
 168
 169         deadline := time.Now().Add(5 * time.Second)
 170         for range time.NewTicker(10 * time.Millisecond).C {
 171                 insts, err := s.stubDriver.InstanceSets()[0].Instances(nil)
 172                 c.Check(err, check.IsNil)
 173                 queue.Update()
 174                 ents, _ := queue.Entries()
 175                 if len(ents) == 0 && len(insts) == 0 {
 176                         break
 177                 }
 178                 if time.Now().After(deadline) {
 179                         c.Fatalf("timed out with %d containers (%v), %d instances (%+v)", len(ents), ents, len(insts), insts)
 180                 }
 181         }
 182 }
 183
 184 func (s *DispatcherSuite) TestAPIPermissions(c *check.C) {
 185         s.cluster.ManagementToken = "abcdefgh"
 186         drivers["test"] = s.stubDriver
 187         s.disp.setupOnce.Do(s.disp.initialize)
 188         s.disp.queue = &test.Queue{}
 189         go s.disp.run()
 190
 191         for _, token := range []string{"abc", ""} {
 192                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 193                 if token != "" {
 194                         req.Header.Set("Authorization", "Bearer "+token)
 195                 }
 196                 resp := httptest.NewRecorder()
 197                 s.disp.ServeHTTP(resp, req)
 198                 if token == "" {
 199                         c.Check(resp.Code, check.Equals, http.StatusUnauthorized)
 200                 } else {
 201                         c.Check(resp.Code, check.Equals, http.StatusForbidden)
 202                 }
 203         }
 204 }
 205
 206 func (s *DispatcherSuite) TestAPIDisabled(c *check.C) {
 207         s.cluster.ManagementToken = ""
 208         drivers["test"] = s.stubDriver
 209         s.disp.setupOnce.Do(s.disp.initialize)
 210         s.disp.queue = &test.Queue{}
 211         go s.disp.run()
 212
 213         for _, token := range []string{"abc", ""} {
 214                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 215                 if token != "" {
 216                         req.Header.Set("Authorization", "Bearer "+token)
 217                 }
 218                 resp := httptest.NewRecorder()
 219                 s.disp.ServeHTTP(resp, req)
 220                 c.Check(resp.Code, check.Equals, http.StatusForbidden)
 221         }
 222 }
 223
 224 func (s *DispatcherSuite) TestInstancesAPI(c *check.C) {
 225         s.cluster.ManagementToken = "abcdefgh"
 226         s.cluster.CloudVMs.TimeoutBooting = arvados.Duration(time.Second)
 227         drivers["test"] = s.stubDriver
 228         s.disp.setupOnce.Do(s.disp.initialize)
 229         s.disp.queue = &test.Queue{}
 230         go s.disp.run()
 231
 232         type instance struct {
 233                 Instance             string
 234                 WorkerState          string `json:"worker_state"`
 235                 Price                float64
 236                 LastContainerUUID    string `json:"last_container_uuid"`
 237                 ArvadosInstanceType  string `json:"arvados_instance_type"`
 238                 ProviderInstanceType string `json:"provider_instance_type"`
 239         }
 240         type instancesResponse struct {
 241                 Items []instance
 242         }
 243         getInstances := func() instancesResponse {
 244                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 245                 req.Header.Set("Authorization", "Bearer abcdefgh")
 246                 resp := httptest.NewRecorder()
 247                 s.disp.ServeHTTP(resp, req)
 248                 var sr instancesResponse
 249                 c.Check(resp.Code, check.Equals, http.StatusOK)
 250                 err := json.Unmarshal(resp.Body.Bytes(), &sr)
 251                 c.Check(err, check.IsNil)
 252                 return sr
 253         }
 254
 255         sr := getInstances()
 256         c.Check(len(sr.Items), check.Equals, 0)
 257
 258         ch := s.disp.pool.Subscribe()
 259         defer s.disp.pool.Unsubscribe(ch)
 260         ok := s.disp.pool.Create(test.InstanceType(1))
 261         c.Check(ok, check.Equals, true)
 262         <-ch
 263
 264         for deadline := time.Now().Add(time.Second); time.Now().Before(deadline); {
 265                 sr = getInstances()
 266                 if len(sr.Items) > 0 {
 267                         break
 268                 }
 269                 time.Sleep(time.Millisecond)
 270         }
 271         c.Assert(len(sr.Items), check.Equals, 1)
 272         c.Check(sr.Items[0].Instance, check.Matches, "stub.*")
 273         c.Check(sr.Items[0].WorkerState, check.Equals, "booting")
 274         c.Check(sr.Items[0].Price, check.Equals, 0.123)
 275         c.Check(sr.Items[0].LastContainerUUID, check.Equals, "")
 276         c.Check(sr.Items[0].ProviderInstanceType, check.Equals, test.InstanceType(1).ProviderType)
 277         c.Check(sr.Items[0].ArvadosInstanceType, check.Equals, test.InstanceType(1).Name)
 278 }