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/arvadostest"
  21         "git.curoverse.com/arvados.git/sdk/go/ctxlog"
  22         "golang.org/x/crypto/ssh"
  23         check "gopkg.in/check.v1"
  24 )
  25
  26 var _ = check.Suite(&DispatcherSuite{})
  27
  28 type DispatcherSuite struct {
  29         ctx        context.Context
  30         cancel     context.CancelFunc
  31         cluster    *arvados.Cluster
  32         stubDriver *test.StubDriver
  33         disp       *dispatcher
  34 }
  35
  36 func (s *DispatcherSuite) SetUpTest(c *check.C) {
  37         s.ctx, s.cancel = context.WithCancel(context.Background())
  38         s.ctx = ctxlog.Context(s.ctx, ctxlog.TestLogger(c))
  39         dispatchpub, _ := test.LoadTestKey(c, "test/sshkey_dispatch")
  40         dispatchprivraw, err := ioutil.ReadFile("test/sshkey_dispatch")
  41         c.Assert(err, check.IsNil)
  42
  43         _, hostpriv := test.LoadTestKey(c, "test/sshkey_vm")
  44         s.stubDriver = &test.StubDriver{
  45                 HostKey:                   hostpriv,
  46                 AuthorizedKeys:            []ssh.PublicKey{dispatchpub},
  47                 ErrorRateDestroy:          0.1,
  48                 MinTimeBetweenCreateCalls: time.Millisecond,
  49         }
  50
  51         s.cluster = &arvados.Cluster{
  52                 CloudVMs: arvados.CloudVMs{
  53                         Driver:               "test",
  54                         SyncInterval:         arvados.Duration(10 * time.Millisecond),
  55                         TimeoutIdle:          arvados.Duration(150 * time.Millisecond),
  56                         TimeoutBooting:       arvados.Duration(150 * time.Millisecond),
  57                         TimeoutProbe:         arvados.Duration(15 * time.Millisecond),
  58                         TimeoutShutdown:      arvados.Duration(5 * time.Millisecond),
  59                         MaxCloudOpsPerSecond: 500,
  60                 },
  61                 Dispatch: arvados.Dispatch{
  62                         PrivateKey:         string(dispatchprivraw),
  63                         PollInterval:       arvados.Duration(5 * time.Millisecond),
  64                         ProbeInterval:      arvados.Duration(5 * time.Millisecond),
  65                         StaleLockTimeout:   arvados.Duration(5 * time.Millisecond),
  66                         MaxProbesPerSecond: 1000,
  67                         TimeoutSignal:      arvados.Duration(3 * time.Millisecond),
  68                         TimeoutTERM:        arvados.Duration(20 * time.Millisecond),
  69                 },
  70                 InstanceTypes: arvados.InstanceTypeMap{
  71                         test.InstanceType(1).Name:  test.InstanceType(1),
  72                         test.InstanceType(2).Name:  test.InstanceType(2),
  73                         test.InstanceType(3).Name:  test.InstanceType(3),
  74                         test.InstanceType(4).Name:  test.InstanceType(4),
  75                         test.InstanceType(6).Name:  test.InstanceType(6),
  76                         test.InstanceType(8).Name:  test.InstanceType(8),
  77                         test.InstanceType(16).Name: test.InstanceType(16),
  78                 },
  79                 NodeProfiles: map[string]arvados.NodeProfile{
  80                         "*": {
  81                                 Controller:    arvados.SystemServiceInstance{Listen: os.Getenv("ARVADOS_API_HOST")},
  82                                 DispatchCloud: arvados.SystemServiceInstance{Listen: ":"},
  83                         },
  84                 },
  85                 Services: arvados.Services{
  86                         Controller: arvados.Service{ExternalURL: arvados.URL{Scheme: "https", Host: os.Getenv("ARVADOS_API_HOST")}},
  87                 },
  88         }
  89
  90         arvClient, err := arvados.NewClientFromConfig(s.cluster)
  91         c.Check(err, check.IsNil)
  92
  93         s.disp = &dispatcher{
  94                 Cluster:   s.cluster,
  95                 Context:   s.ctx,
  96                 ArvClient: arvClient,
  97                 AuthToken: arvadostest.AdminToken,
  98         }
  99         // Test cases can modify s.cluster before calling
 100         // initialize(), and then modify private state before calling
 101         // go run().
 102 }
 103
 104 func (s *DispatcherSuite) TearDownTest(c *check.C) {
 105         s.cancel()
 106         s.disp.Close()
 107 }
 108
 109 // DispatchToStubDriver checks that the dispatcher wires everything
 110 // together effectively. It uses a real scheduler and worker pool with
 111 // a fake queue and cloud driver. The fake cloud driver injects
 112 // artificial errors in order to exercise a variety of code paths.
 113 func (s *DispatcherSuite) TestDispatchToStubDriver(c *check.C) {
 114         drivers["test"] = s.stubDriver
 115         s.disp.setupOnce.Do(s.disp.initialize)
 116         queue := &test.Queue{
 117                 ChooseType: func(ctr *arvados.Container) (arvados.InstanceType, error) {
 118                         return ChooseInstanceType(s.cluster, ctr)
 119                 },
 120         }
 121         for i := 0; i < 200; i++ {
 122                 queue.Containers = append(queue.Containers, arvados.Container{
 123                         UUID:     test.ContainerUUID(i + 1),
 124                         State:    arvados.ContainerStateQueued,
 125                         Priority: int64(i%20 + 1),
 126                         RuntimeConstraints: arvados.RuntimeConstraints{
 127                                 RAM:   int64(i%3+1) << 30,
 128                                 VCPUs: i%8 + 1,
 129                         },
 130                 })
 131         }
 132         s.disp.queue = queue
 133
 134         var mtx sync.Mutex
 135         done := make(chan struct{})
 136         waiting := map[string]struct{}{}
 137         for _, ctr := range queue.Containers {
 138                 waiting[ctr.UUID] = struct{}{}
 139         }
 140         finishContainer := func(ctr arvados.Container) {
 141                 mtx.Lock()
 142                 defer mtx.Unlock()
 143                 if _, ok := waiting[ctr.UUID]; !ok {
 144                         c.Errorf("container completed twice: %s", ctr.UUID)
 145                         return
 146                 }
 147                 delete(waiting, ctr.UUID)
 148                 if len(waiting) == 0 {
 149                         close(done)
 150                 }
 151         }
 152         executeContainer := func(ctr arvados.Container) int {
 153                 finishContainer(ctr)
 154                 return int(rand.Uint32() & 0x3)
 155         }
 156         n := 0
 157         s.stubDriver.Queue = queue
 158         s.stubDriver.SetupVM = func(stubvm *test.StubVM) {
 159                 n++
 160                 stubvm.Boot = time.Now().Add(time.Duration(rand.Int63n(int64(5 * time.Millisecond))))
 161                 stubvm.CrunchRunDetachDelay = time.Duration(rand.Int63n(int64(10 * time.Millisecond)))
 162                 stubvm.ExecuteContainer = executeContainer
 163                 stubvm.CrashRunningContainer = finishContainer
 164                 switch n % 7 {
 165                 case 0:
 166                         stubvm.Broken = time.Now().Add(time.Duration(rand.Int63n(90)) * time.Millisecond)
 167                 case 1:
 168                         stubvm.CrunchRunMissing = true
 169                 case 2:
 170                         stubvm.ReportBroken = time.Now().Add(time.Duration(rand.Int63n(200)) * time.Millisecond)
 171                 default:
 172                         stubvm.CrunchRunCrashRate = 0.1
 173                 }
 174         }
 175
 176         start := time.Now()
 177         go s.disp.run()
 178         err := s.disp.CheckHealth()
 179         c.Check(err, check.IsNil)
 180
 181         select {
 182         case <-done:
 183                 c.Logf("containers finished (%s), waiting for instances to shutdown and queue to clear", time.Since(start))
 184         case <-time.After(10 * time.Second):
 185                 c.Fatalf("timed out; still waiting for %d containers: %q", len(waiting), waiting)
 186         }
 187
 188         deadline := time.Now().Add(5 * time.Second)
 189         for range time.NewTicker(10 * time.Millisecond).C {
 190                 insts, err := s.stubDriver.InstanceSets()[0].Instances(nil)
 191                 c.Check(err, check.IsNil)
 192                 queue.Update()
 193                 ents, _ := queue.Entries()
 194                 if len(ents) == 0 && len(insts) == 0 {
 195                         break
 196                 }
 197                 if time.Now().After(deadline) {
 198                         c.Fatalf("timed out with %d containers (%v), %d instances (%+v)", len(ents), ents, len(insts), insts)
 199                 }
 200         }
 201 }
 202
 203 func (s *DispatcherSuite) TestAPIPermissions(c *check.C) {
 204         s.cluster.ManagementToken = "abcdefgh"
 205         drivers["test"] = s.stubDriver
 206         s.disp.setupOnce.Do(s.disp.initialize)
 207         s.disp.queue = &test.Queue{}
 208         go s.disp.run()
 209
 210         for _, token := range []string{"abc", ""} {
 211                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 212                 if token != "" {
 213                         req.Header.Set("Authorization", "Bearer "+token)
 214                 }
 215                 resp := httptest.NewRecorder()
 216                 s.disp.ServeHTTP(resp, req)
 217                 if token == "" {
 218                         c.Check(resp.Code, check.Equals, http.StatusUnauthorized)
 219                 } else {
 220                         c.Check(resp.Code, check.Equals, http.StatusForbidden)
 221                 }
 222         }
 223 }
 224
 225 func (s *DispatcherSuite) TestAPIDisabled(c *check.C) {
 226         s.cluster.ManagementToken = ""
 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         for _, token := range []string{"abc", ""} {
 233                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 234                 if token != "" {
 235                         req.Header.Set("Authorization", "Bearer "+token)
 236                 }
 237                 resp := httptest.NewRecorder()
 238                 s.disp.ServeHTTP(resp, req)
 239                 c.Check(resp.Code, check.Equals, http.StatusForbidden)
 240         }
 241 }
 242
 243 func (s *DispatcherSuite) TestInstancesAPI(c *check.C) {
 244         s.cluster.ManagementToken = "abcdefgh"
 245         s.cluster.CloudVMs.TimeoutBooting = arvados.Duration(time.Second)
 246         drivers["test"] = s.stubDriver
 247         s.disp.setupOnce.Do(s.disp.initialize)
 248         s.disp.queue = &test.Queue{}
 249         go s.disp.run()
 250
 251         type instance struct {
 252                 Instance             string
 253                 WorkerState          string `json:"worker_state"`
 254                 Price                float64
 255                 LastContainerUUID    string `json:"last_container_uuid"`
 256                 ArvadosInstanceType  string `json:"arvados_instance_type"`
 257                 ProviderInstanceType string `json:"provider_instance_type"`
 258         }
 259         type instancesResponse struct {
 260                 Items []instance
 261         }
 262         getInstances := func() instancesResponse {
 263                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 264                 req.Header.Set("Authorization", "Bearer abcdefgh")
 265                 resp := httptest.NewRecorder()
 266                 s.disp.ServeHTTP(resp, req)
 267                 var sr instancesResponse
 268                 c.Check(resp.Code, check.Equals, http.StatusOK)
 269                 err := json.Unmarshal(resp.Body.Bytes(), &sr)
 270                 c.Check(err, check.IsNil)
 271                 return sr
 272         }
 273
 274         sr := getInstances()
 275         c.Check(len(sr.Items), check.Equals, 0)
 276
 277         ch := s.disp.pool.Subscribe()
 278         defer s.disp.pool.Unsubscribe(ch)
 279         ok := s.disp.pool.Create(test.InstanceType(1))
 280         c.Check(ok, check.Equals, true)
 281         <-ch
 282
 283         for deadline := time.Now().Add(time.Second); time.Now().Before(deadline); {
 284                 sr = getInstances()
 285                 if len(sr.Items) > 0 {
 286                         break
 287                 }
 288                 time.Sleep(time.Millisecond)
 289         }
 290         c.Assert(len(sr.Items), check.Equals, 1)
 291         c.Check(sr.Items[0].Instance, check.Matches, "stub.*")
 292         c.Check(sr.Items[0].WorkerState, check.Equals, "booting")
 293         c.Check(sr.Items[0].Price, check.Equals, 0.123)
 294         c.Check(sr.Items[0].LastContainerUUID, check.Equals, "")
 295         c.Check(sr.Items[0].ProviderInstanceType, check.Equals, test.InstanceType(1).ProviderType)
 296         c.Check(sr.Items[0].ArvadosInstanceType, check.Equals, test.InstanceType(1).Name)
 297 }