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.arvados.org/arvados.git/lib/dispatchcloud/test"
  19         "git.arvados.org/arvados.git/sdk/go/arvados"
  20         "git.arvados.org/arvados.git/sdk/go/arvadostest"
  21         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  22         "github.com/prometheus/client_golang/prometheus"
  23         "golang.org/x/crypto/ssh"
  24         check "gopkg.in/check.v1"
  25 )
  26
  27 var _ = check.Suite(&DispatcherSuite{})
  28
  29 type DispatcherSuite struct {
  30         ctx        context.Context
  31         cancel     context.CancelFunc
  32         cluster    *arvados.Cluster
  33         stubDriver *test.StubDriver
  34         disp       *dispatcher
  35 }
  36
  37 func (s *DispatcherSuite) SetUpTest(c *check.C) {
  38         s.ctx, s.cancel = context.WithCancel(context.Background())
  39         s.ctx = ctxlog.Context(s.ctx, ctxlog.TestLogger(c))
  40         dispatchpub, _ := test.LoadTestKey(c, "test/sshkey_dispatch")
  41         dispatchprivraw, err := ioutil.ReadFile("test/sshkey_dispatch")
  42         c.Assert(err, check.IsNil)
  43
  44         _, hostpriv := test.LoadTestKey(c, "test/sshkey_vm")
  45         s.stubDriver = &test.StubDriver{
  46                 HostKey:                   hostpriv,
  47                 AuthorizedKeys:            []ssh.PublicKey{dispatchpub},
  48                 ErrorRateDestroy:          0.1,
  49                 MinTimeBetweenCreateCalls: time.Millisecond,
  50         }
  51
  52         s.cluster = &arvados.Cluster{
  53                 ManagementToken: "test-management-token",
  54                 Containers: arvados.ContainersConfig{
  55                         CrunchRunCommand:       "crunch-run",
  56                         CrunchRunArgumentsList: []string{"--foo", "--extra='args'"},
  57                         DispatchPrivateKey:     string(dispatchprivraw),
  58                         StaleLockTimeout:       arvados.Duration(5 * time.Millisecond),
  59                         RuntimeEngine:          "stub",
  60                         CloudVMs: arvados.CloudVMsConfig{
  61                                 Driver:               "test",
  62                                 SyncInterval:         arvados.Duration(10 * time.Millisecond),
  63                                 TimeoutIdle:          arvados.Duration(150 * time.Millisecond),
  64                                 TimeoutBooting:       arvados.Duration(150 * time.Millisecond),
  65                                 TimeoutProbe:         arvados.Duration(15 * time.Millisecond),
  66                                 TimeoutShutdown:      arvados.Duration(5 * time.Millisecond),
  67                                 MaxCloudOpsPerSecond: 500,
  68                                 PollInterval:         arvados.Duration(5 * time.Millisecond),
  69                                 ProbeInterval:        arvados.Duration(5 * time.Millisecond),
  70                                 MaxProbesPerSecond:   1000,
  71                                 TimeoutSignal:        arvados.Duration(3 * time.Millisecond),
  72                                 TimeoutStaleRunLock:  arvados.Duration(3 * time.Millisecond),
  73                                 TimeoutTERM:          arvados.Duration(20 * time.Millisecond),
  74                                 ResourceTags:         map[string]string{"testtag": "test value"},
  75                                 TagKeyPrefix:         "test:",
  76                         },
  77                 },
  78                 InstanceTypes: arvados.InstanceTypeMap{
  79                         test.InstanceType(1).Name:  test.InstanceType(1),
  80                         test.InstanceType(2).Name:  test.InstanceType(2),
  81                         test.InstanceType(3).Name:  test.InstanceType(3),
  82                         test.InstanceType(4).Name:  test.InstanceType(4),
  83                         test.InstanceType(6).Name:  test.InstanceType(6),
  84                         test.InstanceType(8).Name:  test.InstanceType(8),
  85                         test.InstanceType(16).Name: test.InstanceType(16),
  86                 },
  87         }
  88         arvadostest.SetServiceURL(&s.cluster.Services.DispatchCloud, "http://localhost:/")
  89         arvadostest.SetServiceURL(&s.cluster.Services.Controller, "https://"+os.Getenv("ARVADOS_API_HOST")+"/")
  90
  91         arvClient, err := arvados.NewClientFromConfig(s.cluster)
  92         c.Check(err, check.IsNil)
  93
  94         s.disp = &dispatcher{
  95                 Cluster:   s.cluster,
  96                 Context:   s.ctx,
  97                 ArvClient: arvClient,
  98                 AuthToken: arvadostest.AdminToken,
  99                 Registry:  prometheus.NewRegistry(),
 100         }
 101         // Test cases can modify s.cluster before calling
 102         // initialize(), and then modify private state before calling
 103         // go run().
 104 }
 105
 106 func (s *DispatcherSuite) TearDownTest(c *check.C) {
 107         s.cancel()
 108         s.disp.Close()
 109 }
 110
 111 // DispatchToStubDriver checks that the dispatcher wires everything
 112 // together effectively. It uses a real scheduler and worker pool with
 113 // a fake queue and cloud driver. The fake cloud driver injects
 114 // artificial errors in order to exercise a variety of code paths.
 115 func (s *DispatcherSuite) TestDispatchToStubDriver(c *check.C) {
 116         Drivers["test"] = s.stubDriver
 117         s.disp.setupOnce.Do(s.disp.initialize)
 118         queue := &test.Queue{
 119                 ChooseType: func(ctr *arvados.Container) (arvados.InstanceType, error) {
 120                         return ChooseInstanceType(s.cluster, ctr)
 121                 },
 122                 Logger: ctxlog.TestLogger(c),
 123         }
 124         for i := 0; i < 200; i++ {
 125                 queue.Containers = append(queue.Containers, arvados.Container{
 126                         UUID:     test.ContainerUUID(i + 1),
 127                         State:    arvados.ContainerStateQueued,
 128                         Priority: int64(i%20 + 1),
 129                         RuntimeConstraints: arvados.RuntimeConstraints{
 130                                 RAM:   int64(i%3+1) << 30,
 131                                 VCPUs: i%8 + 1,
 132                         },
 133                 })
 134         }
 135         s.disp.queue = queue
 136
 137         var mtx sync.Mutex
 138         done := make(chan struct{})
 139         waiting := map[string]struct{}{}
 140         for _, ctr := range queue.Containers {
 141                 waiting[ctr.UUID] = struct{}{}
 142         }
 143         finishContainer := func(ctr arvados.Container) {
 144                 mtx.Lock()
 145                 defer mtx.Unlock()
 146                 if _, ok := waiting[ctr.UUID]; !ok {
 147                         c.Errorf("container completed twice: %s", ctr.UUID)
 148                         return
 149                 }
 150                 delete(waiting, ctr.UUID)
 151                 if len(waiting) == 0 {
 152                         close(done)
 153                 }
 154         }
 155         executeContainer := func(ctr arvados.Container) int {
 156                 finishContainer(ctr)
 157                 return int(rand.Uint32() & 0x3)
 158         }
 159         n := 0
 160         s.stubDriver.Queue = queue
 161         s.stubDriver.SetupVM = func(stubvm *test.StubVM) {
 162                 n++
 163                 stubvm.Boot = time.Now().Add(time.Duration(rand.Int63n(int64(5 * time.Millisecond))))
 164                 stubvm.CrunchRunDetachDelay = time.Duration(rand.Int63n(int64(10 * time.Millisecond)))
 165                 stubvm.ExecuteContainer = executeContainer
 166                 stubvm.CrashRunningContainer = finishContainer
 167                 stubvm.ExtraCrunchRunArgs = "'--runtime-engine=stub' '--foo' '--extra='\\''args'\\'''"
 168                 switch n % 7 {
 169                 case 0:
 170                         stubvm.Broken = time.Now().Add(time.Duration(rand.Int63n(90)) * time.Millisecond)
 171                 case 1:
 172                         stubvm.CrunchRunMissing = true
 173                 case 2:
 174                         stubvm.ReportBroken = time.Now().Add(time.Duration(rand.Int63n(200)) * time.Millisecond)
 175                 default:
 176                         stubvm.CrunchRunCrashRate = 0.1
 177                         stubvm.ArvMountDeadlockRate = 0.1
 178                 }
 179         }
 180         s.stubDriver.Bugf = c.Errorf
 181
 182         start := time.Now()
 183         go s.disp.run()
 184         err := s.disp.CheckHealth()
 185         c.Check(err, check.IsNil)
 186
 187         select {
 188         case <-done:
 189                 c.Logf("containers finished (%s), waiting for instances to shutdown and queue to clear", time.Since(start))
 190         case <-time.After(10 * time.Second):
 191                 c.Fatalf("timed out; still waiting for %d containers: %q", len(waiting), waiting)
 192         }
 193
 194         deadline := time.Now().Add(5 * time.Second)
 195         for range time.NewTicker(10 * time.Millisecond).C {
 196                 insts, err := s.stubDriver.InstanceSets()[0].Instances(nil)
 197                 c.Check(err, check.IsNil)
 198                 queue.Update()
 199                 ents, _ := queue.Entries()
 200                 if len(ents) == 0 && len(insts) == 0 {
 201                         break
 202                 }
 203                 if time.Now().After(deadline) {
 204                         c.Fatalf("timed out with %d containers (%v), %d instances (%+v)", len(ents), ents, len(insts), insts)
 205                 }
 206         }
 207
 208         req := httptest.NewRequest("GET", "/metrics", nil)
 209         req.Header.Set("Authorization", "Bearer "+s.cluster.ManagementToken)
 210         resp := httptest.NewRecorder()
 211         s.disp.ServeHTTP(resp, req)
 212         c.Check(resp.Code, check.Equals, http.StatusOK)
 213         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="Create"} [^0].*`)
 214         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="List"} [^0].*`)
 215         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="Destroy"} [^0].*`)
 216         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="1",operation="Create"} [^0].*`)
 217         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="1",operation="List"} 0\n.*`)
 218         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="aborted"} 0.*`)
 219         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="disappeared"} [^0].*`)
 220         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="failure"} [^0].*`)
 221         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="success"} [^0].*`)
 222         c.Check(resp.Body.String(), check.Matches, `(?ms).*instances_disappeared{state="shutdown"} [^0].*`)
 223         c.Check(resp.Body.String(), check.Matches, `(?ms).*instances_disappeared{state="unknown"} 0\n.*`)
 224         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds{quantile="0.95"} [0-9.]*`)
 225         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds_count [0-9]*`)
 226         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds_sum [0-9.]*`)
 227         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds{quantile="0.95"} [0-9.]*`)
 228         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds_count [0-9]*`)
 229         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds_sum [0-9.]*`)
 230         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_shutdown_request_to_disappearance_seconds_count [0-9]*`)
 231         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_shutdown_request_to_disappearance_seconds_sum [0-9.]*`)
 232         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_queue_to_crunch_run_seconds_count [0-9]*`)
 233         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_queue_to_crunch_run_seconds_sum [0-9e+.]*`)
 234         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_count{outcome="success"} [0-9]*`)
 235         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_sum{outcome="success"} [0-9e+.]*`)
 236         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_count{outcome="fail"} [0-9]*`)
 237         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_sum{outcome="fail"} [0-9e+.]*`)
 238 }
 239
 240 func (s *DispatcherSuite) TestAPIPermissions(c *check.C) {
 241         s.cluster.ManagementToken = "abcdefgh"
 242         Drivers["test"] = s.stubDriver
 243         s.disp.setupOnce.Do(s.disp.initialize)
 244         s.disp.queue = &test.Queue{}
 245         go s.disp.run()
 246
 247         for _, token := range []string{"abc", ""} {
 248                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 249                 if token != "" {
 250                         req.Header.Set("Authorization", "Bearer "+token)
 251                 }
 252                 resp := httptest.NewRecorder()
 253                 s.disp.ServeHTTP(resp, req)
 254                 if token == "" {
 255                         c.Check(resp.Code, check.Equals, http.StatusUnauthorized)
 256                 } else {
 257                         c.Check(resp.Code, check.Equals, http.StatusForbidden)
 258                 }
 259         }
 260 }
 261
 262 func (s *DispatcherSuite) TestAPIDisabled(c *check.C) {
 263         s.cluster.ManagementToken = ""
 264         Drivers["test"] = s.stubDriver
 265         s.disp.setupOnce.Do(s.disp.initialize)
 266         s.disp.queue = &test.Queue{}
 267         go s.disp.run()
 268
 269         for _, token := range []string{"abc", ""} {
 270                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 271                 if token != "" {
 272                         req.Header.Set("Authorization", "Bearer "+token)
 273                 }
 274                 resp := httptest.NewRecorder()
 275                 s.disp.ServeHTTP(resp, req)
 276                 c.Check(resp.Code, check.Equals, http.StatusForbidden)
 277         }
 278 }
 279
 280 func (s *DispatcherSuite) TestInstancesAPI(c *check.C) {
 281         s.cluster.ManagementToken = "abcdefgh"
 282         s.cluster.Containers.CloudVMs.TimeoutBooting = arvados.Duration(time.Second)
 283         Drivers["test"] = s.stubDriver
 284         s.disp.setupOnce.Do(s.disp.initialize)
 285         s.disp.queue = &test.Queue{}
 286         go s.disp.run()
 287
 288         type instance struct {
 289                 Instance             string
 290                 WorkerState          string `json:"worker_state"`
 291                 Price                float64
 292                 LastContainerUUID    string `json:"last_container_uuid"`
 293                 ArvadosInstanceType  string `json:"arvados_instance_type"`
 294                 ProviderInstanceType string `json:"provider_instance_type"`
 295         }
 296         type instancesResponse struct {
 297                 Items []instance
 298         }
 299         getInstances := func() instancesResponse {
 300                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 301                 req.Header.Set("Authorization", "Bearer abcdefgh")
 302                 resp := httptest.NewRecorder()
 303                 s.disp.ServeHTTP(resp, req)
 304                 var sr instancesResponse
 305                 c.Check(resp.Code, check.Equals, http.StatusOK)
 306                 err := json.Unmarshal(resp.Body.Bytes(), &sr)
 307                 c.Check(err, check.IsNil)
 308                 return sr
 309         }
 310
 311         sr := getInstances()
 312         c.Check(len(sr.Items), check.Equals, 0)
 313
 314         ch := s.disp.pool.Subscribe()
 315         defer s.disp.pool.Unsubscribe(ch)
 316         ok := s.disp.pool.Create(test.InstanceType(1))
 317         c.Check(ok, check.Equals, true)
 318         <-ch
 319
 320         for deadline := time.Now().Add(time.Second); time.Now().Before(deadline); {
 321                 sr = getInstances()
 322                 if len(sr.Items) > 0 {
 323                         break
 324                 }
 325                 time.Sleep(time.Millisecond)
 326         }
 327         c.Assert(len(sr.Items), check.Equals, 1)
 328         c.Check(sr.Items[0].Instance, check.Matches, "inst.*")
 329         c.Check(sr.Items[0].WorkerState, check.Equals, "booting")
 330         c.Check(sr.Items[0].Price, check.Equals, 0.123)
 331         c.Check(sr.Items[0].LastContainerUUID, check.Equals, "")
 332         c.Check(sr.Items[0].ProviderInstanceType, check.Equals, test.InstanceType(1).ProviderType)
 333         c.Check(sr.Items[0].ArvadosInstanceType, check.Equals, test.InstanceType(1).Name)
 334 }