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