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         "crypto/tls"
  10         "encoding/json"
  11         "io/ioutil"
  12         "math/rand"
  13         "net/http"
  14         "net/http/httptest"
  15         "net/url"
  16         "os"
  17         "sync"
  18         "time"
  19
  20         "git.arvados.org/arvados.git/lib/config"
  21         "git.arvados.org/arvados.git/lib/dispatchcloud/test"
  22         "git.arvados.org/arvados.git/sdk/go/arvados"
  23         "git.arvados.org/arvados.git/sdk/go/arvadostest"
  24         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  25         "github.com/prometheus/client_golang/prometheus"
  26         "golang.org/x/crypto/ssh"
  27         check "gopkg.in/check.v1"
  28 )
  29
  30 var _ = check.Suite(&DispatcherSuite{})
  31
  32 type DispatcherSuite struct {
  33         ctx            context.Context
  34         cancel         context.CancelFunc
  35         cluster        *arvados.Cluster
  36         stubDriver     *test.StubDriver
  37         disp           *dispatcher
  38         error503Server *httptest.Server
  39 }
  40
  41 func (s *DispatcherSuite) SetUpTest(c *check.C) {
  42         s.ctx, s.cancel = context.WithCancel(context.Background())
  43         s.ctx = ctxlog.Context(s.ctx, ctxlog.TestLogger(c))
  44         dispatchpub, _ := test.LoadTestKey(c, "test/sshkey_dispatch")
  45         dispatchprivraw, err := ioutil.ReadFile("test/sshkey_dispatch")
  46         c.Assert(err, check.IsNil)
  47
  48         _, hostpriv := test.LoadTestKey(c, "test/sshkey_vm")
  49         s.stubDriver = &test.StubDriver{
  50                 HostKey:                   hostpriv,
  51                 AuthorizedKeys:            []ssh.PublicKey{dispatchpub},
  52                 ErrorRateDestroy:          0.1,
  53                 MinTimeBetweenCreateCalls: time.Millisecond,
  54         }
  55
  56         // We need the postgresql connection info from the integration
  57         // test config.
  58         cfg, err := config.NewLoader(nil, ctxlog.FromContext(s.ctx)).Load()
  59         c.Assert(err, check.IsNil)
  60         testcluster, err := cfg.GetCluster("")
  61         c.Assert(err, check.IsNil)
  62
  63         s.cluster = &arvados.Cluster{
  64                 ManagementToken: "test-management-token",
  65                 PostgreSQL:      testcluster.PostgreSQL,
  66                 Containers: arvados.ContainersConfig{
  67                         CrunchRunCommand:       "crunch-run",
  68                         CrunchRunArgumentsList: []string{"--foo", "--extra='args'"},
  69                         DispatchPrivateKey:     string(dispatchprivraw),
  70                         StaleLockTimeout:       arvados.Duration(5 * time.Millisecond),
  71                         RuntimeEngine:          "stub",
  72                         MaxDispatchAttempts:    10,
  73                         CloudVMs: arvados.CloudVMsConfig{
  74                                 Driver:               "test",
  75                                 SyncInterval:         arvados.Duration(10 * time.Millisecond),
  76                                 TimeoutIdle:          arvados.Duration(150 * time.Millisecond),
  77                                 TimeoutBooting:       arvados.Duration(150 * time.Millisecond),
  78                                 TimeoutProbe:         arvados.Duration(15 * time.Millisecond),
  79                                 TimeoutShutdown:      arvados.Duration(5 * time.Millisecond),
  80                                 MaxCloudOpsPerSecond: 500,
  81                                 PollInterval:         arvados.Duration(5 * time.Millisecond),
  82                                 ProbeInterval:        arvados.Duration(5 * time.Millisecond),
  83                                 MaxProbesPerSecond:   1000,
  84                                 TimeoutSignal:        arvados.Duration(3 * time.Millisecond),
  85                                 TimeoutStaleRunLock:  arvados.Duration(3 * time.Millisecond),
  86                                 TimeoutTERM:          arvados.Duration(20 * time.Millisecond),
  87                                 ResourceTags:         map[string]string{"testtag": "test value"},
  88                                 TagKeyPrefix:         "test:",
  89                         },
  90                 },
  91                 InstanceTypes: arvados.InstanceTypeMap{
  92                         test.InstanceType(1).Name:  test.InstanceType(1),
  93                         test.InstanceType(2).Name:  test.InstanceType(2),
  94                         test.InstanceType(3).Name:  test.InstanceType(3),
  95                         test.InstanceType(4).Name:  test.InstanceType(4),
  96                         test.InstanceType(6).Name:  test.InstanceType(6),
  97                         test.InstanceType(8).Name:  test.InstanceType(8),
  98                         test.InstanceType(16).Name: test.InstanceType(16),
  99                 },
 100         }
 101         arvadostest.SetServiceURL(&s.cluster.Services.DispatchCloud, "http://localhost:/")
 102         arvadostest.SetServiceURL(&s.cluster.Services.Controller, "https://"+os.Getenv("ARVADOS_API_HOST")+"/")
 103
 104         arvClient, err := arvados.NewClientFromConfig(s.cluster)
 105         c.Assert(err, check.IsNil)
 106         // Disable auto-retry
 107         arvClient.Timeout = 0
 108
 109         s.error503Server = httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 110                 c.Logf("503 stub: returning 503")
 111                 w.WriteHeader(http.StatusServiceUnavailable)
 112         }))
 113         arvClient.Client = &http.Client{
 114                 Transport: &http.Transport{
 115                         Proxy: s.arvClientProxy(c),
 116                         TLSClientConfig: &tls.Config{
 117                                 InsecureSkipVerify: true}}}
 118
 119         s.disp = &dispatcher{
 120                 Cluster:   s.cluster,
 121                 Context:   s.ctx,
 122                 ArvClient: arvClient,
 123                 AuthToken: arvadostest.AdminToken,
 124                 Registry:  prometheus.NewRegistry(),
 125         }
 126         // Test cases can modify s.cluster before calling
 127         // initialize(), and then modify private state before calling
 128         // go run().
 129 }
 130
 131 func (s *DispatcherSuite) TearDownTest(c *check.C) {
 132         s.cancel()
 133         s.disp.Close()
 134         s.error503Server.Close()
 135 }
 136
 137 // Intercept outgoing API requests for "/503" and respond HTTP
 138 // 503. This lets us force (*arvados.Client)Last503() to return
 139 // something.
 140 func (s *DispatcherSuite) arvClientProxy(c *check.C) func(*http.Request) (*url.URL, error) {
 141         return func(req *http.Request) (*url.URL, error) {
 142                 if req.URL.Path == "/503" {
 143                         c.Logf("arvClientProxy: proxying to 503 stub")
 144                         return url.Parse(s.error503Server.URL)
 145                 } else {
 146                         return nil, nil
 147                 }
 148         }
 149 }
 150
 151 // DispatchToStubDriver checks that the dispatcher wires everything
 152 // together effectively. It uses a real scheduler and worker pool with
 153 // a fake queue and cloud driver. The fake cloud driver injects
 154 // artificial errors in order to exercise a variety of code paths.
 155 func (s *DispatcherSuite) TestDispatchToStubDriver(c *check.C) {
 156         Drivers["test"] = s.stubDriver
 157         s.disp.setupOnce.Do(s.disp.initialize)
 158         queue := &test.Queue{
 159                 MaxDispatchAttempts: 5,
 160                 ChooseType: func(ctr *arvados.Container) (arvados.InstanceType, error) {
 161                         return ChooseInstanceType(s.cluster, ctr)
 162                 },
 163                 Logger: ctxlog.TestLogger(c),
 164         }
 165         for i := 0; i < 200; i++ {
 166                 queue.Containers = append(queue.Containers, arvados.Container{
 167                         UUID:     test.ContainerUUID(i + 1),
 168                         State:    arvados.ContainerStateQueued,
 169                         Priority: int64(i%20 + 1),
 170                         RuntimeConstraints: arvados.RuntimeConstraints{
 171                                 RAM:   int64(i%3+1) << 30,
 172                                 VCPUs: i%8 + 1,
 173                         },
 174                 })
 175         }
 176         s.disp.queue = queue
 177
 178         var mtx sync.Mutex
 179         done := make(chan struct{})
 180         waiting := map[string]struct{}{}
 181         for _, ctr := range queue.Containers {
 182                 waiting[ctr.UUID] = struct{}{}
 183         }
 184         finishContainer := func(ctr arvados.Container) {
 185                 mtx.Lock()
 186                 defer mtx.Unlock()
 187                 if _, ok := waiting[ctr.UUID]; !ok {
 188                         c.Errorf("container completed twice: %s", ctr.UUID)
 189                         return
 190                 }
 191                 delete(waiting, ctr.UUID)
 192                 if len(waiting) == 100 {
 193                         // trigger scheduler maxConcurrency limit
 194                         c.Logf("test: requesting 503 in order to trigger maxConcurrency limit")
 195                         s.disp.ArvClient.RequestAndDecode(nil, "GET", "503", nil, nil)
 196                 }
 197                 if len(waiting) == 0 {
 198                         close(done)
 199                 }
 200         }
 201         executeContainer := func(ctr arvados.Container) int {
 202                 finishContainer(ctr)
 203                 return int(rand.Uint32() & 0x3)
 204         }
 205         n := 0
 206         s.stubDriver.Queue = queue
 207         s.stubDriver.SetupVM = func(stubvm *test.StubVM) {
 208                 n++
 209                 stubvm.Boot = time.Now().Add(time.Duration(rand.Int63n(int64(5 * time.Millisecond))))
 210                 stubvm.CrunchRunDetachDelay = time.Duration(rand.Int63n(int64(10 * time.Millisecond)))
 211                 stubvm.ExecuteContainer = executeContainer
 212                 stubvm.CrashRunningContainer = finishContainer
 213                 stubvm.ExtraCrunchRunArgs = "'--runtime-engine=stub' '--foo' '--extra='\\''args'\\'''"
 214                 switch n % 7 {
 215                 case 0:
 216                         stubvm.Broken = time.Now().Add(time.Duration(rand.Int63n(90)) * time.Millisecond)
 217                 case 1:
 218                         stubvm.CrunchRunMissing = true
 219                 case 2:
 220                         stubvm.ReportBroken = time.Now().Add(time.Duration(rand.Int63n(200)) * time.Millisecond)
 221                 default:
 222                         stubvm.CrunchRunCrashRate = 0.1
 223                         stubvm.ArvMountDeadlockRate = 0.1
 224                 }
 225         }
 226         s.stubDriver.Bugf = c.Errorf
 227
 228         start := time.Now()
 229         go s.disp.run()
 230         err := s.disp.CheckHealth()
 231         c.Check(err, check.IsNil)
 232
 233         for len(waiting) > 0 {
 234                 waswaiting := len(waiting)
 235                 select {
 236                 case <-done:
 237                         // loop will end because len(waiting)==0
 238                 case <-time.After(5 * time.Second):
 239                         if len(waiting) >= waswaiting {
 240                                 c.Fatalf("timed out; no progress in 5 s while waiting for %d containers: %q", len(waiting), waiting)
 241                         }
 242                 }
 243         }
 244         c.Logf("containers finished (%s), waiting for instances to shutdown and queue to clear", time.Since(start))
 245
 246         deadline := time.Now().Add(5 * time.Second)
 247         for range time.NewTicker(10 * time.Millisecond).C {
 248                 insts, err := s.stubDriver.InstanceSets()[0].Instances(nil)
 249                 c.Check(err, check.IsNil)
 250                 queue.Update()
 251                 ents, _ := queue.Entries()
 252                 if len(ents) == 0 && len(insts) == 0 {
 253                         break
 254                 }
 255                 if time.Now().After(deadline) {
 256                         c.Fatalf("timed out with %d containers (%v), %d instances (%+v)", len(ents), ents, len(insts), insts)
 257                 }
 258         }
 259
 260         req := httptest.NewRequest("GET", "/metrics", nil)
 261         req.Header.Set("Authorization", "Bearer "+s.cluster.ManagementToken)
 262         resp := httptest.NewRecorder()
 263         s.disp.ServeHTTP(resp, req)
 264         c.Check(resp.Code, check.Equals, http.StatusOK)
 265         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="Create"} [^0].*`)
 266         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="List"} [^0].*`)
 267         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="0",operation="Destroy"} [^0].*`)
 268         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="1",operation="Create"} [^0].*`)
 269         c.Check(resp.Body.String(), check.Matches, `(?ms).*driver_operations{error="1",operation="List"} 0\n.*`)
 270         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="aborted"} [0-9]+\n.*`)
 271         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="disappeared"} [^0].*`)
 272         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="failure"} [^0].*`)
 273         c.Check(resp.Body.String(), check.Matches, `(?ms).*boot_outcomes{outcome="success"} [^0].*`)
 274         c.Check(resp.Body.String(), check.Matches, `(?ms).*instances_disappeared{state="shutdown"} [^0].*`)
 275         c.Check(resp.Body.String(), check.Matches, `(?ms).*instances_disappeared{state="unknown"} 0\n.*`)
 276         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds{quantile="0.95"} [0-9.]*`)
 277         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds_count [0-9]*`)
 278         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ssh_seconds_sum [0-9.]*`)
 279         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds{quantile="0.95"} [0-9.]*`)
 280         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds_count [0-9]*`)
 281         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_to_ready_for_container_seconds_sum [0-9.]*`)
 282         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_shutdown_request_to_disappearance_seconds_count [0-9]*`)
 283         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_shutdown_request_to_disappearance_seconds_sum [0-9.]*`)
 284         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_queue_to_crunch_run_seconds_count [0-9]*`)
 285         c.Check(resp.Body.String(), check.Matches, `(?ms).*time_from_queue_to_crunch_run_seconds_sum [0-9e+.]*`)
 286         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_count{outcome="success"} [0-9]*`)
 287         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_sum{outcome="success"} [0-9e+.]*`)
 288         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_count{outcome="fail"} [0-9]*`)
 289         c.Check(resp.Body.String(), check.Matches, `(?ms).*run_probe_duration_seconds_sum{outcome="fail"} [0-9e+.]*`)
 290         c.Check(resp.Body.String(), check.Matches, `(?ms).*last_503_time [1-9][0-9e+.]*`)
 291         c.Check(resp.Body.String(), check.Matches, `(?ms).*max_concurrent_containers [1-9][0-9e+.]*`)
 292 }
 293
 294 func (s *DispatcherSuite) TestAPIPermissions(c *check.C) {
 295         s.cluster.ManagementToken = "abcdefgh"
 296         Drivers["test"] = s.stubDriver
 297         s.disp.setupOnce.Do(s.disp.initialize)
 298         s.disp.queue = &test.Queue{}
 299         go s.disp.run()
 300
 301         for _, token := range []string{"abc", ""} {
 302                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 303                 if token != "" {
 304                         req.Header.Set("Authorization", "Bearer "+token)
 305                 }
 306                 resp := httptest.NewRecorder()
 307                 s.disp.ServeHTTP(resp, req)
 308                 if token == "" {
 309                         c.Check(resp.Code, check.Equals, http.StatusUnauthorized)
 310                 } else {
 311                         c.Check(resp.Code, check.Equals, http.StatusForbidden)
 312                 }
 313         }
 314 }
 315
 316 func (s *DispatcherSuite) TestAPIDisabled(c *check.C) {
 317         s.cluster.ManagementToken = ""
 318         Drivers["test"] = s.stubDriver
 319         s.disp.setupOnce.Do(s.disp.initialize)
 320         s.disp.queue = &test.Queue{}
 321         go s.disp.run()
 322
 323         for _, token := range []string{"abc", ""} {
 324                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 325                 if token != "" {
 326                         req.Header.Set("Authorization", "Bearer "+token)
 327                 }
 328                 resp := httptest.NewRecorder()
 329                 s.disp.ServeHTTP(resp, req)
 330                 c.Check(resp.Code, check.Equals, http.StatusForbidden)
 331         }
 332 }
 333
 334 func (s *DispatcherSuite) TestInstancesAPI(c *check.C) {
 335         s.cluster.ManagementToken = "abcdefgh"
 336         s.cluster.Containers.CloudVMs.TimeoutBooting = arvados.Duration(time.Second)
 337         Drivers["test"] = s.stubDriver
 338         s.disp.setupOnce.Do(s.disp.initialize)
 339         s.disp.queue = &test.Queue{}
 340         go s.disp.run()
 341
 342         type instance struct {
 343                 Instance             string
 344                 WorkerState          string `json:"worker_state"`
 345                 Price                float64
 346                 LastContainerUUID    string `json:"last_container_uuid"`
 347                 ArvadosInstanceType  string `json:"arvados_instance_type"`
 348                 ProviderInstanceType string `json:"provider_instance_type"`
 349         }
 350         type instancesResponse struct {
 351                 Items []instance
 352         }
 353         getInstances := func() instancesResponse {
 354                 req := httptest.NewRequest("GET", "/arvados/v1/dispatch/instances", nil)
 355                 req.Header.Set("Authorization", "Bearer abcdefgh")
 356                 resp := httptest.NewRecorder()
 357                 s.disp.ServeHTTP(resp, req)
 358                 var sr instancesResponse
 359                 c.Check(resp.Code, check.Equals, http.StatusOK)
 360                 err := json.Unmarshal(resp.Body.Bytes(), &sr)
 361                 c.Check(err, check.IsNil)
 362                 return sr
 363         }
 364
 365         sr := getInstances()
 366         c.Check(len(sr.Items), check.Equals, 0)
 367
 368         ch := s.disp.pool.Subscribe()
 369         defer s.disp.pool.Unsubscribe(ch)
 370         ok := s.disp.pool.Create(test.InstanceType(1))
 371         c.Check(ok, check.Equals, true)
 372         <-ch
 373
 374         for deadline := time.Now().Add(time.Second); time.Now().Before(deadline); {
 375                 sr = getInstances()
 376                 if len(sr.Items) > 0 {
 377                         break
 378                 }
 379                 time.Sleep(time.Millisecond)
 380         }
 381         c.Assert(len(sr.Items), check.Equals, 1)
 382         c.Check(sr.Items[0].Instance, check.Matches, "inst.*")
 383         c.Check(sr.Items[0].WorkerState, check.Equals, "booting")
 384         c.Check(sr.Items[0].Price, check.Equals, 0.123)
 385         c.Check(sr.Items[0].LastContainerUUID, check.Equals, "")
 386         c.Check(sr.Items[0].ProviderInstanceType, check.Equals, test.InstanceType(1).ProviderType)
 387         c.Check(sr.Items[0].ArvadosInstanceType, check.Equals, test.InstanceType(1).Name)
 388 }