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