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