@BeforeClass
public static void setupJobManager() {
Configuration config = new Configuration();
int port = NetUtils.getAvailablePort();
config.setString(ConfigConstants.JOB_MANAGER_IPC_ADDRESS_KEY, "localhost");
config.setInteger(ConfigConstants.JOB_MANAGER_IPC_PORT_KEY, port);
scala.Option<Tuple2<String, Object>> listeningAddress =
scala.Option.apply(new Tuple2<String, Object>("localhost", port));
jobManagerSystem = AkkaUtils.createActorSystem(config, listeningAddress);
ActorRef jobManagerActorRef =
JobManager.startJobManagerActors(
config,
jobManagerSystem,
StreamingMode.BATCH_ONLY,
JobManager.class,
MemoryArchivist.class)
._1();
try {
LeaderRetrievalService lrs = LeaderRetrievalUtils.createLeaderRetrievalService(config);
jmGateway = LeaderRetrievalUtils.retrieveLeaderGateway(lrs, jobManagerSystem, timeout);
} catch (Exception e) {
fail("Could not retrieve the JobManager gateway. " + e.getMessage());
}
}
private FiniteDuration getTimeout() {
final Configuration configuration = GlobalConfiguration.getConfiguration();
if (this.timeout != null) {
configuration.setString(ConfigConstants.AKKA_ASK_TIMEOUT, this.timeout);
}
return AkkaUtils.getClientTimeout(configuration);
}
private ActorRef getJobManager() throws IOException {
final Configuration configuration = GlobalConfiguration.getConfiguration();
ActorSystem actorSystem;
try {
final scala.Tuple2<String, Object> systemEndpoint = new scala.Tuple2<String, Object>("", 0);
actorSystem =
AkkaUtils.createActorSystem(
configuration, new Some<scala.Tuple2<String, Object>>(systemEndpoint));
} catch (final Exception e) {
throw new RuntimeException("Could not start actor system to communicate with JobManager", e);
}
return JobManager.getJobManagerActorRef(
new InetSocketAddress(this.jobManagerHost, this.jobManagerPort),
actorSystem,
AkkaUtils.getLookupTimeout(configuration));
}
@BeforeClass
public static void startActorSystem() {
config = new Configuration();
config.setString(ConfigConstants.AKKA_ASK_TIMEOUT, "5 s");
config.setString(ConfigConstants.AKKA_WATCH_HEARTBEAT_INTERVAL, "200 ms");
config.setString(ConfigConstants.AKKA_WATCH_HEARTBEAT_PAUSE, "2 s");
config.setDouble(ConfigConstants.AKKA_WATCH_THRESHOLD, 2.0);
actorSystem = AkkaUtils.createLocalActorSystem(config);
}
private Map<ExecutionAttemptID, Execution> setupExecution(
JobVertex v1, int dop1, JobVertex v2, int dop2) throws Exception {
final JobID jobId = new JobID();
v1.setParallelism(dop1);
v2.setParallelism(dop2);
v1.setInvokableClass(BatchTask.class);
v2.setInvokableClass(BatchTask.class);
// execution graph that executes actions synchronously
ExecutionGraph eg =
new ExecutionGraph(
TestingUtils.directExecutionContext(),
jobId,
"some job",
new Configuration(),
new SerializedValue<>(new ExecutionConfig()),
AkkaUtils.getDefaultTimeout(),
new NoRestartStrategy());
eg.setQueuedSchedulingAllowed(false);
List<JobVertex> ordered = Arrays.asList(v1, v2);
eg.attachJobGraph(ordered);
Scheduler scheduler = new Scheduler(TestingUtils.defaultExecutionContext());
for (int i = 0; i < dop1 + dop2; i++) {
scheduler.newInstanceAvailable(
ExecutionGraphTestUtils.getInstance(
new ExecutionGraphTestUtils.SimpleActorGateway(
TestingUtils.directExecutionContext())));
}
assertEquals(dop1 + dop2, scheduler.getNumberOfAvailableSlots());
// schedule, this triggers mock deployment
eg.scheduleForExecution(scheduler);
Map<ExecutionAttemptID, Execution> executions = eg.getRegisteredExecutions();
assertEquals(dop1 + dop2, executions.size());
return executions;
}
/**
* Create a new Flink on YARN cluster.
*
* @param yarnClient
* @param appId the YARN application ID
* @param hadoopConfig
* @param flinkConfig
* @param sessionFilesDir
* @param detached Set to true if no actor system or RPC communication with the cluster should be
* established
* @throws IOException
* @throws YarnException
*/
public FlinkYarnCluster(
final YarnClient yarnClient,
final ApplicationId appId,
Configuration hadoopConfig,
org.apache.flink.configuration.Configuration flinkConfig,
Path sessionFilesDir,
boolean detached)
throws IOException, YarnException {
this.akkaDuration = AkkaUtils.getTimeout(flinkConfig);
this.akkaTimeout = Timeout.durationToTimeout(akkaDuration);
this.yarnClient = yarnClient;
this.hadoopConfig = hadoopConfig;
this.sessionFilesDir = sessionFilesDir;
this.applicationId = appId;
this.detached = detached;
this.flinkConfig = flinkConfig;
this.appId = appId;
// get one application report manually
intialAppReport = yarnClient.getApplicationReport(appId);
String jobManagerHost = intialAppReport.getHost();
int jobManagerPort = intialAppReport.getRpcPort();
this.jobManagerAddress = new InetSocketAddress(jobManagerHost, jobManagerPort);
}
public WebRuntimeMonitor(
Configuration config, LeaderRetrievalService leaderRetrievalService, ActorSystem actorSystem)
throws IOException, InterruptedException {
this.leaderRetrievalService = checkNotNull(leaderRetrievalService);
this.timeout = AkkaUtils.getTimeout(config);
this.retriever =
new JobManagerRetriever(this, actorSystem, AkkaUtils.getTimeout(config), timeout);
final WebMonitorConfig cfg = new WebMonitorConfig(config);
final int configuredPort = cfg.getWebFrontendPort();
if (configuredPort < 0) {
throw new IllegalArgumentException("Web frontend port is invalid: " + configuredPort);
}
final WebMonitorUtils.LogFileLocation logFiles = WebMonitorUtils.LogFileLocation.find(config);
// create an empty directory in temp for the web server
String rootDirFileName = "flink-web-" + UUID.randomUUID();
webRootDir = new File(getBaseDir(config), rootDirFileName);
LOG.info("Using directory {} for the web interface files", webRootDir);
final boolean webSubmitAllow = cfg.isProgramSubmitEnabled();
if (webSubmitAllow) {
// create storage for uploads
String uploadDirName = "flink-web-upload-" + UUID.randomUUID();
this.uploadDir = new File(getBaseDir(config), uploadDirName);
if (!uploadDir.mkdir() || !uploadDir.canWrite()) {
throw new IOException("Unable to create temporary directory to support jar uploads.");
}
LOG.info("Using directory {} for web frontend JAR file uploads", uploadDir);
} else {
this.uploadDir = null;
}
ExecutionGraphHolder currentGraphs = new ExecutionGraphHolder();
// - Back pressure stats ----------------------------------------------
stackTraceSamples = new StackTraceSampleCoordinator(actorSystem, 60000);
// Back pressure stats tracker config
int cleanUpInterval =
config.getInteger(
ConfigConstants.JOB_MANAGER_WEB_BACK_PRESSURE_CLEAN_UP_INTERVAL,
ConfigConstants.DEFAULT_JOB_MANAGER_WEB_BACK_PRESSURE_CLEAN_UP_INTERVAL);
int refreshInterval =
config.getInteger(
ConfigConstants.JOB_MANAGER_WEB_BACK_PRESSURE_REFRESH_INTERVAL,
ConfigConstants.DEFAULT_JOB_MANAGER_WEB_BACK_PRESSURE_REFRESH_INTERVAL);
int numSamples =
config.getInteger(
ConfigConstants.JOB_MANAGER_WEB_BACK_PRESSURE_NUM_SAMPLES,
ConfigConstants.DEFAULT_JOB_MANAGER_WEB_BACK_PRESSURE_NUM_SAMPLES);
int delay =
config.getInteger(
ConfigConstants.JOB_MANAGER_WEB_BACK_PRESSURE_DELAY,
ConfigConstants.DEFAULT_JOB_MANAGER_WEB_BACK_PRESSURE_DELAY);
FiniteDuration delayBetweenSamples = new FiniteDuration(delay, TimeUnit.MILLISECONDS);
backPressureStatsTracker =
new BackPressureStatsTracker(
stackTraceSamples, cleanUpInterval, numSamples, delayBetweenSamples);
// --------------------------------------------------------------------
executorService = new ForkJoinPool();
ExecutionContextExecutor context = ExecutionContext$.MODULE$.fromExecutor(executorService);
router =
new Router()
// config how to interact with this web server
.GET("/config", handler(new DashboardConfigHandler(cfg.getRefreshInterval())))
// the overview - how many task managers, slots, free slots, ...
.GET("/overview", handler(new ClusterOverviewHandler(DEFAULT_REQUEST_TIMEOUT)))
// job manager configuration
.GET("/jobmanager/config", handler(new JobManagerConfigHandler(config)))
// overview over jobs
.GET(
"/joboverview",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, true, true)))
.GET(
"/joboverview/running",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, true, false)))
.GET(
"/joboverview/completed",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, false, true)))
.GET("/jobs", handler(new CurrentJobIdsHandler(DEFAULT_REQUEST_TIMEOUT)))
.GET("/jobs/:jobid", handler(new JobDetailsHandler(currentGraphs)))
.GET("/jobs/:jobid/vertices", handler(new JobDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid",
handler(new JobVertexDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasktimes",
handler(new SubtasksTimesHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/taskmanagers",
handler(new JobVertexTaskManagersHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/accumulators",
handler(new JobVertexAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/checkpoints",
handler(new JobVertexCheckpointsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/backpressure",
handler(
new JobVertexBackPressureHandler(
currentGraphs, backPressureStatsTracker, refreshInterval)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/accumulators",
handler(new SubtasksAllAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum",
handler(new SubtaskCurrentAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt",
handler(new SubtaskExecutionAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt/accumulators",
handler(new SubtaskExecutionAttemptAccumulatorsHandler(currentGraphs)))
.GET("/jobs/:jobid/plan", handler(new JobPlanHandler(currentGraphs)))
.GET("/jobs/:jobid/config", handler(new JobConfigHandler(currentGraphs)))
.GET("/jobs/:jobid/exceptions", handler(new JobExceptionsHandler(currentGraphs)))
.GET("/jobs/:jobid/accumulators", handler(new JobAccumulatorsHandler(currentGraphs)))
.GET("/jobs/:jobid/checkpoints", handler(new JobCheckpointsHandler(currentGraphs)))
.GET("/taskmanagers", handler(new TaskManagersHandler(DEFAULT_REQUEST_TIMEOUT)))
.GET(
"/taskmanagers/:" + TaskManagersHandler.TASK_MANAGER_ID_KEY + "/metrics",
handler(new TaskManagersHandler(DEFAULT_REQUEST_TIMEOUT)))
.GET(
"/taskmanagers/:" + TaskManagersHandler.TASK_MANAGER_ID_KEY + "/log",
new TaskManagerLogHandler(
retriever,
context,
jobManagerAddressPromise.future(),
timeout,
TaskManagerLogHandler.FileMode.LOG,
config))
.GET(
"/taskmanagers/:" + TaskManagersHandler.TASK_MANAGER_ID_KEY + "/stdout",
new TaskManagerLogHandler(
retriever,
context,
jobManagerAddressPromise.future(),
timeout,
TaskManagerLogHandler.FileMode.STDOUT,
config))
// log and stdout
.GET(
"/jobmanager/log",
logFiles.logFile == null
? new ConstantTextHandler("(log file unavailable)")
: new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, logFiles.logFile))
.GET(
"/jobmanager/stdout",
logFiles.stdOutFile == null
? new ConstantTextHandler("(stdout file unavailable)")
: new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, logFiles.stdOutFile))
// Cancel a job via GET (for proper integration with YARN this has to be performed via
// GET)
.GET("/jobs/:jobid/yarn-cancel", handler(new JobCancellationHandler()))
// DELETE is the preferred way of canceling a job (Rest-conform)
.DELETE("/jobs/:jobid/cancel", handler(new JobCancellationHandler()))
// stop a job via GET (for proper integration with YARN this has to be performed via
// GET)
.GET("/jobs/:jobid/yarn-stop", handler(new JobStoppingHandler()))
// DELETE is the preferred way of stopping a job (Rest-conform)
.DELETE("/jobs/:jobid/stop", handler(new JobStoppingHandler()));
if (webSubmitAllow) {
router
// fetch the list of uploaded jars.
.GET("/jars", handler(new JarListHandler(uploadDir)))
// get plan for an uploaded jar
.GET("/jars/:jarid/plan", handler(new JarPlanHandler(uploadDir)))
// run a jar
.POST("/jars/:jarid/run", handler(new JarRunHandler(uploadDir, timeout)))
// upload a jar
.POST("/jars/upload", handler(new JarUploadHandler(uploadDir)))
// delete an uploaded jar from submission interface
.DELETE("/jars/:jarid", handler(new JarDeleteHandler(uploadDir)));
} else {
router
// send an Access Denied message (sort of)
// Every other GET request will go to the File Server, which will not provide
// access to the jar directory anyway, because it doesn't exist in webRootDir.
.GET("/jars", handler(new JarAccessDeniedHandler()));
}
// this handler serves all the static contents
router.GET(
"/:*",
new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, webRootDir));
// add shutdown hook for deleting the directories and remaining temp files on shutdown
try {
Runtime.getRuntime()
.addShutdownHook(
new Thread() {
@Override
public void run() {
cleanup();
}
});
} catch (IllegalStateException e) {
// race, JVM is in shutdown already, we can safely ignore this
LOG.debug("Unable to add shutdown hook, shutdown already in progress", e);
} catch (Throwable t) {
// these errors usually happen when the shutdown is already in progress
LOG.warn("Error while adding shutdown hook", t);
}
ChannelInitializer<SocketChannel> initializer =
new ChannelInitializer<SocketChannel>() {
@Override
protected void initChannel(SocketChannel ch) {
Handler handler = new Handler(router);
ch.pipeline()
.addLast(new HttpServerCodec())
.addLast(new HttpRequestHandler(uploadDir))
.addLast(handler.name(), handler)
.addLast(new PipelineErrorHandler(LOG));
}
};
NioEventLoopGroup bossGroup = new NioEventLoopGroup(1);
NioEventLoopGroup workerGroup = new NioEventLoopGroup();
this.bootstrap = new ServerBootstrap();
this.bootstrap
.group(bossGroup, workerGroup)
.channel(NioServerSocketChannel.class)
.childHandler(initializer);
Channel ch = this.bootstrap.bind(configuredPort).sync().channel();
this.serverChannel = ch;
InetSocketAddress bindAddress = (InetSocketAddress) ch.localAddress();
String address = bindAddress.getAddress().getHostAddress();
int port = bindAddress.getPort();
LOG.info("Web frontend listening at " + address + ':' + port);
}
public WebRuntimeMonitor(
Configuration config, LeaderRetrievalService leaderRetrievalService, ActorSystem actorSystem)
throws IOException, InterruptedException {
this.leaderRetrievalService = checkNotNull(leaderRetrievalService);
final WebMonitorConfig cfg = new WebMonitorConfig(config);
// create an empty directory in temp for the web server
String fileName = String.format("flink-web-%s", UUID.randomUUID().toString());
webRootDir = new File(System.getProperty("java.io.tmpdir"), fileName);
LOG.info("Using directory {} for the web interface files", webRootDir);
// figure out where our logs are
final String flinkRoot = config.getString(ConfigConstants.FLINK_BASE_DIR_PATH_KEY, null);
final String defaultLogDirectory = flinkRoot + "/log";
final String logDirectories =
config.getString(ConfigConstants.JOB_MANAGER_WEB_LOG_PATH_KEY, defaultLogDirectory);
// find out which directory holds the path for log and stdout
final ArrayList<String> logPaths = new ArrayList<>();
final ArrayList<String> outPaths = new ArrayList<>();
// yarn allows for multiple log directories. Search in all.
for (String paths : logDirectories.split(",")) {
File dir = new File(paths);
if (dir.exists() && dir.isDirectory() && dir.canRead()) {
if (dir.listFiles(LOG_FILE_PATTERN).length == 1) {
logPaths.add(paths);
}
if (dir.listFiles(STDOUT_FILE_PATTERN).length == 1) {
outPaths.add(paths);
}
}
}
// we don't want any ambiguities. There must be only one log and out file.
if (logPaths.size() != 1 || outPaths.size() != 1) {
throw new IllegalConfigurationException(
"The path to the log and out files (" + logDirectories + ") is not valid.");
}
final File logDir = new File(logPaths.get(0));
final File outDir = new File(outPaths.get(0));
LOG.info("Serving job manager logs from {}", logDir.getAbsolutePath());
LOG.info("Serving job manager stdout from {}", outDir.getAbsolutePath());
// port configuration
this.configuredPort = cfg.getWebFrontendPort();
if (this.configuredPort < 0) {
throw new IllegalArgumentException("Web frontend port is invalid: " + this.configuredPort);
}
timeout = AkkaUtils.getTimeout(config);
FiniteDuration lookupTimeout = AkkaUtils.getTimeout(config);
retriever = new JobManagerRetriever(this, actorSystem, lookupTimeout, timeout);
ExecutionGraphHolder currentGraphs = new ExecutionGraphHolder();
router =
new Router()
// config how to interact with this web server
.GET("/config", handler(new DashboardConfigHandler(cfg.getRefreshInterval())))
// the overview - how many task managers, slots, free slots, ...
.GET("/overview", handler(new ClusterOverviewHandler(DEFAULT_REQUEST_TIMEOUT)))
// job manager configuration, log and stdout
.GET("/jobmanager/config", handler(new JobManagerConfigHandler(config)))
// overview over jobs
.GET(
"/joboverview",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, true, true)))
.GET(
"/joboverview/running",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, true, false)))
.GET(
"/joboverview/completed",
handler(new CurrentJobsOverviewHandler(DEFAULT_REQUEST_TIMEOUT, false, true)))
.GET("/jobs", handler(new CurrentJobIdsHandler(retriever, DEFAULT_REQUEST_TIMEOUT)))
.GET("/jobs/:jobid", handler(new JobDetailsHandler(currentGraphs)))
.GET("/jobs/:jobid/vertices", handler(new JobDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid",
handler(new JobVertexDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasktimes",
handler(new SubtasksTimesHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/accumulators",
handler(new JobVertexAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/accumulators",
handler(new SubtasksAllAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum",
handler(new SubtaskCurrentAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt",
handler(new SubtaskExecutionAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt/accumulators",
handler(new SubtaskExecutionAttemptAccumulatorsHandler(currentGraphs)))
.GET("/jobs/:jobid/plan", handler(new JobPlanHandler(currentGraphs)))
.GET("/jobs/:jobid/config", handler(new JobConfigHandler(currentGraphs)))
.GET("/jobs/:jobid/exceptions", handler(new JobExceptionsHandler(currentGraphs)))
.GET("/jobs/:jobid/accumulators", handler(new JobAccumulatorsHandler(currentGraphs)))
.GET("/taskmanagers", handler(new TaskManagersHandler(DEFAULT_REQUEST_TIMEOUT)))
.GET(
"/taskmanagers/:" + TaskManagersHandler.TASK_MANAGER_ID_KEY,
handler(new TaskManagersHandler(DEFAULT_REQUEST_TIMEOUT)))
.GET(
"/jobmanager/log",
new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, logDir))
.GET(
"/jobmanager/stdout",
new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, outDir))
// this handler serves all the static contents
.GET(
"/:*",
new StaticFileServerHandler(
retriever, jobManagerAddressPromise.future(), timeout, webRootDir));
synchronized (startupShutdownLock) {
// add shutdown hook for deleting the directory
try {
Runtime.getRuntime()
.addShutdownHook(
new Thread() {
@Override
public void run() {
shutdown();
}
});
} catch (IllegalStateException e) {
// race, JVM is in shutdown already, we can safely ignore this
LOG.debug("Unable to add shutdown hook, shutdown already in progress", e);
} catch (Throwable t) {
// these errors usually happen when the shutdown is already in progress
LOG.warn("Error while adding shutdown hook", t);
}
ChannelInitializer<SocketChannel> initializer =
new ChannelInitializer<SocketChannel>() {
@Override
protected void initChannel(SocketChannel ch) {
Handler handler = new Handler(router);
ch.pipeline()
.addLast(new HttpServerCodec())
.addLast(new HttpObjectAggregator(65536))
.addLast(new ChunkedWriteHandler())
.addLast(handler.name(), handler);
}
};
NioEventLoopGroup bossGroup = new NioEventLoopGroup(1);
NioEventLoopGroup workerGroup = new NioEventLoopGroup();
this.bootstrap = new ServerBootstrap();
this.bootstrap
.group(bossGroup, workerGroup)
.channel(NioServerSocketChannel.class)
.childHandler(initializer);
Channel ch = this.bootstrap.bind(configuredPort).sync().channel();
this.serverChannel = ch;
InetSocketAddress bindAddress = (InetSocketAddress) ch.localAddress();
String address = bindAddress.getAddress().getHostAddress();
int port = bindAddress.getPort();
LOG.info("Web frontend listening at " + address + ':' + port);
}
}
public WebRuntimeMonitor(
Configuration config, LeaderRetrievalService leaderRetrievalService, ActorSystem actorSystem)
throws IOException {
this.leaderRetrievalService = checkNotNull(leaderRetrievalService);
final WebMonitorConfig cfg = new WebMonitorConfig(config);
// create an empty directory in temp for the web server
String fileName = String.format("flink-web-%s", UUID.randomUUID().toString());
webRootDir = new File(System.getProperty("java.io.tmpdir"), fileName);
LOG.info("Using directory {} for the web interface files", webRootDir);
// port configuration
this.configuredPort = cfg.getWebFrontendPort();
if (this.configuredPort < 0) {
throw new IllegalArgumentException("Web frontend port is invalid: " + this.configuredPort);
}
FiniteDuration timeout = AkkaUtils.getTimeout(config);
FiniteDuration lookupTimeout = AkkaUtils.getTimeout(config);
retriever = new JobManagerArchiveRetriever(this, actorSystem, lookupTimeout, timeout);
ExecutionGraphHolder currentGraphs = new ExecutionGraphHolder(retriever);
router =
new Router()
// config how to interact with this web server
.GET("/config", handler(new DashboardConfigHandler(cfg.getRefreshInterval())))
// the overview - how many task managers, slots, free slots, ...
.GET(
"/overview",
handler(new ClusterOverviewHandler(retriever, DEFAULT_REQUEST_TIMEOUT)))
// job manager configuration
.GET("/jobmanager/config", handler(new JobManagerConfigHandler(config)))
// overview over jobs
.GET(
"/joboverview",
handler(
new CurrentJobsOverviewHandler(retriever, DEFAULT_REQUEST_TIMEOUT, true, true)))
.GET(
"/joboverview/running",
handler(
new CurrentJobsOverviewHandler(
retriever, DEFAULT_REQUEST_TIMEOUT, true, false)))
.GET(
"/joboverview/completed",
handler(
new CurrentJobsOverviewHandler(
retriever, DEFAULT_REQUEST_TIMEOUT, false, true)))
.GET("/jobs", handler(new CurrentJobIdsHandler(retriever, DEFAULT_REQUEST_TIMEOUT)))
.GET("/jobs/:jobid", handler(new JobDetailsHandler(currentGraphs)))
.GET("/jobs/:jobid/vertices", handler(new JobDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid",
handler(new JobVertexDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasktimes",
handler(new SubtasksTimesHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/accumulators",
handler(new JobVertexAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/accumulators",
handler(new SubtasksAllAccumulatorsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum",
handler(new SubtaskCurrentAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt",
handler(new SubtaskExecutionAttemptDetailsHandler(currentGraphs)))
.GET(
"/jobs/:jobid/vertices/:vertexid/subtasks/:subtasknum/attempts/:attempt/accumulators",
handler(new SubtaskExecutionAttemptAccumulatorsHandler(currentGraphs)))
.GET("/jobs/:jobid/plan", handler(new JobPlanHandler(currentGraphs)))
.GET("/jobs/:jobid/config", handler(new JobConfigHandler(currentGraphs)))
.GET("/jobs/:jobid/exceptions", handler(new JobExceptionsHandler(currentGraphs)))
.GET("/jobs/:jobid/accumulators", handler(new JobAccumulatorsHandler(currentGraphs)))
.GET(
"/taskmanagers",
handler(new TaskManagersHandler(retriever, DEFAULT_REQUEST_TIMEOUT)))
.GET(
"/taskmanagers/:" + TaskManagersHandler.TASK_MANAGER_ID_KEY,
handler(new TaskManagersHandler(retriever, DEFAULT_REQUEST_TIMEOUT)))
// this handler serves all the static contents
.GET("/:*", new StaticFileServerHandler(webRootDir));
}
@Test
public void testBuildDeploymentDescriptor() {
try {
final JobID jobId = new JobID();
final JobVertexID jid1 = new JobVertexID();
final JobVertexID jid2 = new JobVertexID();
final JobVertexID jid3 = new JobVertexID();
final JobVertexID jid4 = new JobVertexID();
JobVertex v1 = new JobVertex("v1", jid1);
JobVertex v2 = new JobVertex("v2", jid2);
JobVertex v3 = new JobVertex("v3", jid3);
JobVertex v4 = new JobVertex("v4", jid4);
v1.setParallelism(10);
v2.setParallelism(10);
v3.setParallelism(10);
v4.setParallelism(10);
v1.setInvokableClass(BatchTask.class);
v2.setInvokableClass(BatchTask.class);
v3.setInvokableClass(BatchTask.class);
v4.setInvokableClass(BatchTask.class);
v2.connectNewDataSetAsInput(v1, DistributionPattern.ALL_TO_ALL);
v3.connectNewDataSetAsInput(v2, DistributionPattern.ALL_TO_ALL);
v4.connectNewDataSetAsInput(v2, DistributionPattern.ALL_TO_ALL);
ExecutionGraph eg =
new ExecutionGraph(
TestingUtils.defaultExecutionContext(),
jobId,
"some job",
new Configuration(),
new SerializedValue<>(new ExecutionConfig()),
AkkaUtils.getDefaultTimeout(),
new NoRestartStrategy());
List<JobVertex> ordered = Arrays.asList(v1, v2, v3, v4);
eg.attachJobGraph(ordered);
ExecutionJobVertex ejv = eg.getAllVertices().get(jid2);
ExecutionVertex vertex = ejv.getTaskVertices()[3];
ExecutionGraphTestUtils.SimpleActorGateway instanceGateway =
new ExecutionGraphTestUtils.SimpleActorGateway(TestingUtils.directExecutionContext());
final Instance instance = getInstance(instanceGateway);
final SimpleSlot slot = instance.allocateSimpleSlot(jobId);
assertEquals(ExecutionState.CREATED, vertex.getExecutionState());
vertex.deployToSlot(slot);
assertEquals(ExecutionState.DEPLOYING, vertex.getExecutionState());
TaskDeploymentDescriptor descr = instanceGateway.lastTDD;
assertNotNull(descr);
assertEquals(jobId, descr.getJobID());
assertEquals(jid2, descr.getVertexID());
assertEquals(3, descr.getIndexInSubtaskGroup());
assertEquals(10, descr.getNumberOfSubtasks());
assertEquals(BatchTask.class.getName(), descr.getInvokableClassName());
assertEquals("v2", descr.getTaskName());
List<ResultPartitionDeploymentDescriptor> producedPartitions = descr.getProducedPartitions();
List<InputGateDeploymentDescriptor> consumedPartitions = descr.getInputGates();
assertEquals(2, producedPartitions.size());
assertEquals(1, consumedPartitions.size());
assertEquals(10, producedPartitions.get(0).getNumberOfSubpartitions());
assertEquals(10, producedPartitions.get(1).getNumberOfSubpartitions());
assertEquals(10, consumedPartitions.get(0).getInputChannelDeploymentDescriptors().length);
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
@Test
/**
* Tests that a blocking batch job fails if there are not enough resources left to schedule the
* succeeding tasks. This test case is related to [FLINK-4296] where finished producing tasks
* swallow the fail exception when scheduling a consumer task.
*/
public void testNoResourceAvailableFailure() throws Exception {
final JobID jobId = new JobID();
JobVertex v1 = new JobVertex("source");
JobVertex v2 = new JobVertex("sink");
int dop1 = 1;
int dop2 = 1;
v1.setParallelism(dop1);
v2.setParallelism(dop2);
v1.setInvokableClass(BatchTask.class);
v2.setInvokableClass(BatchTask.class);
v2.connectNewDataSetAsInput(
v1, DistributionPattern.POINTWISE, ResultPartitionType.BLOCKING, false);
// execution graph that executes actions synchronously
ExecutionGraph eg =
new ExecutionGraph(
TestingUtils.directExecutionContext(),
jobId,
"failing test job",
new Configuration(),
new SerializedValue<>(new ExecutionConfig()),
AkkaUtils.getDefaultTimeout(),
new NoRestartStrategy());
eg.setQueuedSchedulingAllowed(false);
List<JobVertex> ordered = Arrays.asList(v1, v2);
eg.attachJobGraph(ordered);
Scheduler scheduler = new Scheduler(TestingUtils.directExecutionContext());
for (int i = 0; i < dop1; i++) {
scheduler.newInstanceAvailable(
ExecutionGraphTestUtils.getInstance(
new ExecutionGraphTestUtils.SimpleActorGateway(
TestingUtils.directExecutionContext())));
}
assertEquals(dop1, scheduler.getNumberOfAvailableSlots());
// schedule, this triggers mock deployment
eg.scheduleForExecution(scheduler);
ExecutionAttemptID attemptID =
eg.getJobVertex(v1.getID())
.getTaskVertices()[0]
.getCurrentExecutionAttempt()
.getAttemptId();
eg.updateState(new TaskExecutionState(jobId, attemptID, ExecutionState.RUNNING));
eg.updateState(
new TaskExecutionState(
jobId,
attemptID,
ExecutionState.FINISHED,
null,
new AccumulatorSnapshot(
jobId,
attemptID,
new HashMap<AccumulatorRegistry.Metric, Accumulator<?, ?>>(),
new HashMap<String, Accumulator<?, ?>>())));
assertEquals(JobStatus.FAILED, eg.getState());
}
/**
* Connect the FlinkYarnCluster to the ApplicationMaster.
*
* <p>Detached YARN sessions don't need to connect to the ApplicationMaster. Detached per job YARN
* sessions need to connect until the required number of TaskManagers have been started.
*
* @throws IOException
*/
public void connectToCluster() throws IOException {
if (isConnected) {
throw new IllegalStateException("Can not connect to the cluster again");
}
// start actor system
LOG.info("Start actor system.");
InetAddress ownHostname =
NetUtils.resolveAddress(
jobManagerAddress); // find name of own public interface, able to connect to the JM
actorSystem =
AkkaUtils.createActorSystem(
flinkConfig,
new Some(new Tuple2<String, Integer>(ownHostname.getCanonicalHostName(), 0)));
// start application client
LOG.info("Start application client.");
applicationClient =
actorSystem.actorOf(
Props.create(ApplicationClient.class, flinkConfig), "applicationClient");
// instruct ApplicationClient to start a periodical status polling
applicationClient.tell(
new Messages.LocalRegisterClient(this.jobManagerAddress), applicationClient);
actorRunner =
new Thread(
new Runnable() {
@Override
public void run() {
// blocks until ApplicationMaster has been stopped
actorSystem.awaitTermination();
// get final application report
try {
ApplicationReport appReport = yarnClient.getApplicationReport(appId);
LOG.info(
"Application "
+ appId
+ " finished with state "
+ appReport.getYarnApplicationState()
+ " and final state "
+ appReport.getFinalApplicationStatus()
+ " at "
+ appReport.getFinishTime());
if (appReport.getYarnApplicationState() == YarnApplicationState.FAILED
|| appReport.getYarnApplicationState() == YarnApplicationState.KILLED) {
LOG.warn("Application failed. Diagnostics " + appReport.getDiagnostics());
LOG.warn(
"If log aggregation is activated in the Hadoop cluster, we recommend to retrieve "
+ "the full application log using this command:\n"
+ "\tyarn logs -applicationId "
+ appReport.getApplicationId()
+ "\n"
+ "(It sometimes takes a few seconds until the logs are aggregated)");
}
} catch (Exception e) {
LOG.warn("Error while getting final application report", e);
}
}
});
actorRunner.setDaemon(true);
actorRunner.start();
pollingRunner = new PollingThread(yarnClient, appId);
pollingRunner.setDaemon(true);
pollingRunner.start();
Runtime.getRuntime().addShutdownHook(clientShutdownHook);
isConnected = true;
}
@Test
public void testJobManagerProcessFailure() throws Exception {
// Config
final int numberOfJobManagers = 2;
final int numberOfTaskManagers = 2;
final int numberOfSlotsPerTaskManager = 2;
assertEquals(PARALLELISM, numberOfTaskManagers * numberOfSlotsPerTaskManager);
// Setup
// Test actor system
ActorSystem testActorSystem;
// Job managers
final JobManagerProcess[] jmProcess = new JobManagerProcess[numberOfJobManagers];
// Task managers
final ActorSystem[] tmActorSystem = new ActorSystem[numberOfTaskManagers];
// Leader election service
LeaderRetrievalService leaderRetrievalService = null;
// Coordination between the processes goes through a directory
File coordinateTempDir = null;
try {
final Deadline deadline = TestTimeOut.fromNow();
// Coordination directory
coordinateTempDir = createTempDirectory();
// Job Managers
Configuration config =
ZooKeeperTestUtils.createZooKeeperRecoveryModeConfig(
ZooKeeper.getConnectString(), FileStateBackendBasePath.getPath());
// Start first process
jmProcess[0] = new JobManagerProcess(0, config);
jmProcess[0].createAndStart();
// Task manager configuration
config.setInteger(ConfigConstants.TASK_MANAGER_MEMORY_SIZE_KEY, 4);
config.setInteger(ConfigConstants.TASK_MANAGER_NETWORK_NUM_BUFFERS_KEY, 100);
config.setInteger(ConfigConstants.TASK_MANAGER_NUM_TASK_SLOTS, 2);
// Start the task manager process
for (int i = 0; i < numberOfTaskManagers; i++) {
tmActorSystem[i] = AkkaUtils.createActorSystem(AkkaUtils.getDefaultAkkaConfig());
TaskManager.startTaskManagerComponentsAndActor(
config,
tmActorSystem[i],
"localhost",
Option.<String>empty(),
Option.<LeaderRetrievalService>empty(),
false,
TaskManager.class);
}
// Test actor system
testActorSystem = AkkaUtils.createActorSystem(AkkaUtils.getDefaultAkkaConfig());
jmProcess[0].getActorRef(testActorSystem, deadline.timeLeft());
// Leader listener
TestingListener leaderListener = new TestingListener();
leaderRetrievalService = ZooKeeperUtils.createLeaderRetrievalService(config);
leaderRetrievalService.start(leaderListener);
// Initial submission
leaderListener.waitForNewLeader(deadline.timeLeft().toMillis());
String leaderAddress = leaderListener.getAddress();
UUID leaderId = leaderListener.getLeaderSessionID();
// Get the leader ref
ActorRef leaderRef =
AkkaUtils.getActorRef(leaderAddress, testActorSystem, deadline.timeLeft());
ActorGateway leaderGateway = new AkkaActorGateway(leaderRef, leaderId);
// Wait for all task managers to connect to the leading job manager
JobManagerActorTestUtils.waitForTaskManagers(
numberOfTaskManagers, leaderGateway, deadline.timeLeft());
final File coordinateDirClosure = coordinateTempDir;
final Throwable[] errorRef = new Throwable[1];
// we trigger program execution in a separate thread
Thread programTrigger =
new Thread("Program Trigger") {
@Override
public void run() {
try {
testJobManagerFailure(ZooKeeper.getConnectString(), coordinateDirClosure);
} catch (Throwable t) {
t.printStackTrace();
errorRef[0] = t;
}
}
};
// start the test program
programTrigger.start();
// wait until all marker files are in place, indicating that all tasks have started
AbstractTaskManagerProcessFailureRecoveryTest.waitForMarkerFiles(
coordinateTempDir, READY_MARKER_FILE_PREFIX, PARALLELISM, deadline.timeLeft().toMillis());
// Kill one of the job managers and trigger recovery
jmProcess[0].destroy();
jmProcess[1] = new JobManagerProcess(1, config);
jmProcess[1].createAndStart();
jmProcess[1].getActorRef(testActorSystem, deadline.timeLeft());
// we create the marker file which signals the program functions tasks that they can complete
AbstractTaskManagerProcessFailureRecoveryTest.touchFile(
new File(coordinateTempDir, PROCEED_MARKER_FILE));
programTrigger.join(deadline.timeLeft().toMillis());
// We wait for the finish marker file. We don't wait for the program trigger, because
// we submit in detached mode.
AbstractTaskManagerProcessFailureRecoveryTest.waitForMarkerFiles(
coordinateTempDir, FINISH_MARKER_FILE_PREFIX, 1, deadline.timeLeft().toMillis());
// check that the program really finished
assertFalse("The program did not finish in time", programTrigger.isAlive());
// check whether the program encountered an error
if (errorRef[0] != null) {
Throwable error = errorRef[0];
error.printStackTrace();
fail(
"The program encountered a "
+ error.getClass().getSimpleName()
+ " : "
+ error.getMessage());
}
} catch (Throwable t) {
// Print early (in some situations the process logs get too big
// for Travis and the root problem is not shown)
t.printStackTrace();
for (JobManagerProcess p : jmProcess) {
if (p != null) {
p.printProcessLog();
}
}
throw t;
} finally {
for (int i = 0; i < numberOfTaskManagers; i++) {
if (tmActorSystem[i] != null) {
tmActorSystem[i].shutdown();
}
}
if (leaderRetrievalService != null) {
leaderRetrievalService.stop();
}
for (JobManagerProcess jmProces : jmProcess) {
if (jmProces != null) {
jmProces.destroy();
}
}
// Delete coordination directory
if (coordinateTempDir != null) {
try {
FileUtils.deleteDirectory(coordinateTempDir);
} catch (Throwable ignored) {
}
}
}
}
@BeforeClass
public static void setup() {
testActorSystem = AkkaUtils.createLocalActorSystem(new Configuration());
}
/*
* Test setup:
* - v1 is isolated, no slot sharing
* - v2 and v3 (not connected) share slots
* - v4 and v5 (connected) share slots
*/
@Test
public void testAssignSlotSharingGroup() {
try {
JobVertex v1 = new JobVertex("v1");
JobVertex v2 = new JobVertex("v2");
JobVertex v3 = new JobVertex("v3");
JobVertex v4 = new JobVertex("v4");
JobVertex v5 = new JobVertex("v5");
v1.setParallelism(4);
v2.setParallelism(5);
v3.setParallelism(7);
v4.setParallelism(1);
v5.setParallelism(11);
v2.connectNewDataSetAsInput(v1, DistributionPattern.POINTWISE);
v5.connectNewDataSetAsInput(v4, DistributionPattern.POINTWISE);
SlotSharingGroup jg1 = new SlotSharingGroup();
v2.setSlotSharingGroup(jg1);
v3.setSlotSharingGroup(jg1);
SlotSharingGroup jg2 = new SlotSharingGroup();
v4.setSlotSharingGroup(jg2);
v5.setSlotSharingGroup(jg2);
List<JobVertex> vertices = new ArrayList<JobVertex>(Arrays.asList(v1, v2, v3, v4, v5));
ExecutionGraph eg =
new ExecutionGraph(
TestingUtils.defaultExecutionContext(),
new JobID(),
"test job",
new Configuration(),
ExecutionConfigTest.getSerializedConfig(),
AkkaUtils.getDefaultTimeout(),
new NoRestartStrategy());
eg.attachJobGraph(vertices);
// verify that the vertices are all in the same slot sharing group
SlotSharingGroup group1 = null;
SlotSharingGroup group2 = null;
// verify that v1 tasks have no slot sharing group
assertNull(eg.getJobVertex(v1.getID()).getSlotSharingGroup());
// v2 and v3 are shared
group1 = eg.getJobVertex(v2.getID()).getSlotSharingGroup();
assertNotNull(group1);
assertEquals(group1, eg.getJobVertex(v3.getID()).getSlotSharingGroup());
assertEquals(2, group1.getJobVertexIds().size());
assertTrue(group1.getJobVertexIds().contains(v2.getID()));
assertTrue(group1.getJobVertexIds().contains(v3.getID()));
// v4 and v5 are shared
group2 = eg.getJobVertex(v4.getID()).getSlotSharingGroup();
assertNotNull(group2);
assertEquals(group2, eg.getJobVertex(v5.getID()).getSlotSharingGroup());
assertEquals(2, group1.getJobVertexIds().size());
assertTrue(group2.getJobVertexIds().contains(v4.getID()));
assertTrue(group2.getJobVertexIds().contains(v5.getID()));
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
