/**
* Verify if the NodeManager could identify disk failures.
*
* @param localORLogDirs <em>true</em> represent nm-local-dirs and <em>false </em> means
* nm-log-dirs
* @param expectedDirs expected nm-local-dirs/nm-log-dirs as a string
* @param isHealthy <em>true</em> if the overall node should be healthy
*/
private void verifyDisksHealth(boolean localORLogDirs, String expectedDirs, boolean isHealthy) {
// Wait for the NodeManager to identify disk failures.
waitForDiskHealthCheck();
List<String> list = localORLogDirs ? dirsHandler.getLocalDirs() : dirsHandler.getLogDirs();
String seenDirs = StringUtils.join(",", list);
LOG.info("ExpectedDirs=" + expectedDirs);
LOG.info("SeenDirs=" + seenDirs);
Assert.assertTrue(
"NodeManager could not identify disk failure.", expectedDirs.equals(seenDirs));
Assert.assertEquals(
"Node's health in terms of disks is wrong", isHealthy, dirsHandler.areDisksHealthy());
for (int i = 0; i < 10; i++) {
Iterator<RMNode> iter =
yarnCluster.getResourceManager().getRMContext().getRMNodes().values().iterator();
if (iter.next().getNodeHealthStatus().getIsNodeHealthy() == isHealthy) {
break;
}
// wait for the node health info to go to RM
try {
Thread.sleep(1000);
} catch (InterruptedException e) {
LOG.error("Interrupted while waiting for NM->RM heartbeat.");
}
}
Iterator<RMNode> iter =
yarnCluster.getResourceManager().getRMContext().getRMNodes().values().iterator();
Assert.assertEquals(
"RM is not updated with the health status of a node",
isHealthy,
iter.next().getNodeHealthStatus().getIsNodeHealthy());
}
private void testDirsFailures(boolean localORLogDirs) throws IOException {
String dirType = localORLogDirs ? "local" : "log";
String dirsProperty =
localORLogDirs ? YarnConfiguration.NM_LOCAL_DIRS : YarnConfiguration.NM_LOG_DIRS;
Configuration conf = new Configuration();
// set disk health check interval to a small value (say 1 sec).
conf.setLong(YarnConfiguration.NM_DISK_HEALTH_CHECK_INTERVAL_MS, DISK_HEALTH_CHECK_INTERVAL);
// If 2 out of the total 4 local-dirs fail OR if 2 Out of the total 4
// log-dirs fail, then the node's health status should become unhealthy.
conf.setFloat(YarnConfiguration.NM_MIN_HEALTHY_DISKS_FRACTION, 0.60F);
if (yarnCluster != null) {
yarnCluster.stop();
FileUtil.fullyDelete(localFSDirBase);
localFSDirBase.mkdirs();
}
LOG.info("Starting up YARN cluster");
yarnCluster =
new MiniYARNCluster(TestDiskFailures.class.getName(), 1, numLocalDirs, numLogDirs);
yarnCluster.init(conf);
yarnCluster.start();
NodeManager nm = yarnCluster.getNodeManager(0);
LOG.info("Configured nm-" + dirType + "-dirs=" + nm.getConfig().get(dirsProperty));
dirsHandler = nm.getNodeHealthChecker().getDiskHandler();
List<String> list = localORLogDirs ? dirsHandler.getLocalDirs() : dirsHandler.getLogDirs();
String[] dirs = list.toArray(new String[list.size()]);
Assert.assertEquals("Number of nm-" + dirType + "-dirs is wrong.", numLocalDirs, dirs.length);
String expectedDirs = StringUtils.join(",", list);
// validate the health of disks initially
verifyDisksHealth(localORLogDirs, expectedDirs, true);
// Make 1 nm-local-dir fail and verify if "the nodemanager can identify
// the disk failure(s) and can update the list of good nm-local-dirs.
prepareDirToFail(dirs[2]);
expectedDirs = dirs[0] + "," + dirs[1] + "," + dirs[3];
verifyDisksHealth(localORLogDirs, expectedDirs, true);
// Now, make 1 more nm-local-dir/nm-log-dir fail and verify if "the
// nodemanager can identify the disk failures and can update the list of
// good nm-local-dirs/nm-log-dirs and can update the overall health status
// of the node to unhealthy".
prepareDirToFail(dirs[0]);
expectedDirs = dirs[1] + "," + dirs[3];
verifyDisksHealth(localORLogDirs, expectedDirs, false);
// Fail the remaining 2 local-dirs/log-dirs and verify if NM remains with
// empty list of local-dirs/log-dirs and the overall health status is
// unhealthy.
prepareDirToFail(dirs[1]);
prepareDirToFail(dirs[3]);
expectedDirs = "";
verifyDisksHealth(localORLogDirs, expectedDirs, false);
}
private void uploadLogsForContainer(ContainerId containerId) {
if (this.logAggregationDisabled) {
return;
}
// Lazy creation of the writer
if (this.writer == null) {
LOG.info(
"Starting aggregate log-file for app "
+ this.applicationId
+ " at "
+ this.remoteNodeTmpLogFileForApp);
try {
this.writer = new LogWriter(this.conf, this.remoteNodeTmpLogFileForApp, this.userUgi);
// Write ACLs once when and if the writer is created.
this.writer.writeApplicationACLs(appAcls);
this.writer.writeApplicationOwner(this.userUgi.getShortUserName());
} catch (IOException e) {
LOG.error(
"Cannot create writer for app "
+ this.applicationId
+ ". Disabling log-aggregation for this app.",
e);
this.logAggregationDisabled = true;
return;
}
}
LOG.info(
"Uploading logs for container "
+ containerId
+ ". Current good log dirs are "
+ StringUtils.join(",", dirsHandler.getLogDirs()));
LogKey logKey = new LogKey(containerId);
LogValue logValue =
new LogValue(dirsHandler.getLogDirs(), containerId, userUgi.getShortUserName());
try {
this.writer.append(logKey, logValue);
} catch (IOException e) {
LOG.error("Couldn't upload logs for " + containerId + ". Skipping this container.");
}
}
/**
* Make a local and log directory inaccessible during initialization and verify those bad
* directories are recognized and removed from the list of available local and log directories.
*
* @throws IOException
*/
@Test
public void testDirFailuresOnStartup() throws IOException {
Configuration conf = new YarnConfiguration();
String localDir1 = new File(testDir, "localDir1").getPath();
String localDir2 = new File(testDir, "localDir2").getPath();
String logDir1 = new File(testDir, "logDir1").getPath();
String logDir2 = new File(testDir, "logDir2").getPath();
conf.set(YarnConfiguration.NM_LOCAL_DIRS, localDir1 + "," + localDir2);
conf.set(YarnConfiguration.NM_LOG_DIRS, logDir1 + "," + logDir2);
prepareDirToFail(localDir1);
prepareDirToFail(logDir2);
LocalDirsHandlerService dirSvc = new LocalDirsHandlerService();
dirSvc.init(conf);
List<String> localDirs = dirSvc.getLocalDirs();
Assert.assertEquals(1, localDirs.size());
Assert.assertEquals(localDir2, localDirs.get(0));
List<String> logDirs = dirSvc.getLogDirs();
Assert.assertEquals(1, logDirs.size());
Assert.assertEquals(logDir1, logDirs.get(0));
}
@SuppressWarnings("unchecked")
private void doAppLogAggregation() {
ContainerId containerId;
while (!this.appFinishing.get()) {
synchronized (this) {
try {
wait(THREAD_SLEEP_TIME);
} catch (InterruptedException e) {
LOG.warn("PendingContainers queue is interrupted");
this.appFinishing.set(true);
}
}
}
// Application is finished. Finish pending-containers
while ((containerId = this.pendingContainers.poll()) != null) {
uploadLogsForContainer(containerId);
}
// Remove the local app-log-dirs
List<String> rootLogDirs = dirsHandler.getLogDirs();
Path[] localAppLogDirs = new Path[rootLogDirs.size()];
int index = 0;
for (String rootLogDir : rootLogDirs) {
localAppLogDirs[index] = new Path(rootLogDir, this.applicationId);
index++;
}
this.delService.delete(this.userUgi.getShortUserName(), null, localAppLogDirs);
if (this.writer != null) {
this.writer.closeWriter();
LOG.info("Finished aggregate log-file for app " + this.applicationId);
}
try {
userUgi.doAs(
new PrivilegedExceptionAction<Object>() {
@Override
public Object run() throws Exception {
FileSystem remoteFS = FileSystem.get(conf);
remoteFS.rename(remoteNodeTmpLogFileForApp, remoteNodeLogFileForApp);
return null;
}
});
} catch (Exception e) {
LOG.error(
"Failed to move temporary log file to final location: ["
+ remoteNodeTmpLogFileForApp
+ "] to ["
+ remoteNodeLogFileForApp
+ "]",
e);
}
this.dispatcher
.getEventHandler()
.handle(
new ApplicationEvent(
this.appId, ApplicationEventType.APPLICATION_LOG_HANDLING_FINISHED));
this.appAggregationFinished.set(true);
}