use of org.apache.tez.dag.api.DAG in project hive by apache.
the class GenericUDTFGetSplits method getSplits.
public InputSplit[] getSplits(JobConf job, int numSplits, TezWork work, Schema schema) throws IOException {
DAG dag = DAG.create(work.getName());
dag.setCredentials(job.getCredentials());
DagUtils utils = DagUtils.getInstance();
Context ctx = new Context(job);
MapWork mapWork = (MapWork) work.getAllWork().get(0);
// bunch of things get setup in the context based on conf but we need only the MR tmp directory
// for the following method.
JobConf wxConf = utils.initializeVertexConf(job, ctx, mapWork);
// TODO: should we also whitelist input formats here? from mapred.input.format.class
Path scratchDir = utils.createTezDir(ctx.getMRScratchDir(), job);
FileSystem fs = scratchDir.getFileSystem(job);
try {
LocalResource appJarLr = createJarLocalResource(utils.getExecJarPathLocal(), utils, job);
Vertex wx = utils.createVertex(wxConf, mapWork, scratchDir, appJarLr, new ArrayList<LocalResource>(), fs, ctx, false, work, work.getVertexType(mapWork));
String vertexName = wx.getName();
dag.addVertex(wx);
utils.addCredentials(mapWork, dag);
// we have the dag now proceed to get the splits:
Preconditions.checkState(HiveConf.getBoolVar(wxConf, ConfVars.HIVE_TEZ_GENERATE_CONSISTENT_SPLITS));
Preconditions.checkState(HiveConf.getBoolVar(wxConf, ConfVars.LLAP_CLIENT_CONSISTENT_SPLITS));
HiveSplitGenerator splitGenerator = new HiveSplitGenerator(wxConf, mapWork);
List<Event> eventList = splitGenerator.initialize();
InputSplit[] result = new InputSplit[eventList.size() - 1];
InputConfigureVertexTasksEvent configureEvent = (InputConfigureVertexTasksEvent) eventList.get(0);
List<TaskLocationHint> hints = configureEvent.getLocationHint().getTaskLocationHints();
Preconditions.checkState(hints.size() == eventList.size() - 1);
if (LOG.isDebugEnabled()) {
LOG.debug("NumEvents=" + eventList.size() + ", NumSplits=" + result.length);
}
LlapCoordinator coordinator = LlapCoordinator.getInstance();
if (coordinator == null) {
throw new IOException("LLAP coordinator is not initialized; must be running in HS2 with " + ConfVars.LLAP_HS2_ENABLE_COORDINATOR.varname + " enabled");
}
// See the discussion in the implementation as to why we generate app ID.
ApplicationId applicationId = coordinator.createExtClientAppId();
// This assumes LLAP cluster owner is always the HS2 user.
String llapUser = UserGroupInformation.getLoginUser().getShortUserName();
String queryUser = null;
byte[] tokenBytes = null;
LlapSigner signer = null;
if (UserGroupInformation.isSecurityEnabled()) {
signer = coordinator.getLlapSigner(job);
// 1. Generate the token for query user (applies to all splits).
queryUser = SessionState.getUserFromAuthenticator();
if (queryUser == null) {
queryUser = UserGroupInformation.getCurrentUser().getUserName();
LOG.warn("Cannot determine the session user; using " + queryUser + " instead");
}
LlapTokenLocalClient tokenClient = coordinator.getLocalTokenClient(job, llapUser);
// We put the query user, not LLAP user, into the message and token.
Token<LlapTokenIdentifier> token = tokenClient.createToken(applicationId.toString(), queryUser, true);
LOG.info("Created the token for remote user: {}", token);
bos.reset();
token.write(dos);
tokenBytes = bos.toByteArray();
} else {
queryUser = UserGroupInformation.getCurrentUser().getUserName();
}
LOG.info("Number of splits: " + (eventList.size() - 1));
SignedMessage signedSvs = null;
for (int i = 0; i < eventList.size() - 1; i++) {
TaskSpec taskSpec = new TaskSpecBuilder().constructTaskSpec(dag, vertexName, eventList.size() - 1, applicationId, i);
// 2. Generate the vertex/submit information for all events.
if (i == 0) {
// The queryId could either be picked up from the current request being processed, or
// generated. The current request isn't exactly correct since the query is 'done' once we
// return the results. Generating a new one has the added benefit of working once this
// is moved out of a UDTF into a proper API.
// Setting this to the generated AppId which is unique.
// Despite the differences in TaskSpec, the vertex spec should be the same.
signedSvs = createSignedVertexSpec(signer, taskSpec, applicationId, queryUser, applicationId.toString());
}
SubmitWorkInfo submitWorkInfo = new SubmitWorkInfo(applicationId, System.currentTimeMillis(), taskSpec.getVertexParallelism(), signedSvs.message, signedSvs.signature);
byte[] submitWorkBytes = SubmitWorkInfo.toBytes(submitWorkInfo);
// 3. Generate input event.
SignedMessage eventBytes = makeEventBytes(wx, vertexName, eventList.get(i + 1), signer);
// 4. Make location hints.
SplitLocationInfo[] locations = makeLocationHints(hints.get(i));
result[i] = new LlapInputSplit(i, submitWorkBytes, eventBytes.message, eventBytes.signature, locations, schema, llapUser, tokenBytes);
}
return result;
} catch (Exception e) {
throw new IOException(e);
}
}
use of org.apache.tez.dag.api.DAG in project hive by apache.
the class TestTezTask method testEmptyWork.
@Test
public void testEmptyWork() throws IllegalArgumentException, IOException, Exception {
DAG dag = task.build(conf, new TezWork("", null), path, appLr, null, new Context(conf));
assertEquals(dag.getVertices().size(), 0);
}
use of org.apache.tez.dag.api.DAG in project hive by apache.
the class TestTezTask method testSubmit.
@Test
public void testSubmit() throws Exception {
DAG dag = DAG.create("test");
task.submit(conf, dag, path, appLr, sessionState, Collections.<LocalResource>emptyList(), new String[0], Collections.<String, LocalResource>emptyMap());
// validate close/reopen
verify(sessionState, times(1)).open(any(HiveConf.class), any(String[].class));
// now uses pool after HIVE-7043
verify(sessionState, times(1)).close(eq(true));
verify(session, times(2)).submitDAG(any(DAG.class));
}
use of org.apache.tez.dag.api.DAG in project hive by apache.
the class TezTask method execute.
@Override
public int execute(DriverContext driverContext) {
int rc = 1;
boolean cleanContext = false;
Context ctx = null;
TezSessionState session = null;
try {
// Get or create Context object. If we create it we have to clean it later as well.
ctx = driverContext.getCtx();
if (ctx == null) {
ctx = new Context(conf);
cleanContext = true;
}
// Need to remove this static hack. But this is the way currently to get a session.
SessionState ss = SessionState.get();
session = ss.getTezSession();
if (session != null && !session.isOpen()) {
LOG.warn("The session: " + session + " has not been opened");
}
session = TezSessionPoolManager.getInstance().getSession(session, conf, false, getWork().getLlapMode());
ss.setTezSession(session);
try {
// jobConf will hold all the configuration for hadoop, tez, and hive
JobConf jobConf = utils.createConfiguration(conf);
// Get all user jars from work (e.g. input format stuff).
String[] inputOutputJars = work.configureJobConfAndExtractJars(jobConf);
// we will localize all the files (jars, plans, hashtables) to the
// scratch dir. let's create this and tmp first.
Path scratchDir = ctx.getMRScratchDir();
// create the tez tmp dir
scratchDir = utils.createTezDir(scratchDir, conf);
Map<String, LocalResource> inputOutputLocalResources = getExtraLocalResources(jobConf, scratchDir, inputOutputJars);
// Ensure the session is open and has the necessary local resources
updateSession(session, jobConf, scratchDir, inputOutputJars, inputOutputLocalResources);
List<LocalResource> additionalLr = session.getLocalizedResources();
logResources(additionalLr);
// unless already installed on all the cluster nodes, we'll have to
// localize hive-exec.jar as well.
LocalResource appJarLr = session.getAppJarLr();
// next we translate the TezWork to a Tez DAG
DAG dag = build(jobConf, work, scratchDir, appJarLr, additionalLr, ctx);
CallerContext callerContext = CallerContext.create("HIVE", queryPlan.getQueryId(), "HIVE_QUERY_ID", queryPlan.getQueryStr());
dag.setCallerContext(callerContext);
// Add the extra resources to the dag
addExtraResourcesToDag(session, dag, inputOutputJars, inputOutputLocalResources);
// submit will send the job to the cluster and start executing
dagClient = submit(jobConf, dag, scratchDir, appJarLr, session, additionalLr, inputOutputJars, inputOutputLocalResources);
// finally monitor will print progress until the job is done
TezJobMonitor monitor = new TezJobMonitor(work.getWorkMap(), dagClient, conf, dag, ctx);
rc = monitor.monitorExecution();
if (rc != 0) {
this.setException(new HiveException(monitor.getDiagnostics()));
}
// fetch the counters
try {
Set<StatusGetOpts> statusGetOpts = EnumSet.of(StatusGetOpts.GET_COUNTERS);
counters = dagClient.getDAGStatus(statusGetOpts).getDAGCounters();
} catch (Exception err) {
// Don't fail execution due to counters - just don't print summary info
LOG.warn("Failed to get counters. Ignoring, summary info will be incomplete. " + err, err);
counters = null;
}
} finally {
// We return this to the pool even if it's unusable; reopen is supposed to handle this.
try {
TezSessionPoolManager.getInstance().returnSession(session, getWork().getLlapMode());
} catch (Exception e) {
LOG.error("Failed to return session: {} to pool", session, e);
throw e;
}
}
if (LOG.isInfoEnabled() && counters != null && (HiveConf.getBoolVar(conf, HiveConf.ConfVars.TEZ_EXEC_SUMMARY) || Utilities.isPerfOrAboveLogging(conf))) {
for (CounterGroup group : counters) {
LOG.info(group.getDisplayName() + ":");
for (TezCounter counter : group) {
LOG.info(" " + counter.getDisplayName() + ": " + counter.getValue());
}
}
}
} catch (Exception e) {
LOG.error("Failed to execute tez graph.", e);
// rc will be 1 at this point indicating failure.
} finally {
Utilities.clearWork(conf);
// Clear gWorkMap
for (BaseWork w : work.getAllWork()) {
JobConf workCfg = workToConf.get(w);
if (workCfg != null) {
Utilities.clearWorkMapForConf(workCfg);
}
}
if (cleanContext) {
try {
ctx.clear();
} catch (Exception e) {
/*best effort*/
LOG.warn("Failed to clean up after tez job", e);
}
}
// need to either move tmp files or remove them
if (dagClient != null) {
// rc will only be overwritten if close errors out
rc = close(work, rc);
}
}
return rc;
}
use of org.apache.tez.dag.api.DAG in project hive by apache.
the class TezTask method build.
DAG build(JobConf conf, TezWork work, Path scratchDir, LocalResource appJarLr, List<LocalResource> additionalLr, Context ctx) throws Exception {
perfLogger.PerfLogBegin(CLASS_NAME, PerfLogger.TEZ_BUILD_DAG);
// getAllWork returns a topologically sorted list, which we use to make
// sure that vertices are created before they are used in edges.
List<BaseWork> ws = work.getAllWork();
Collections.reverse(ws);
FileSystem fs = scratchDir.getFileSystem(conf);
// the name of the dag is what is displayed in the AM/Job UI
String dagName = utils.createDagName(conf, queryPlan);
LOG.info("Dag name: " + dagName);
DAG dag = DAG.create(dagName);
// set some info for the query
JSONObject json = new JSONObject(new LinkedHashMap()).put("context", "Hive").put("description", ctx.getCmd());
String dagInfo = json.toString();
if (LOG.isDebugEnabled()) {
LOG.debug("DagInfo: " + dagInfo);
}
dag.setDAGInfo(dagInfo);
dag.setCredentials(conf.getCredentials());
setAccessControlsForCurrentUser(dag, queryPlan.getQueryId(), conf);
for (BaseWork w : ws) {
boolean isFinal = work.getLeaves().contains(w);
// translate work to vertex
perfLogger.PerfLogBegin(CLASS_NAME, PerfLogger.TEZ_CREATE_VERTEX + w.getName());
if (w instanceof UnionWork) {
// Special case for unions. These items translate to VertexGroups
List<BaseWork> unionWorkItems = new LinkedList<BaseWork>();
List<BaseWork> children = new LinkedList<BaseWork>();
// proper children of the union
for (BaseWork v : work.getChildren(w)) {
EdgeType type = work.getEdgeProperty(w, v).getEdgeType();
if (type == EdgeType.CONTAINS) {
unionWorkItems.add(v);
} else {
children.add(v);
}
}
// create VertexGroup
Vertex[] vertexArray = new Vertex[unionWorkItems.size()];
int i = 0;
for (BaseWork v : unionWorkItems) {
vertexArray[i++] = workToVertex.get(v);
}
VertexGroup group = dag.createVertexGroup(w.getName(), vertexArray);
// For a vertex group, all Outputs use the same Key-class, Val-class and partitioner.
// Pick any one source vertex to figure out the Edge configuration.
JobConf parentConf = workToConf.get(unionWorkItems.get(0));
// now hook up the children
for (BaseWork v : children) {
// finally we can create the grouped edge
GroupInputEdge e = utils.createEdge(group, parentConf, workToVertex.get(v), work.getEdgeProperty(w, v), work.getVertexType(v));
dag.addEdge(e);
}
} else {
// Regular vertices
JobConf wxConf = utils.initializeVertexConf(conf, ctx, w);
Vertex wx = utils.createVertex(wxConf, w, scratchDir, appJarLr, additionalLr, fs, ctx, !isFinal, work, work.getVertexType(w));
if (w.getReservedMemoryMB() > 0) {
// If reversedMemoryMB is set, make memory allocation fraction adjustment as needed
double frac = DagUtils.adjustMemoryReserveFraction(w.getReservedMemoryMB(), super.conf);
LOG.info("Setting " + TEZ_MEMORY_RESERVE_FRACTION + " to " + frac);
wx.setConf(TEZ_MEMORY_RESERVE_FRACTION, Double.toString(frac));
}
// Otherwise just leave it up to Tez to decide how much memory to allocate
dag.addVertex(wx);
utils.addCredentials(w, dag);
perfLogger.PerfLogEnd(CLASS_NAME, PerfLogger.TEZ_CREATE_VERTEX + w.getName());
workToVertex.put(w, wx);
workToConf.put(w, wxConf);
// add all dependencies (i.e.: edges) to the graph
for (BaseWork v : work.getChildren(w)) {
assert workToVertex.containsKey(v);
Edge e = null;
TezEdgeProperty edgeProp = work.getEdgeProperty(w, v);
e = utils.createEdge(wxConf, wx, workToVertex.get(v), edgeProp, work.getVertexType(v));
dag.addEdge(e);
}
}
}
// Clear the work map after build. TODO: remove caching instead?
Utilities.clearWorkMap(conf);
perfLogger.PerfLogEnd(CLASS_NAME, PerfLogger.TEZ_BUILD_DAG);
return dag;
}
Aggregations