Examples with TezJobMonitor org.apache.hadoop.hive.ql.exec.tez.monitoring.TezJobMonitor used on opensource projects

Example 1 with TezJobMonitor

use of org.apache.hadoop.hive.ql.exec.tez.monitoring.TezJobMonitor 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;
    Ref<TezSessionState> sessionRef = Ref.from(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;
            // some DDL task that directly executes a TezTask does not setup Context and hence TriggerContext.
            // Setting queryId is messed up. Some DDL tasks have executionId instead of proper queryId.
            String queryId = HiveConf.getVar(conf, HiveConf.ConfVars.HIVEQUERYID);
            WmContext wmContext = new WmContext(System.currentTimeMillis(), queryId);
            ctx.setWmContext(wmContext);
        }
        // Need to remove this static hack. But this is the way currently to get a session.
        SessionState ss = SessionState.get();
        // Note: given that we return pool sessions to the pool in the finally block below, and that
        // we need to set the global to null to do that, this "reuse" may be pointless.
        TezSessionState session = sessionRef.value = ss.getTezSession();
        if (session != null && !session.isOpen()) {
            LOG.warn("The session: " + session + " has not been opened");
        }
        // We only need a username for UGI to use for groups; getGroups will fetch the groups
        // based on Hadoop configuration, as documented at
        // https://hadoop.apache.org/docs/r2.8.0/hadoop-project-dist/hadoop-common/GroupsMapping.html
        String userName = ss.getUserName();
        List<String> groups = null;
        if (userName == null) {
            userName = "anonymous";
        } else {
            groups = UserGroupInformation.createRemoteUser(ss.getUserName()).getGroups();
        }
        MappingInput mi = new MappingInput(userName, groups, ss.getHiveVariables().get("wmpool"), ss.getHiveVariables().get("wmapp"));
        WmContext wmContext = ctx.getWmContext();
        // 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[] allNonConfFiles = work.configureJobConfAndExtractJars(jobConf);
        // DAG scratch dir. We get a session from the pool so it may be different from Tez one.
        // TODO: we could perhaps reuse the same directory for HiveResources?
        Path scratchDir = utils.createTezDir(ctx.getMRScratchDir(), conf);
        CallerContext callerContext = CallerContext.create("HIVE", queryPlan.getQueryId(), "HIVE_QUERY_ID", queryPlan.getQueryStr());
        perfLogger.PerfLogBegin(CLASS_NAME, PerfLogger.TEZ_GET_SESSION);
        session = sessionRef.value = WorkloadManagerFederation.getSession(sessionRef.value, conf, mi, getWork().getLlapMode(), wmContext);
        perfLogger.PerfLogEnd(CLASS_NAME, PerfLogger.TEZ_GET_SESSION);
        try {
            ss.setTezSession(session);
            LOG.info("Subscribed to counters: {} for queryId: {}", wmContext.getSubscribedCounters(), wmContext.getQueryId());
            // Ensure the session is open and has the necessary local resources.
            // This would refresh any conf resources and also local resources.
            ensureSessionHasResources(session, allNonConfFiles);
            // This is a combination of the jar stuff from conf, and not from conf.
            List<LocalResource> allNonAppResources = session.getLocalizedResources();
            logResources(allNonAppResources);
            Map<String, LocalResource> allResources = DagUtils.createTezLrMap(session.getAppJarLr(), allNonAppResources);
            // next we translate the TezWork to a Tez DAG
            DAG dag = build(jobConf, work, scratchDir, ctx, allResources);
            dag.setCallerContext(callerContext);
            // Check isShutdown opportunistically; it's never unset.
            if (this.isShutdown) {
                throw new HiveException("Operation cancelled");
            }
            DAGClient dagClient = submit(jobConf, dag, sessionRef);
            session = sessionRef.value;
            boolean wasShutdown = false;
            synchronized (dagClientLock) {
                assert this.dagClient == null;
                wasShutdown = this.isShutdown;
                if (!wasShutdown) {
                    this.dagClient = dagClient;
                }
            }
            if (wasShutdown) {
                closeDagClientOnCancellation(dagClient);
                throw new HiveException("Operation cancelled");
            }
            // finally monitor will print progress until the job is done
            TezJobMonitor monitor = new TezJobMonitor(work.getAllWork(), 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 {
            // Note: due to TEZ-3846, the session may actually be invalid in case of some errors.
            // Currently, reopen on an attempted reuse will take care of that; we cannot tell
            // if the session is usable until we try.
            // We return this to the pool even if it's unusable; reopen is supposed to handle this.
            wmContext = ctx.getWmContext();
            try {
                if (sessionRef.value != null) {
                    sessionRef.value.returnToSessionManager();
                }
            } catch (Exception e) {
                LOG.error("Failed to return session: {} to pool", session, e);
                throw e;
            }
            if (!conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("none") && wmContext != null) {
                if (conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("json")) {
                    wmContext.printJson(console);
                } else if (conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("text")) {
                    wmContext.print(console);
                }
            }
        }
        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
        DAGClient dagClient = null;
        synchronized (dagClientLock) {
            dagClient = this.dagClient;
            this.dagClient = null;
        }
        // DagClient as such should have no bearing on jobClose.
        if (dagClient != null) {
            // rc will only be overwritten if close errors out
            rc = close(work, rc, dagClient);
        }
    }
    return rc;
}

Example 2 with TezJobMonitor

use of org.apache.hadoop.hive.ql.exec.tez.monitoring.TezJobMonitor in project hive by apache.

the class TezTask method execute.

@Override
public int execute() {
    int rc = 1;
    boolean cleanContext = false;
    Context ctx = null;
    Ref<TezSessionState> sessionRef = Ref.from(null);
    final String queryId = HiveConf.getVar(conf, HiveConf.ConfVars.HIVEQUERYID);
    try {
        // Get or create Context object. If we create it we have to clean it later as well.
        ctx = context;
        if (ctx == null) {
            ctx = new Context(conf);
            cleanContext = true;
            // some DDL task that directly executes a TezTask does not setup Context and hence TriggerContext.
            // Setting queryId is messed up. Some DDL tasks have executionId instead of proper queryId.
            WmContext wmContext = new WmContext(System.currentTimeMillis(), queryId);
            ctx.setWmContext(wmContext);
        }
        // Need to remove this static hack. But this is the way currently to get a session.
        SessionState ss = SessionState.get();
        // Note: given that we return pool sessions to the pool in the finally block below, and that
        // we need to set the global to null to do that, this "reuse" may be pointless.
        TezSessionState session = sessionRef.value = ss.getTezSession();
        if (session != null && !session.isOpen()) {
            LOG.warn("The session: " + session + " has not been opened");
        }
        // We only need a username for UGI to use for groups; getGroups will fetch the groups
        // based on Hadoop configuration, as documented at
        // https://hadoop.apache.org/docs/r2.8.0/hadoop-project-dist/hadoop-common/GroupsMapping.html
        String userName = getUserNameForGroups(ss);
        List<String> groups = null;
        if (userName == null) {
            userName = "anonymous";
        } else {
            try {
                groups = UserGroupInformation.createRemoteUser(userName).getGroups();
            } catch (Exception ex) {
                LOG.warn("Cannot obtain groups for " + userName, ex);
            }
        }
        MappingInput mi = new MappingInput(userName, groups, ss.getHiveVariables().get("wmpool"), ss.getHiveVariables().get("wmapp"));
        WmContext wmContext = ctx.getWmContext();
        // jobConf will hold all the configuration for hadoop, tez, and hive, which are not set in AM defaults
        JobConf jobConf = utils.createConfiguration(conf, false);
        // Setup the job specific keystore path if exists and put the password into the environment variables of tez am/tasks.
        HiveConfUtil.updateJobCredentialProviders(jobConf);
        // Get all user jars from work (e.g. input format stuff).
        String[] allNonConfFiles = work.configureJobConfAndExtractJars(jobConf);
        // DAG scratch dir. We get a session from the pool so it may be different from Tez one.
        // TODO: we could perhaps reuse the same directory for HiveResources?
        Path scratchDir = utils.createTezDir(ctx.getMRScratchDir(), conf);
        CallerContext callerContext = CallerContext.create("HIVE", queryPlan.getQueryId(), "HIVE_QUERY_ID", queryPlan.getQueryStr());
        perfLogger.perfLogBegin(CLASS_NAME, PerfLogger.TEZ_GET_SESSION);
        session = sessionRef.value = WorkloadManagerFederation.getSession(sessionRef.value, conf, mi, getWork().getLlapMode(), wmContext);
        perfLogger.perfLogEnd(CLASS_NAME, PerfLogger.TEZ_GET_SESSION);
        try {
            ss.setTezSession(session);
            LOG.info("Subscribed to counters: {} for queryId: {}", wmContext.getSubscribedCounters(), wmContext.getQueryId());
            // Ensure the session is open and has the necessary local resources.
            // This would refresh any conf resources and also local resources.
            ensureSessionHasResources(session, allNonConfFiles);
            // This is a combination of the jar stuff from conf, and not from conf.
            List<LocalResource> allNonAppResources = session.getLocalizedResources();
            logResources(allNonAppResources);
            Map<String, LocalResource> allResources = DagUtils.createTezLrMap(session.getAppJarLr(), allNonAppResources);
            // next we translate the TezWork to a Tez DAG
            DAG dag = build(jobConf, work, scratchDir, ctx, allResources);
            dag.setCallerContext(callerContext);
            // Check isShutdown opportunistically; it's never unset.
            if (this.isShutdown) {
                throw new HiveException("Operation cancelled");
            }
            DAGClient dagClient = submit(dag, sessionRef);
            session = sessionRef.value;
            boolean wasShutdown = false;
            synchronized (dagClientLock) {
                assert this.dagClient == null;
                wasShutdown = this.isShutdown;
                if (!wasShutdown) {
                    this.dagClient = dagClient;
                }
            }
            if (wasShutdown) {
                closeDagClientOnCancellation(dagClient);
                throw new HiveException("Operation cancelled");
            }
            // Log all the info required to find the various logs for this query
            LOG.info("HS2 Host: [{}], Query ID: [{}], Dag ID: [{}], DAG Session ID: [{}]", ServerUtils.hostname(), queryId, this.dagClient.getDagIdentifierString(), this.dagClient.getSessionIdentifierString());
            // finally monitor will print progress until the job is done
            TezJobMonitor monitor = new TezJobMonitor(work.getAllWork(), dagClient, conf, dag, ctx, counters);
            rc = monitor.monitorExecution();
            if (rc != 0) {
                this.setException(new HiveException(monitor.getDiagnostics()));
            }
            try {
                // fetch the counters
                Set<StatusGetOpts> statusGetOpts = EnumSet.of(StatusGetOpts.GET_COUNTERS);
                TezCounters dagCounters = dagClient.getDAGStatus(statusGetOpts).getDAGCounters();
                // if initial counters exists, merge it with dag counters to get aggregated view
                TezCounters mergedCounters = counters == null ? dagCounters : Utils.mergeTezCounters(dagCounters, counters);
                counters = mergedCounters;
            } 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);
                counters = null;
            }
            // save useful commit information into query state, e.g. for custom commit hooks, like Iceberg
            if (rc == 0) {
                collectCommitInformation(work);
            }
        } finally {
            // Note: due to TEZ-3846, the session may actually be invalid in case of some errors.
            // Currently, reopen on an attempted reuse will take care of that; we cannot tell
            // if the session is usable until we try.
            // We return this to the pool even if it's unusable; reopen is supposed to handle this.
            wmContext = ctx.getWmContext();
            try {
                if (sessionRef.value != null) {
                    sessionRef.value.returnToSessionManager();
                }
            } catch (Exception e) {
                LOG.error("Failed to return session: {} to pool", session, e);
                throw e;
            }
            if (!conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("none") && wmContext != null) {
                if (conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("json")) {
                    wmContext.printJson(console);
                } else if (conf.getVar(HiveConf.ConfVars.TEZ_SESSION_EVENTS_SUMMARY).equalsIgnoreCase("text")) {
                    wmContext.print(console);
                }
            }
        }
        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());
                }
            }
        }
        updateNumRows();
    } catch (Exception e) {
        LOG.error("Failed to execute tez graph.", e);
        setException(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
        DAGClient dagClient = null;
        synchronized (dagClientLock) {
            dagClient = this.dagClient;
            this.dagClient = null;
        }
        // DagClient as such should have no bearing on jobClose.
        if (dagClient != null) {
            // rc will only be overwritten if close errors out
            rc = close(work, rc, dagClient);
        }
    }
    return rc;
}