Examples with SplitLocationInfo org.apache.hadoop.mapred.SplitLocationInfo used on opensource projects

Example 1 with SplitLocationInfo

use of org.apache.hadoop.mapred.SplitLocationInfo 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);
    }
}

Example 2 with SplitLocationInfo

use of org.apache.hadoop.mapred.SplitLocationInfo in project hive by apache.

the class GenericUDTFGetSplits method getSplits.

public InputSplit[] getSplits(JobConf job, int numSplits, TezWork work, Schema schema, ApplicationId applicationId) 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(ctx.getConf()), utils, job);
        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");
        }
        // Update the queryId to use the generated applicationId. See comment below about
        // why this is done.
        HiveConf.setVar(wxConf, HiveConf.ConfVars.HIVEQUERYID, applicationId.toString());
        Vertex wx = utils.createVertex(wxConf, mapWork, scratchDir, fs, ctx, false, work, work.getVertexType(mapWork), DagUtils.createTezLrMap(appJarLr, null));
        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);
        }
        // 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();
        }
        // Generate umbilical token (applies to all splits)
        Token<JobTokenIdentifier> umbilicalToken = JobTokenCreator.createJobToken(applicationId);
        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, umbilicalToken);
            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);
    }
}

Example 3 with SplitLocationInfo

use of org.apache.hadoop.mapred.SplitLocationInfo in project hive by apache.

the class LlapInputSplit method readFields.

@Override
public void readFields(DataInput in) throws IOException {
    splitNum = in.readInt();
    int length = in.readInt();
    planBytes = new byte[length];
    in.readFully(planBytes);
    length = in.readInt();
    fragmentBytes = new byte[length];
    in.readFully(fragmentBytes);
    length = in.readInt();
    if (length > 0) {
        fragmentBytesSignature = new byte[length];
        in.readFully(fragmentBytesSignature);
    }
    length = in.readInt();
    locations = new SplitLocationInfo[length];
    for (int i = 0; i < length; ++i) {
        locations[i] = new SplitLocationInfo(in.readUTF(), false);
    }
    llapDaemonInfos = new LlapDaemonInfo[in.readInt()];
    for (int i = 0; i < llapDaemonInfos.length; i++) {
        llapDaemonInfos[i] = new LlapDaemonInfo();
        llapDaemonInfos[i].readFields(in);
    }
    schema = new Schema();
    schema.readFields(in);
    llapUser = in.readUTF();
    length = in.readInt();
    if (length > 0) {
        tokenBytes = new byte[length];
        in.readFully(tokenBytes);
    }
    jwt = in.readUTF();
}

Example 4 with SplitLocationInfo

use of org.apache.hadoop.mapred.SplitLocationInfo in project hive by apache.

the class GenericUDTFGetSplits method getSplits.

// generateLightWeightSplits - if true then
// 1) schema and planBytes[] in each LlapInputSplit are not populated
// 2) schemaSplit(contains only schema) and planSplit(contains only planBytes[]) are populated in SplitResult
private SplitResult getSplits(JobConf job, TezWork work, Schema schema, ApplicationId extClientAppId, boolean generateLightWeightSplits) throws IOException {
    SplitResult splitResult = new SplitResult();
    splitResult.schemaSplit = new LlapInputSplit(0, new byte[0], new byte[0], new byte[0], new SplitLocationInfo[0], new LlapDaemonInfo[0], schema, "", new byte[0], "");
    if (schemaSplitOnly) {
        // schema only
        return splitResult;
    }
    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);
    try {
        LocalResource appJarLr = createJarLocalResource(utils.getExecJarPathLocal(ctx.getConf()), utils, job);
        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");
        }
        // Update the queryId to use the generated extClientAppId. See comment below about
        // why this is done.
        HiveConf.setVar(wxConf, HiveConf.ConfVars.HIVEQUERYID, extClientAppId.toString());
        Vertex wx = utils.createVertex(wxConf, mapWork, scratchDir, work, DagUtils.createTezLrMap(appJarLr, null));
        String vertexName = wx.getName();
        dag.addVertex(wx);
        utils.addCredentials(mapWork, dag, job);
        // 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, false, inputArgNumSplits);
        List<Event> eventList = splitGenerator.initialize();
        int numGroupedSplitsGenerated = eventList.size() - 1;
        InputSplit[] result = new InputSplit[numGroupedSplitsGenerated];
        InputConfigureVertexTasksEvent configureEvent = (InputConfigureVertexTasksEvent) eventList.get(0);
        List<TaskLocationHint> hints = configureEvent.getLocationHint().getTaskLocationHints();
        Preconditions.checkState(hints.size() == numGroupedSplitsGenerated);
        if (LOG.isDebugEnabled()) {
            LOG.debug("NumEvents=" + eventList.size() + ", NumSplits=" + result.length);
        }
        // This assumes LLAP cluster owner is always the HS2 user.
        String llapUser = LlapRegistryService.currentUser();
        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(extClientAppId.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();
        }
        // Generate umbilical token (applies to all splits)
        Token<JobTokenIdentifier> umbilicalToken = JobTokenCreator.createJobToken(extClientAppId);
        LOG.info("Number of splits: " + numGroupedSplitsGenerated);
        SignedMessage signedSvs = null;
        byte[] submitWorkBytes = null;
        final byte[] emptySubmitWorkBytes = new byte[0];
        final Schema emptySchema = new Schema();
        for (int i = 0; i < numGroupedSplitsGenerated; i++) {
            TaskSpec taskSpec = new TaskSpecBuilder().constructTaskSpec(dag, vertexName, numGroupedSplitsGenerated, extClientAppId, 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, extClientAppId, queryUser, extClientAppId.toString());
                SubmitWorkInfo submitWorkInfo = new SubmitWorkInfo(extClientAppId, System.currentTimeMillis(), numGroupedSplitsGenerated, signedSvs.message, signedSvs.signature, umbilicalToken);
                submitWorkBytes = SubmitWorkInfo.toBytes(submitWorkInfo);
                if (generateLightWeightSplits) {
                    splitResult.planSplit = new LlapInputSplit(0, submitWorkBytes, new byte[0], new byte[0], new SplitLocationInfo[0], new LlapDaemonInfo[0], new Schema(), "", new byte[0], "");
                }
            }
            // 3. Generate input event.
            SignedMessage eventBytes = makeEventBytes(wx, vertexName, eventList.get(i + 1), signer);
            // 4. Make location hints.
            SplitLocationInfo[] locations = makeLocationHints(hints.get(i));
            // 5. populate info about llap daemons(to help client submit request and read data)
            LlapDaemonInfo[] llapDaemonInfos = populateLlapDaemonInfos(job, locations);
            // 6. Generate JWT for external clients if it's a cloud deployment
            // we inject extClientAppId in JWT which is same as what fragment contains.
            // extClientAppId in JWT and in fragment are compared on LLAP when a fragment is submitted.
            // see method ContainerRunnerImpl#verifyJwtForExternalClient
            String jwt = "";
            if (LlapUtil.isCloudDeployment(job)) {
                LlapExtClientJwtHelper llapExtClientJwtHelper = new LlapExtClientJwtHelper(job);
                jwt = llapExtClientJwtHelper.buildJwtForLlap(extClientAppId);
            }
            if (generateLightWeightSplits) {
                result[i] = new LlapInputSplit(i, emptySubmitWorkBytes, eventBytes.message, eventBytes.signature, locations, llapDaemonInfos, emptySchema, llapUser, tokenBytes, jwt);
            } else {
                result[i] = new LlapInputSplit(i, submitWorkBytes, eventBytes.message, eventBytes.signature, locations, llapDaemonInfos, schema, llapUser, tokenBytes, jwt);
            }
        }
        splitResult.actualSplits = result;
        return splitResult;
    } catch (Exception e) {
        throw new IOException(e);
    }
}

Example 5 with SplitLocationInfo

use of org.apache.hadoop.mapred.SplitLocationInfo in project hive by apache.

the class GenericUDTFGetSplits method makeLocationHints.

private SplitLocationInfo[] makeLocationHints(TaskLocationHint hint) {
    Set<String> hosts = hint.getHosts();
    if (hosts == null) {
        LOG.warn("No hosts");
        return new SplitLocationInfo[0];
    }
    if (hosts.size() != 1) {
        LOG.warn("Bad # of locations: " + hosts.size());
    }
    SplitLocationInfo[] locations = new SplitLocationInfo[hosts.size()];
    int j = 0;
    for (String host : hosts) {
        locations[j++] = new SplitLocationInfo(host, false);
    }
    return locations;
}