Version: 2.1.1

Storm HDFS Integration

Storm components for interacting with HDFS file systems

HDFS Bolt

Usage

The following example will write pipe("|")-delimited files to the HDFS path hdfs://localhost:54310/foo. After every 1,000 tuples it will sync filesystem, making that data visible to other HDFS clients. It will rotate files when they reach 5 megabytes in size.

// use "|" instead of "," for field delimiter
RecordFormat format = new DelimitedRecordFormat()
        .withFieldDelimiter("|");

// sync the filesystem after every 1k tuples
SyncPolicy syncPolicy = new CountSyncPolicy(1000);

// rotate files when they reach 5MB
FileRotationPolicy rotationPolicy = new FileSizeRotationPolicy(5.0f, Units.MB);

FileNameFormat fileNameFormat = new DefaultFileNameFormat()
        .withPath("/foo/");

HdfsBolt bolt = new HdfsBolt()
        .withFsUrl("hdfs://localhost:54310")
        .withFileNameFormat(fileNameFormat)
        .withRecordFormat(format)
        .withRotationPolicy(rotationPolicy)
        .withSyncPolicy(syncPolicy);

Packaging a Topology

When packaging your topology, it's important that you use the maven-shade-plugin as opposed to the maven-assembly-plugin.

The shade plugin provides facilities for merging JAR manifest entries, which the hadoop client leverages for URL scheme resolution.

If you experience errors such as the following:

java.lang.RuntimeException: Error preparing HdfsBolt: No FileSystem for scheme: hdfs

it's an indication that your topology jar file isn't packaged properly.

If you are using maven to create your topology jar, you should use the following maven-shade-plugin configuration to create your topology jar:

<plugin>
    <groupId>org.apache.maven.plugins</groupId>
    <artifactId>maven-shade-plugin</artifactId>
    <version>1.4</version>
    <configuration>
        <createDependencyReducedPom>true</createDependencyReducedPom>
    </configuration>
    <executions>
        <execution>
            <phase>package</phase>
            <goals>
                <goal>shade</goal>
            </goals>
            <configuration>
                <transformers>
                    <transformer
                            implementation="org.apache.maven.plugins.shade.resource.ServicesResourceTransformer"/>
                    <transformer
                            implementation="org.apache.maven.plugins.shade.resource.ManifestResourceTransformer">
                        <mainClass></mainClass>
                    </transformer>
                </transformers>
            </configuration>
        </execution>
    </executions>
</plugin>

Specifying a Hadoop Version

By default, storm-hdfs uses the following Hadoop dependencies:

<dependency>
    <groupId>org.apache.hadoop</groupId>
    <artifactId>hadoop-client</artifactId>
    <version>2.6.1</version>
    <exclusions>
        <exclusion>
            <groupId>org.slf4j</groupId>
            <artifactId>slf4j-log4j12</artifactId>
        </exclusion>
    </exclusions>
</dependency>
<dependency>
    <groupId>org.apache.hadoop</groupId>
    <artifactId>hadoop-hdfs</artifactId>
    <version>2.6.1</version>
    <exclusions>
        <exclusion>
            <groupId>org.slf4j</groupId>
            <artifactId>slf4j-log4j12</artifactId>
        </exclusion>
    </exclusions>
</dependency>

If you are using a different version of Hadoop, you should exclude the Hadoop libraries from the storm-hdfs dependency and add the dependencies for your preferred version in your pom.

Hadoop client version incompatibilites can manifest as errors like:

com.google.protobuf.InvalidProtocolBufferException: Protocol message contained an invalid tag (zero)

HDFS Bolt Customization

Record Formats

Record format can be controlled by providing an implementation of the org.apache.storm.hdfs.format.RecordFormat interface:

public interface RecordFormat extends Serializable {
    byte[] format(Tuple tuple);
}

The provided org.apache.storm.hdfs.format.DelimitedRecordFormat is capable of producing formats such as CSV and tab-delimited files.

File Naming

File naming can be controlled by providing an implementation of the org.apache.storm.hdfs.format.FileNameFormat interface:

public interface FileNameFormat extends Serializable {
    void prepare(Map conf, TopologyContext topologyContext);
    String getName(long rotation, long timeStamp);
    String getPath();
}

The provided org.apache.storm.hdfs.format.DefaultFileNameFormat will create file names with the following format:

 {prefix}{componentId}-{taskId}-{rotationNum}-{timestamp}{extension}

For example:

 MyBolt-5-7-1390579837830.txt

By default, prefix is empty and extenstion is ".txt".

New FileNameFormat:

The new provided org.apache.storm.hdfs.format.SimpleFileNameFormat and org.apache.storm.hdfs.trident.format.SimpleFileNameFormat are more flexible, and the withName method support parameters as following:

  • $TIME - current time. use withTimeFormat to format.
  • $NUM - rotation number
  • $HOST - local host name
  • $PARTITION - partition index (org.apache.storm.hdfs.trident.format.SimpleFileNameFormat only)
  • $COMPONENT - component id (org.apache.storm.hdfs.format.SimpleFileNameFormat only)
  • $TASK - task id (org.apache.storm.hdfs.format.SimpleFileNameFormat only)

eg: seq.$TIME.$HOST.$COMPONENT.$NUM.dat

The default file name is $TIME.$NUM.txt, and the default timeFormat is yyyyMMddHHmmss.

Sync Policies

Sync policies allow you to control when buffered data is flushed to the underlying filesystem (thus making it available to clients reading the data) by implementing the org.apache.storm.hdfs.sync.SyncPolicy interface:

public interface SyncPolicy extends Serializable {
    boolean mark(Tuple tuple, long offset);
    void reset();
}

The HdfsBolt will call the mark() method for every tuple it processes. Returning true will trigger the HdfsBolt to perform a sync/flush, after which it will call the reset() method.

The org.apache.storm.hdfs.sync.CountSyncPolicy class simply triggers a sync after the specified number of tuples have been processed.

File Rotation Policies

Similar to sync policies, file rotation policies allow you to control when data files are rotated by providing a org.apache.storm.hdfs.rotation.FileRotation interface:

public interface FileRotationPolicy extends Serializable {
    boolean mark(Tuple tuple, long offset);
    void reset();
    FileRotationPolicy copy();
}

The org.apache.storm.hdfs.rotation.FileSizeRotationPolicy implementation allows you to trigger file rotation when data files reach a specific file size:

FileRotationPolicy rotationPolicy = new FileSizeRotationPolicy(5.0f, Units.MB);

File Rotation Actions

Both the HDFS bolt and Trident State implementation allow you to register any number of RotationActions. What RotationActions do is provide a hook to allow you to perform some action right after a file is rotated. For example, moving a file to a different location or renaming it.

public interface RotationAction extends Serializable {
    void execute(FileSystem fileSystem, Path filePath) throws IOException;
}

Storm-HDFS includes a simple action that will move a file after rotation:

public class MoveFileAction implements RotationAction {
    private static final Logger LOG = LoggerFactory.getLogger(MoveFileAction.class);

    private String destination;

    public MoveFileAction withDestination(String destDir){
        destination = destDir;
        return this;
    }

    @Override
    public void execute(FileSystem fileSystem, Path filePath) throws IOException {
        Path destPath = new Path(destination, filePath.getName());
        LOG.info("Moving file {} to {}", filePath, destPath);
        boolean success = fileSystem.rename(filePath, destPath);
        return;
    }
}

If you are using Trident and sequence files you can do something like this:

        HdfsState.Options seqOpts = new HdfsState.SequenceFileOptions()
                .withFileNameFormat(fileNameFormat)
                .withSequenceFormat(new DefaultSequenceFormat("key", "data"))
                .withRotationPolicy(rotationPolicy)
                .withFsUrl("hdfs://localhost:54310")
                .addRotationAction(new MoveFileAction().withDestination("/dest2/"));

Data Partitioning

Data can be partitioned to different HDFS directories based on characteristics of the tuple being processed or purely external factors, such as system time. To partition your your data, write a class that implements the Partitioner interface and pass it to the withPartitioner() method of your bolt. The getPartitionPath() method returns a partition path for a given tuple.

Here's an example of a Partitioner that operates on a specific field of data:


    Partitioner partitoner = new Partitioner() {
            @Override
            public String getPartitionPath(Tuple tuple) {
                return Path.SEPARATOR + tuple.getStringByField("city");
            }
     };

HDFS Bolt Support for HDFS Sequence Files

The org.apache.storm.hdfs.bolt.SequenceFileBolt class allows you to write storm data to HDFS sequence files:

        // sync the filesystem after every 1k tuples
        SyncPolicy syncPolicy = new CountSyncPolicy(1000);

        // rotate files when they reach 5MB
        FileRotationPolicy rotationPolicy = new FileSizeRotationPolicy(5.0f, Units.MB);

        FileNameFormat fileNameFormat = new DefaultFileNameFormat()
                .withExtension(".seq")
                .withPath("/data/");

        // create sequence format instance.
        DefaultSequenceFormat format = new DefaultSequenceFormat("timestamp", "sentence");

        SequenceFileBolt bolt = new SequenceFileBolt()
                .withFsUrl("hdfs://localhost:54310")
                .withFileNameFormat(fileNameFormat)
                .withSequenceFormat(format)
                .withRotationPolicy(rotationPolicy)
                .withSyncPolicy(syncPolicy)
                .withCompressionType(SequenceFile.CompressionType.RECORD)
                .withCompressionCodec("deflate");

The SequenceFileBolt requires that you provide a org.apache.storm.hdfs.bolt.format.SequenceFormat that maps tuples to key/value pairs:

public interface SequenceFormat extends Serializable {
    Class keyClass();
    Class valueClass();

    Writable key(Tuple tuple);
    Writable value(Tuple tuple);
}

HDFS Bolt Support for Avro Files

The org.apache.storm.hdfs.bolt.AvroGenericRecordBolt class allows you to write Avro objects directly to HDFS:

        // sync the filesystem after every 1k tuples
        SyncPolicy syncPolicy = new CountSyncPolicy(1000);

        // rotate files when they reach 5MB
        FileRotationPolicy rotationPolicy = new FileSizeRotationPolicy(5.0f, Units.MB);

        FileNameFormat fileNameFormat = new DefaultFileNameFormat()
                .withExtension(".avro")
                .withPath("/data/");

        // create sequence format instance.
        DefaultSequenceFormat format = new DefaultSequenceFormat("timestamp", "sentence");

        AvroGenericRecordBolt bolt = new AvroGenericRecordBolt()
                .withFsUrl("hdfs://localhost:54310")
                .withFileNameFormat(fileNameFormat)
                .withRotationPolicy(rotationPolicy)
                .withSyncPolicy(syncPolicy);

The avro bolt will write records to separate files based on the schema of the record being processed. In other words, if the bolt receives records with two different schemas, it will write to two separate files. Each file will be rotatated in accordance with the specified rotation policy. If a large number of Avro schemas are expected, then the bolt should be configured with a maximum number of open files at least equal to the number of schemas expected to prevent excessive file open/close/create operations.

To use this bolt you must register the appropriate Kryo serializers with your topology configuration. A convenience method is provided for this:

AvroUtils.addAvroKryoSerializations(conf);

By default Storm will use the GenericAvroSerializer to handle serialization. This will work, but there are much faster options available if you can pre-define the schemas you will be using or utilize an external schema registry. An implementation using the Confluent Schema Registry is provided, but others can be implemented and provided to Storm. Please see the javadoc for classes in org.apache.storm.hdfs.avro for information about using the built-in options or creating your own.

HDFS Bolt support for Trident API

storm-hdfs also includes a Trident state implementation for writing data to HDFS, with an API that closely mirrors that of the bolts.

         Fields hdfsFields = new Fields("field1", "field2");

         FileNameFormat fileNameFormat = new DefaultFileNameFormat()
                 .withPath("/trident")
                 .withPrefix("trident")
                 .withExtension(".txt");

         RecordFormat recordFormat = new DelimitedRecordFormat()
                 .withFields(hdfsFields);

         FileRotationPolicy rotationPolicy = new FileSizeRotationPolicy(5.0f, FileSizeRotationPolicy.Units.MB);

        HdfsState.Options options = new HdfsState.HdfsFileOptions()
                .withFileNameFormat(fileNameFormat)
                .withRecordFormat(recordFormat)
                .withRotationPolicy(rotationPolicy)
                .withFsUrl("hdfs://localhost:54310");

         StateFactory factory = new HdfsStateFactory().withOptions(options);

         TridentState state = stream
                 .partitionPersist(factory, hdfsFields, new HdfsUpdater(), new Fields());

To use the sequence file State implementation, use the HdfsState.SequenceFileOptions:

        HdfsState.Options seqOpts = new HdfsState.SequenceFileOptions()
                .withFileNameFormat(fileNameFormat)
                .withSequenceFormat(new DefaultSequenceFormat("key", "data"))
                .withRotationPolicy(rotationPolicy)
                .withFsUrl("hdfs://localhost:54310")
                .addRotationAction(new MoveFileAction().toDestination("/dest2/"));

Note

Whenever a batch is replayed by storm (due to failures), the trident state implementation automatically removes duplicates from the current data file by copying the data up to the last transaction to another file. Since this operation involves a lot of data copy, ensure that the data files are rotated at reasonable sizes with FileSizeRotationPolicy and at reasonable intervals with TimedRotationPolicy so that the recovery can complete within topology.message.timeout.secs.

Also note with TimedRotationPolicy the files are never rotated in the middle of a batch even if the timer ticks, but only when a batch completes so that complete batches can be efficiently recovered in case of failures.

Working with Secure HDFS

If your topology is going to interact with secure HDFS, your bolts/states needs to be authenticated by NameNode. We currently have 2 options to support this:

Using HDFS delegation tokens

Your administrator can configure nimbus to automatically get delegation tokens on behalf of the topology submitter user. The nimbus should be started with following configurations:

nimbus.autocredential.plugins.classes : ["org.apache.storm.hdfs.security.AutoHDFS"]
nimbus.credential.renewers.classes : ["org.apache.storm.hdfs.security.AutoHDFS"]
hdfs.keytab.file: "/path/to/keytab/on/nimbus" (This is the keytab of hdfs super user that can impersonate other users.)
hdfs.kerberos.principal: "superuser@EXAMPLE.com" 
nimbus.credential.renewers.freq.secs : 82800 (23 hours, hdfs tokens needs to be renewed every 24 hours so this value should be less then 24 hours.)
topology.hdfs.uri:"hdfs://host:port" (This is an optional config, by default we will use value of "fs.defaultFS" property specified in hadoop's core-site.xml)

Your topology configuration should have:

topology.auto-credentials :["org.apache.storm.hdfs.common.security.AutoHDFS"]

If nimbus did not have the above configuration you need to add and then restart it. Ensure the hadoop configuration files (core-site.xml and hdfs-site.xml) and the storm-hdfs jar with all the dependencies is present in nimbus's classpath.

As an alternative to adding the configuration files (core-site.xml and hdfs-site.xml) to the classpath, you could specify the configurations as a part of the topology configuration. E.g. in you custom storm.yaml (or -c option while submitting the topology),

hdfsCredentialsConfigKeys : ["cluster1", "cluster2"] (the hdfs clusters you want to fetch the tokens from)
"cluster1": {"config1": "value1", "config2": "value2", ... } (A map of config key-values specific to cluster1)
"cluster2": {"config1": "value1", "hdfs.keytab.file": "/path/to/keytab/for/cluster2/on/nimubs", "hdfs.kerberos.principal": "cluster2user@EXAMPLE.com"} (here along with other configs, we have custom keytab and principal for "cluster2" which will override the keytab/principal specified at topology level)

Instead of specifying key values you may also directly specify the resource files for e.g.,

"cluster1": {"resources": ["/path/to/core-site1.xml", "/path/to/hdfs-site1.xml"]}
"cluster2": {"resources": ["/path/to/core-site2.xml", "/path/to/hdfs-site2.xml"]}

Storm will download the tokens separately for each of the clusters and populate it into the subject and also renew the tokens periodically. This way it would be possible to run multiple bolts connecting to separate HDFS cluster within the same topology.

Nimbus will use the keytab and principal specified in the config to authenticate with Namenode. From then on for every topology submission, nimbus will impersonate the topology submitter user and acquire delegation tokens on behalf of the topology submitter user. If topology was started with topology.auto-credentials set to AutoHDFS, nimbus will push the delegation tokens to all the workers for your topology and the hdfs bolt/state will authenticate with namenode using these tokens.

As nimbus is impersonating topology submitter user, you need to ensure the user specified in hdfs.kerberos.principal has permissions to acquire tokens on behalf of other users. To achieve this you need to follow configuration directions listed on this link http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-common/Superusers.html

You can read about setting up secure HDFS here: http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-common/SecureMode.html.

Using keytabs on all worker hosts

If you have distributed the keytab files for hdfs user on all potential worker hosts then you can use this method. You should specify a hdfs config key using the method HdfsBolt/State.withconfigKey("somekey") and the value map of this key should have following 2 properties:

hdfs.keytab.file: "/path/to/keytab/" hdfs.kerberos.principal: "user@EXAMPLE.com"

On worker hosts the bolt/trident-state code will use the keytab file with principal provided in the config to authenticate with Namenode. This method is little dangerous as you need to ensure all workers have the keytab file at the same location and you need to remember this as you bring up new hosts in the cluster.


HDFS Spout

Hdfs spout is intended to allow feeding data into Storm from a HDFS directory. It will actively monitor the directory to consume any new files that appear in the directory. HDFS spout does not support Trident currently.

Impt: Hdfs spout assumes that the files being made visible to it in the monitored directory are NOT actively being written to. Only after a file is completely written should it be made visible to the spout. This can be achieved by either writing the files out to another directory and once completely written, move it to the monitored directory. Alternatively the file can be created with a '.ignore' suffix in the monitored directory and after data is completely written, rename it without the suffix. File names with a '.ignore' suffix are ignored by the spout.

When the spout is actively consuming a file, it renames the file with a '.inprogress' suffix. After consuming all the contents in the file, the file will be moved to a configurable done directory and the '.inprogress' suffix will be dropped.

Concurrency If multiple spout instances are used in the topology, each instance will consume a different file. Synchronization among spout instances is done using lock files created in a (by default) '.lock' subdirectory under the monitored directory. A file with the same name as the file being consumed (without the in progress suffix) is created in the lock directory. Once the file is completely consumed, the corresponding lock file is deleted.

Recovery from failure Periodically, the spout also records progress information wrt to how much of the file has been consumed in the lock file. In case of an crash of the spout instance (or force kill of topology) another spout can take over the file and resume from the location recorded in the lock file.

Certain error conditions (such spout crashing) can leave behind lock files without deleting them. Such a stale lock file also indicates that the corresponding input file has also not been completely processed. When detected, ownership of such stale lock files will be transferred to another spout.
The configuration 'hdfsspout.lock.timeout.sec' is used to specify the duration of inactivity after which lock files should be considered stale. For lock file ownership transfer to succeed, the HDFS lease on the file (from prev lock owner) should have expired. Spouts scan for stale lock files before selecting the next file for consumption.

Lock on .lock Directory Hdfs spout instances create a DIRLOCK file in the .lock directory to co-ordinate certain accesses to the .lock dir itself. A spout will try to create it when it needs access to the .lock directory and then delete it when done. In error conditions such as a topology crash, force kill or untimely death of a spout, this file may not get deleted. Future running instances of the spout will eventually recover this once the DIRLOCK file becomes stale due to inactivity for hdfsspout.lock.timeout.sec seconds.

Usage

The following example creates an HDFS spout that reads text files from HDFS path hdfs://localhost:54310/source.

// Instantiate spout to read text files
HdfsSpout textReaderSpout = new HdfsSpout().setReaderType("text")
                                           .withOutputFields(TextFileReader.defaultFields)                                      
                                           .setHdfsUri("hdfs://localhost:54310")  // required
                                           .setSourceDir("/data/in")              // required                                      
                                           .setArchiveDir("/data/done")           // required
                                           .setBadFilesDir("/data/badfiles");     // required                                      
// If using Kerberos
HashMap hdfsSettings = new HashMap();
hdfsSettings.put("hdfs.keytab.file", "/path/to/keytab");
hdfsSettings.put("hdfs.kerberos.principal","user@EXAMPLE.com");

textReaderSpout.setHdfsClientSettings(hdfsSettings);

// Create topology
TopologyBuilder builder = new TopologyBuilder();
builder.setSpout("hdfsspout", textReaderSpout, SPOUT_NUM);

// Setup bolts and wire up topology
     ..snip..

// Submit topology with config
Config conf = new Config();
StormSubmitter.submitTopologyWithProgressBar("topologyName", conf, builder.createTopology());

A sample topology HdfsSpoutTopology is provided in storm-starter module.

Configuration Settings

Below is a list of HdfsSpout member functions used for configuration. The equivalent config is also possible via Config object passed in during submitting topology. However, the later mechanism is deprecated as it does not allow multiple Hdfs spouts with differing settings. :

Only methods mentioned in bold are required.

Method Alternative config name (deprecated) Default Description
.setReaderType() ~~hdfsspout.reader.type~~ Determines which file reader to use. Set to 'seq' for reading sequence files or 'text' for text files. Set to a fully qualified class name if using a custom file reader class (that implements interface org.apache.storm.hdfs.spout.FileReader)
.withOutputFields() Sets the names for the output fields for the spout. The number of fields depends upon the reader being used. For convenience, built-in reader types expose a static member called defaultFields that can be used for setting this.
.setHdfsUri() ~~hdfsspout.hdfs~~ HDFS URI for the hdfs Name node. Example: hdfs://namenodehost:8020
.setSourceDir() ~~hdfsspout.source.dir~~ HDFS directory from where to read files. E.g. /data/inputdir
.setArchiveDir() ~~hdfsspout.archive.dir~~ After a file is processed completely it will be moved to this HDFS directory. If this directory does not exist it will be created. E.g. /data/done
.setBadFilesDir() ~~hdfsspout.badfiles.dir~~ if there is an error parsing a file's contents, the file is moved to this location. If this directory does not exist it will be created. E.g. /data/badfiles
.setLockDir() ~~hdfsspout.lock.dir~~ '.lock' subdirectory under hdfsspout.source.dir Dir in which lock files will be created. Concurrent HDFS spout instances synchronize using lock files. Before processing a file the spout instance creates a lock file in this directory with same name as input file and deletes this lock file after processing the file. Spouts also periodically makes a note of their progress (wrt reading the input file) in the lock file so that another spout instance can resume progress on the same file if the spout dies for any reason.
.setIgnoreSuffix() ~~hdfsspout.ignore.suffix~~ .ignore File names with this suffix in the in the hdfsspout.source.dir location will not be processed
.setCommitFrequencyCount() ~~hdfsspout.commit.count~~ 20000 Record progress in the lock file after these many records are processed. If set to 0, this criterion will not be used.
.setCommitFrequencySec() ~~hdfsspout.commit.sec~~ 10 Record progress in the lock file after these many seconds have elapsed. Must be greater than 0
.setMaxOutstanding() ~~hdfsspout.max.outstanding~~ 10000 Limits the number of unACKed tuples by pausing tuple generation (if ACKers are used in the topology)
.setLockTimeoutSec() ~~hdfsspout.lock.timeout.sec~~ 5 minutes Duration of inactivity after which a lock file is considered to be abandoned and ready for another spout to take ownership
.setClocksInSync() ~~hdfsspout.clocks.insync~~ true Indicates whether clocks on the storm machines are in sync (using services like NTP). Used for detecting stale locks.
.withConfigKey() Optional setting. Overrides the default key name ('hdfs.config', see below) used for specifying HDFS client configs.
.setHdfsClientSettings() ~~hdfs.config~~ (unless changed via withConfigKey) Set it to a Map of Key/value pairs indicating the HDFS settings to be used. For example, keytab and principal could be set using this. See section Using keytabs on all worker hosts under HDFS bolt below.
.withOutputStream() Name of output stream. If set, the tuples will be emited to the specified stream. Else tuples will be emited to the default output stream