Creating extensions

Druid uses a module system that allows for the addition of extensions at runtime.

Writing your own extensions

Druid’s extensions leverage Guice in order to add things at runtime. Basically, Guice is a framework for Dependency Injection, but we use it to hold the expected object graph of the Druid process. Extensions can make any changes they want/need to the object graph via adding Guice bindings. While the extensions actually give you the capability to change almost anything however you want, in general, we expect people to want to extend one of the things listed below. This means that we honor our versioning strategy for changes that affect the interfaces called out on this page, but other interfaces are deemed “internal” and can be changed in an incompatible manner even between patch releases.

  1. Add a new deep storage implementation by extending the org.apache.druid.segment.loading.DataSegment* and org.apache.druid.tasklogs.TaskLog* classes.
  2. Add a new input source by extending org.apache.druid.data.input.InputSource.
  3. Add a new input entity by extending org.apache.druid.data.input.InputEntity.
  4. Add a new input source reader if necessary by extending org.apache.druid.data.input.InputSourceReader. You can use org.apache.druid.data.input.impl.InputEntityIteratingReader in most cases.
  5. Add a new input format by extending org.apache.druid.data.input.InputFormat.
  6. Add a new input entity reader by extending org.apache.druid.data.input.TextReader for text formats or org.apache.druid.data.input.IntermediateRowParsingReader for binary formats.
  7. Add Aggregators by extending org.apache.druid.query.aggregation.AggregatorFactory, org.apache.druid.query.aggregation.Aggregator, and org.apache.druid.query.aggregation.BufferAggregator.
  8. Add PostAggregators by extending org.apache.druid.query.aggregation.PostAggregator.
  9. Add ExtractionFns by extending org.apache.druid.query.extraction.ExtractionFn.
  10. Add Complex metrics by extending org.apache.druid.segment.serde.ComplexMetricSerde.
  11. Add new Query types by extending org.apache.druid.query.QueryRunnerFactory, org.apache.druid.query.QueryToolChest, and org.apache.druid.query.Query.
  12. Add new Jersey resources by calling Jerseys.addResource(binder, clazz).
  13. Add new Jetty filters by extending org.apache.druid.server.initialization.jetty.ServletFilterHolder.
  14. Add new secret providers by extending org.apache.druid.metadata.PasswordProvider.
  15. Bundle your extension with all the other Druid extensions

Extensions are added to the system via an implementation of org.apache.druid.initialization.DruidModule.

Creating a Druid Module

The DruidModule class is has two methods

  1. A configure(Binder) method
  2. A getJacksonModules() method

The configure(Binder) method is the same method that a normal Guice module would have.

The getJacksonModules() method provides a list of Jackson modules that are used to help initialize the Jackson ObjectMapper instances used by Druid. This is how you add extensions that are instantiated via Jackson (like AggregatorFactory and InputSource objects) to Druid.

Registering your Druid Module

Once you have your DruidModule created, you will need to package an extra file in the META-INF/services directory of your jar. This is easiest to accomplish with a maven project by creating files in the src/main/resources directory. There are examples of this in the Druid code under the cassandra-storage, hdfs-storage and s3-extensions modules, for examples.

The file that should exist in your jar is

META-INF/services/org.apache.druid.initialization.DruidModule

It should be a text file with a new-line delimited list of package-qualified classes that implement DruidModule like

  1. org.apache.druid.storage.cassandra.CassandraDruidModule

If your jar has this file, then when it is added to the classpath or as an extension, Druid will notice the file and will instantiate instances of the Module. Your Module should have a default constructor, but if you need access to runtime configuration properties, it can have a method with @Inject on it to get a Properties object injected into it from Guice.

Adding a new deep storage implementation

Check the azure-storage, google-storage, cassandra-storage, hdfs-storage and s3-extensions modules for examples of how to do this.

The basic idea behind the extension is that you need to add bindings for your DataSegmentPusher and DataSegmentPuller objects. The way to add them is something like (taken from HdfsStorageDruidModule)

  1. Binders.dataSegmentPullerBinder(binder)
  2.        .addBinding("hdfs")
  3.        .to(HdfsDataSegmentPuller.class).in(LazySingleton.class);
  5. Binders.dataSegmentPusherBinder(binder)
  6.        .addBinding("hdfs")
  7.        .to(HdfsDataSegmentPusher.class).in(LazySingleton.class);

Binders.dataSegment*Binder() is a call provided by the druid-core jar which sets up a Guice multibind “MapBinder”. If that doesn’t make sense, don’t worry about it, just think of it as a magical incantation.

addBinding("hdfs") for the Puller binder creates a new handler for loadSpec objects of type “hdfs”. For the Pusher binder it creates a new type value that you can specify for the druid.storage.type parameter.

to(...).in(...); is normal Guice stuff.

In addition to DataSegmentPusher and DataSegmentPuller, you can also bind:

  • DataSegmentKiller: Removes segments, used as part of the Kill Task to delete unused segments, i.e. perform garbage collection of segments that are either superseded by newer versions or that have been dropped from the cluster.
  • DataSegmentMover: Allow migrating segments from one place to another, currently this is only used as part of the MoveTask to move unused segments to a different S3 bucket or prefix, typically to reduce storage costs of unused data (e.g. move to glacier or cheaper storage)
  • DataSegmentArchiver: Just a wrapper around Mover, but comes with a pre-configured target bucket/path, so it doesn’t have to be specified at runtime as part of the ArchiveTask.

Validating your deep storage implementation

WARNING! This is not a formal procedure, but a collection of hints to validate if your new deep storage implementation is able do push, pull and kill segments.

It’s recommended to use batch ingestion tasks to validate your implementation. The segment will be automatically rolled up to Historical note after ~20 seconds. In this way, you can validate both push (at realtime process) and pull (at Historical process) segments.

  • DataSegmentPusher

Wherever your data storage (cloud storage service, distributed file system, etc.) is, you should be able to see one new file: index.zip (partitionNum_index.zip for HDFS data storage) after your ingestion task ends.

  • DataSegmentPuller

After ~20 secs your ingestion task ends, you should be able to see your Historical process trying to load the new segment.

The following example was retrieved from a Historical process configured to use Azure for deep storage:

  1. 2015-04-14T02:42:33,450 INFO [ZkCoordinator-0] org.apache.druid.server.coordination.ZkCoordinator - New request[LOAD: dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00
  2. .000Z_2015-04-14T02:41:09.484Z] with zNode[/druid/dev/loadQueue/192.168.33.104:8081/dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z_2015-04-14T02:41:09.
  3. 484Z].
  4. 2015-04-14T02:42:33,451 INFO [ZkCoordinator-0] org.apache.druid.server.coordination.ZkCoordinator - Loading segment dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.0
  5. 00Z_2015-04-14T02:41:09.484Z
  6. 2015-04-14T02:42:33,463 INFO [ZkCoordinator-0] org.apache.druid.guice.JsonConfigurator - Loaded class[class org.apache.druid.storage.azure.AzureAccountConfig] from props[drui
  7. d.azure.] as [org.apache.druid.storage.azure.AzureAccountConfig@759c9ad9]
  8. 2015-04-14T02:49:08,275 INFO [ZkCoordinator-0] org.apache.druid.utils.CompressionUtils - Unzipping file[/opt/druid/tmp/compressionUtilZipCache1263964429587449785.z
  9. ip] to [/opt/druid/zk_druid/dde/2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z/2015-04-14T02:41:09.484Z/0]
  10. 2015-04-14T02:49:08,276 INFO [ZkCoordinator-0] org.apache.druid.storage.azure.AzureDataSegmentPuller - Loaded 1196 bytes from [dde/2015-01-02T00:00:00.000Z_2015-01-03
  11. T00:00:00.000Z/2015-04-14T02:41:09.484Z/0/index.zip] to [/opt/druid/zk_druid/dde/2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z/2015-04-14T02:41:09.484Z/0]
  12. 2015-04-14T02:49:08,277 WARN [ZkCoordinator-0] org.apache.druid.segment.loading.SegmentLoaderLocalCacheManager - Segment [dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z_2015-04-14T02:41:09.484Z] is different than expected size. Expected [0] found [1196]
  13. 2015-04-14T02:49:08,282 INFO [ZkCoordinator-0] org.apache.druid.server.coordination.BatchDataSegmentAnnouncer - Announcing segment[dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z_2015-04-14T02:41:09.484Z] at path[/druid/dev/segments/192.168.33.104:8081/192.168.33.104:8081_historical__default_tier_2015-04-14T02:49:08.282Z_7bb87230ebf940188511dd4a53ffd7351]
  14. 2015-04-14T02:49:08,292 INFO [ZkCoordinator-0] org.apache.druid.server.coordination.ZkCoordinator - Completed request [LOAD: dde_2015-01-02T00:00:00.000Z_2015-01-03T00:00:00.000Z_2015-04-14T02:41:09.484Z]
  • DataSegmentKiller

The easiest way of testing the segment killing is marking a segment as not used and then starting a killing task through the old Coordinator console.

To mark a segment as not used, you need to connect to your metadata storage and update the used column to false on the segment table rows.

To start a segment killing task, you need to access the old Coordinator console http://<COODRINATOR_IP>:<COORDINATOR_PORT/old-console/kill.html then select the appropriate datasource and then input a time range (e.g. 2000/3000).

After the killing task ends, index.zip (partitionNum_index.zip for HDFS data storage) file should be deleted from the data storage.

Adding support for a new input source

Adding support for a new input source requires to implement three interfaces, i.e., InputSource, InputEntity, and InputSourceReader. InputSource is to define where the input data is stored. InputEntity is to define how data can be read in parallel in native parallel indexing. InputSourceReader defines how to read your new input source and you can simply use the provided InputEntityIteratingReader in most cases.

There is an example of this in the druid-s3-extensions module with the S3InputSource and S3Entity.

Adding an InputSource is done almost entirely through the Jackson Modules instead of Guice. Specifically, note the implementation

  1. @Override
  2. public List<? extends Module> getJacksonModules()
  3. {
  4.   return ImmutableList.of(
  5.           new SimpleModule().registerSubtypes(new NamedType(S3InputSource.class, "s3"))
  6.   );
  7. }

This is registering the InputSource with Jackson’s polymorphic serialization/deserialization layer. More concretely, having this will mean that if you specify a "inputSource": { "type": "s3", ... } in your IO config, then the system will load this InputSource for your InputSource implementation.

Note that inside of Druid, we have made the @JacksonInject annotation for Jackson deserialized objects actually use the base Guice injector to resolve the object to be injected. So, if your InputSource needs access to some object, you can add a @JacksonInject annotation on a setter and it will get set on instantiation.

Adding support for a new data format

Adding support for a new data format requires implementing two interfaces, i.e., InputFormat and InputEntityReader. InputFormat is to define how your data is formatted. InputEntityReader is to define how to parse your data and convert into Druid InputRow.

There is an example in the druid-orc-extensions module with the OrcInputFormat and OrcReader.

Adding an InputFormat is very similar to adding an InputSource. They operate purely through Jackson and thus should just be additions to the Jackson modules returned by your DruidModule.

Adding Aggregators

Adding AggregatorFactory objects is very similar to InputSource objects. They operate purely through Jackson and thus should just be additions to the Jackson modules returned by your DruidModule.

Adding Complex Metrics

Adding ComplexMetrics is a little ugly in the current version. The method of getting at complex metrics is through registration with the ComplexMetrics.registerSerde() method. There is no special Guice stuff to get this working, just in your configure(Binder) method register the serialization/deserialization.

Adding new Query types

Adding a new Query type requires the implementation of three interfaces.

  1. org.apache.druid.query.Query
  2. org.apache.druid.query.QueryToolChest
  3. org.apache.druid.query.QueryRunnerFactory

Registering these uses the same general strategy as a deep storage mechanism does. You do something like

  1. DruidBinders.queryToolChestBinder(binder)
  2.             .addBinding(SegmentMetadataQuery.class)
  3.             .to(SegmentMetadataQueryQueryToolChest.class);
  5. DruidBinders.queryRunnerFactoryBinder(binder)
  6.             .addBinding(SegmentMetadataQuery.class)
  7.             .to(SegmentMetadataQueryRunnerFactory.class);

The first one binds the SegmentMetadataQueryQueryToolChest for usage when a SegmentMetadataQuery is used. The second one does the same thing but for the QueryRunnerFactory instead.

Adding new Jersey resources

Adding new Jersey resources to a module requires calling the following code to bind the resource in the module:

  1. Jerseys.addResource(binder, NewResource.class);

Adding a new Password Provider implementation

You will need to implement org.apache.druid.metadata.PasswordProvider interface. For every place where Druid uses PasswordProvider, a new instance of the implementation will be created, thus make sure all the necessary information required for fetching each password is supplied during object instantiation. In your implementation of org.apache.druid.initialization.DruidModule, getJacksonModules should look something like this -

  1.     return ImmutableList.of(
  2.         new SimpleModule("SomePasswordProviderModule")
  3.             .registerSubtypes(
  4.                 new NamedType(SomePasswordProvider.class, "some")
  5.             )
  6.     );

where SomePasswordProvider is the implementation of PasswordProvider interface, you can have a look at org.apache.druid.metadata.EnvironmentVariablePasswordProvider for example.

Bundle your extension with all the other Druid extensions

When you do mvn install, Druid extensions will be packaged within the Druid tarball and extensions directory, which are both underneath distribution/target/.

If you want your extension to be included, you can add your extension’s maven coordinate as an argument at distribution/pom.xml

During mvn install, maven will install your extension to the local maven repository, and then call pull-deps to pull your extension from there. In the end, you should see your extension underneath distribution/target/extensions and within Druid tarball.

Managing dependencies

Managing library collisions can be daunting for extensions which draw in commonly used libraries. Here is a list of group IDs for libraries that are suggested to be specified with a provided scope to prevent collision with versions used in druid:

  1. "org.apache.druid",
  2. "com.metamx.druid",
  3. "asm",
  4. "org.ow2.asm",
  5. "org.jboss.netty",
  6. "com.google.guava",
  7. "com.google.code.findbugs",
  8. "com.google.protobuf",
  9. "com.esotericsoftware.minlog",
  10. "log4j",
  11. "org.slf4j",
  12. "commons-logging",
  13. "org.eclipse.jetty",
  14. "org.mortbay.jetty",
  15. "com.sun.jersey",
  16. "com.sun.jersey.contribs",
  17. "common-beanutils",
  18. "commons-codec",
  19. "commons-lang",
  20. "commons-cli",
  21. "commons-io",
  22. "javax.activation",
  23. "org.apache.httpcomponents",
  24. "org.apache.zookeeper",
  25. "org.codehaus.jackson",
  26. "com.fasterxml.jackson",
  27. "com.fasterxml.jackson.core",
  28. "com.fasterxml.jackson.dataformat",
  29. "com.fasterxml.jackson.datatype",
  30. "org.roaringbitmap",
  31. "net.java.dev.jets3t"

See the documentation in org.apache.druid.cli.PullDependencies for more information.