Skip to main content

Developing on Apache Druid

Druid's codebase consists of several major components. For developers interested in learning the code, this document provides a high level overview of the main components that make up Druid and the relevant classes to start from to learn the code.

Storage format​

Data in Druid is stored in a custom column format known as a segment. Segments are composed of different types of columns. Column.java and the classes that extend it is a great place to looking into the storage format.

Segment creation​

Raw data is ingested in IncrementalIndex.java, and segments are created in IndexMerger.java.

Storage engine​

Druid segments are memory mapped in IndexIO.java to be exposed for querying.

Query engine​

Most of the logic related to Druid queries can be found in the Query* classes. Druid leverages query runners to run queries. Query runners often embed other query runners and each query runner adds on a layer of logic. A good starting point to trace the query logic is to start from QueryResource.java.

Coordination​

Most of the coordination logic for Historical processes is on the Druid Coordinator. The starting point here is DruidCoordinator.java. Most of the coordination logic for (real-time) ingestion is in the Druid indexing service. The starting point here is OverlordResource.java.

Real-time Ingestion​

Druid streaming tasks are based on the 'seekable stream' classes such as SeekableStreamSupervisor.java, SeekableStreamIndexTask.java, and SeekableStreamIndexTaskRunner.java. The data processing happens through StreamAppenderator.java, and the persist and hand-off logic is in StreamAppenderatorDriver.java.

Native Batch Ingestion​

Druid native batch ingestion main task types are based on AbstractBatchTask.java and AbstractBatchSubtask.java. Parallel processing uses ParallelIndexSupervisorTask.java, which spawns subtasks to perform various operations such as data analysis and partitioning depending on the task specification. Segment generation happens in SinglePhaseSubTask.java, PartialHashSegmentGenerateTask.java, or PartialRangeSegmentGenerateTask.java through BatchAppenderator, and the persist and hand-off logic is in BatchAppenderatorDriver.java.

Hadoop-based Batch Ingestion​

The two main Hadoop indexing classes are HadoopDruidDetermineConfigurationJob.java for the job to determine how many Druid segments to create, and HadoopDruidIndexerJob.java, which creates Druid segments.

At some point in the future, we may move the Hadoop ingestion code out of core Druid.

Internal UIs​

Druid currently has two internal UIs. One is for the Coordinator and one is for the Overlord.

At some point in the future, we will likely move the internal UI code out of core Druid.

Client libraries​

We welcome contributions for new client libraries to interact with Druid. See the Community and third-party libraries page for links to existing client libraries.