Druid indexing plugin for using Spark in batch jobs
This repository holds a Druid extension for using Spark as the engine for running batch jobs
To build issue the command sbt clean test publish-local publish-m2
The default properties injected into spark are as follows:
.set("spark.executor.memory", "7G")
.set("spark.executor.cores", "1")
.set("spark.kryo.referenceTracking", "false")
.set("user.timezone", "UTC")
.set("file.encoding", "UTF-8")
.set("java.util.logging.manager", "org.apache.logging.log4j.jul.LogManager")
.set("org.jboss.logging.provider", "slf4j")
There are four key things that need configured to use this extension
- Overlord needs the
druid-spark-batchextension added. - MiddleManager (if present) needs the
druid-spark-batchextension added. - A task json needs configured.
- Spark is included in the default hadoop coordinates similar to
druid.indexer.task.defaultHadoopCoordinates=["org.apache.spark:spark-core_2.10:1.5.2-mmx1"]
To load the extension, use the appropriate coordinates (for druid 0.8.x the following should be added to druid.extensions.coordinates : io.druid.extensions:druid-spark-batch_2.10:jar:assembly:0.0.13) or make certain the extension jars are located in the proper directories (druid 0.9.0 with version 0.9.0.x of this library, druid 0.9.1 with the 0.9.1.x version)
The recommended method of pulling down the extensions is to use pull-deps to pull down the versions of interest. A Hadoop coordinate and an extension should be specified as per -h org.apache.spark:spark-core_2.10:1.5.2-mmx4 and -c io.druid.extensions:druid-spark-batch_2.10:0.9.1-0 (with the appropriate versions of course)
The following is an example spark batch task for the indexing service:
{
"paths":["/<your-druid-spark-batch-dir>/src/test/resources/lineitem.small.tbl"],
"dataSchema": {
"dataSource": "sparkTest",
"granularitySpec": {
"intervals": [
"1992-01-01T00:00:00.000Z/1999-01-01T00:00:00.000Z"
],
"queryGranularity": {
"type": "all"
},
"segmentGranularity": "YEAR",
"type": "uniform"
},
"metricsSpec": [
{
"name": "count",
"type": "count"
},
{
"fieldName": "l_quantity",
"name": "L_QUANTITY_longSum",
"type": "longSum"
},
{
"fieldName": "l_extendedprice",
"name": "L_EXTENDEDPRICE_doubleSum",
"type": "doubleSum"
},
{
"fieldName": "l_discount",
"name": "L_DISCOUNT_doubleSum",
"type": "doubleSum"
},
{
"fieldName": "l_tax",
"name": "L_TAX_doubleSum",
"type": "doubleSum"
}
],
"parser": {
"encoding": "UTF-8",
"parseSpec": {
"columns": [
"l_orderkey",
"l_partkey",
"l_suppkey",
"l_linenumber",
"l_quantity",
"l_extendedprice",
"l_discount",
"l_tax",
"l_returnflag",
"l_linestatus",
"l_shipdate",
"l_commitdate",
"l_receiptdate",
"l_shipinstruct",
"l_shipmode",
"l_comment"
],
"delimiter": "|",
"dimensionsSpec": {
"dimensionExclusions": [
"l_tax",
"l_quantity",
"count",
"l_extendedprice",
"l_shipdate",
"l_discount"
],
"dimensions": [
"l_comment",
"l_commitdate",
"l_linenumber",
"l_linestatus",
"l_orderkey",
"l_receiptdate",
"l_returnflag",
"l_shipinstruct",
"l_shipmode",
"l_suppkey"
],
"spatialDimensions": []
},
"format": "tsv",
"listDelimiter": ",",
"timestampSpec": {
"column": "l_shipdate",
"format": "yyyy-MM-dd",
"missingValue": null
}
},
"type": "string"
}
},
"indexSpec": {
"bitmap": {
"type": "concise"
},
"dimensionCompression": "lz4",
"metricCompression": "lz4"
},
"intervals": ["1992-01-01T00:00:00.000Z/1999-01-01T00:00:00.000Z"],
"master": "local[1]",
"properties": {
"some.property": "someValue",
"spark.io.compression.codec":"org.apache.spark.io.LZ4CompressionCodec"
},
"targetPartitionSize": 10000000,
"type": "index_spark_2.11"
}The json keys accepted by the spark batch indexer are described below
| Field | Type | Required | Default | Description |
|---|---|---|---|---|
type |
String | Yes, index_spark |
N/A | Must be index_spark |
paths |
List of strings | Yes | N/A | A list of hadoop-readable input files. The values are joined with a , and used as a SparkContext.textFile |
dataSchema |
DataSchema | Yes | N/A | The data schema to use |
intervals |
List of strings | Yes | N/A | A list of ISO intervals to be indexed. ALL data for these intervals MUST be present in paths |
maxRowsInMemory |
positive integer | No | 75000 |
Maximum number of rows to store in memory before an intermediate flush to disk |
targetPartitionSize |
positive integer | No | 5000000 |
The target number of rows per partition per segment granularity |
master |
String | No | master[1] |
The spark master URI |
properties |
Map | No | none | A map of string key/value pairs to inject into the SparkContext properties overriding any prior set values |
id |
String | No | Assigned based on dataSource, intervals, and DateTime.now() |
The ID for the task. If not provied it will be assigned |
indexSpec |
InputSpec | No | concise, lz4, lz4 | The InputSpec containing the various compressions to be used |
context |
Map | No | none | The task context |
hadoopDependencyCoordinates |
List of strings | No | null (use default set by druid config) |
The spark dependency coordinates to load in the ClassLoader when launching the task |
buildV9Directly |
Boolean | No | False | Build v9 index directly instead of building v8 index and converting it to v9 format. |
This project uses cross-building in SBT. Both 2.10 and 2.11 versions can be built and deployed with sbt release
For setting repository credentials to be able to publish a release, refer to https://stackoverflow.com/a/19598435
There is now a version for scala 2.10 and scala 2.11. Only ONE of which may be used at any given time.