您可以利用 spark-sql 中的集群功能进行表创建、表连接等,它充当配置单元以避免数据交换和 spark2.1+ 中的排序
见https://issues.apache.org/jira/browse/SPARK-15453
目前 hive 无法识别此功能,因为元数据在 spark 和 hive 之间不兼容,这就是为什么即使在 hive 端识别此表也不能使用相同的语法,这会将所有列视为array
以下示例可能会给您一些想法:
准备源码
val df = (0 until 80000).map(i => (i, i.toString, i.toString)).toDF("item_id", "country", "state").coalesce(1)
从源创建两个桶表
您会看到“与 Hive 不兼容”。向右滚动
df.write.bucketBy(100, "country", "state").sortBy("country", "state").saveAsTable("kofeng.lstg_bucket_test")
17/03/13 15:12:01 WARN HiveExternalCatalog: Persisting bucketed data source table `kofeng`.`lstg_bucket_test` into Hive metastore in Spark SQL specific format, which is NOT compatible with Hive.
df.write.bucketBy(100, "country", "state").sortBy("country", "state").saveAsTable("kofeng.lstg_bucket_test2")
加入他们并解释
由于音量小,请先禁用广播加入。
import org.apache.spark.sql.SparkSession
val spark = SparkSession.builder().appName("Spark SQL basic example").config("spark.sql.autoBroadcastJoinThreshold", "0").getOrCreate()
计划在SPARK 2.1.0中避免交换和排序,在SPARK2.0中避免交换,只过滤和扫描证明数据局部性利用。
val query = """
|SELECT *
|FROM
| kofeng.lstg_bucket_test a
|JOIN
| kofeng.lstg_bucket_test2 b
|ON a.country=b.country AND
| a.state=b.state
""".stripMargin
val joinDF = sql(query)
scala> joinDF.queryExecution.executedPlan
res10: org.apache.spark.sql.execution.SparkPlan =
*SortMergeJoin [country#71, state#72], [country#74, state#75], Inner
:- *Project [item_id#70, country#71, state#72]
: +- *Filter (isnotnull(country#71) && isnotnull(state#72))
: +- *FileScan parquet kofeng.lstg_bucket_test[item_id#70,country#71,state#72] Batched: true, Format: Parquet, Location: InMemoryFileIndex[hdfs://ares-lvs-nn-ha/user/hive/warehouse/kofeng.db/lstg_bucket_test], PartitionFilters: [], PushedFilters: [IsNotNull(country), IsNotNull(state)], ReadSchema: struct<item_id:int,country:int,state:string>
+- *Project [item_id#73, country#74, state#75]
+- *Filter (isnotnull(country#74) && isnotnull(state#75))
+- *FileScan parquet kofeng.lstg_bucket_test2[item_id#73,country#74,state#75] Batched: true, Format: Parquet...