我在 这篇 文章中介绍了 Apache Spark 3.0 动态分区裁剪（Dynamic Partition Pruning），里面涉及到动态分区的优化思路等，但是并没有涉及到如何使用，本文将介绍在什么情况下会启用动态分区裁剪。

并不是什么查询都会启用动态裁剪优化的，必须满足以下几个条件：

如何使用 Dynamic Partition Pruning

我们先使用 Spark 创建两张表：

spark.range(10000)
     .select(col("id"), col("id").as("k"))
     .write.partitionBy("k")
     .format("parquet")
     .mode("overwrite")
     .saveAsTable("iteblog_tab1")


spark.range(100)
	 .select(col("id"), col("id").as("k"))
	 .write.partitionBy("k")
	 .format("parquet")
	 .mode("overwrite")
	 .saveAsTable("iteblog_tab2")

运行完上面的代码之后，iteblog_tab1 表将产生 10000 个分区，iteblog_tab2 表将产生 100 个分区。我们运行下面的查询语句：

spark.sql("SELECT * FROM iteblog_tab1 t1 JOIN iteblog_tab2 t2 ON t1.k = t2.k AND t2.id < 2").show()

在没有启用动态分区裁剪的情况下 Spark 物理执行计划如下：

== Physical Plan ==
CollectLimit 21
+- *(2) Project [cast(id#0L as string) AS id#12, cast(k#1L as string) AS k#13, cast(id#2L as string) AS id#14, cast(k#3L as string) AS k#15]
   +- *(2) BroadcastHashJoin [k#1L], [k#3L], Inner, BuildRight
      :- *(2) ColumnarToRow
      :  +- FileScan parquet default.iteblog_tab1[id#0L,k#1L] Batched: true, DataFilters: [], Format: Parquet, Location: PrunedInMemoryFileIndex[file:/user/hive/warehouse/iteblog_tab1/k=0, file:/user/hive/warehouse/ite..., PartitionFilters: [isnotnull(k#1L)], PushedFilters: [], ReadSchema: struct<id:bigint>
      +- BroadcastExchange HashedRelationBroadcastMode(List(input[1, bigint, true])), [id=#41]
         +- *(1) Project [id#2L, k#3L]
            +- *(1) Filter (isnotnull(id#2L) AND (id#2L < 2))
               +- *(1) ColumnarToRow
                  +- FileScan parquet default.iteblog_tab2[id#2L,k#3L] Batched: true, DataFilters: [isnotnull(id#2L), (id#2L < 2)], Format: Parquet, Location: PrunedInMemoryFileIndex[file:/user/hive/warehouse/iteblog_tab2/k=0, file:/user/hive/warehouse/ite..., PartitionFilters: [isnotnull(k#3L)], PushedFilters: [IsNotNull(id), LessThan(id,2)], ReadSchema: struct<id:bigint>

DAG 图如下：

如果想及时了解Spark、Hadoop或者Hbase相关的文章，欢迎关注微信公共帐号：iteblog_hadoop

从上面的 物理执行计划和 DAG 执行图可以看出，这个就是正常的算子下推的结果。

如果我们使用 Apache Spark 3.0 并启用动态分区裁减功能，物理执行计划如下：

== Physical Plan ==
CollectLimit 21
+- *(2) Project [cast(id#0L as string) AS id#12, cast(k#1L as string) AS k#13, cast(id#2L as string) AS id#14, cast(k#3L as string) AS k#15]
   +- *(2) BroadcastHashJoin [k#1L], [k#3L], Inner, BuildRight
      :- *(2) ColumnarToRow
      :  +- FileScan parquet default.iteblog_tab1[id#0L,k#1L] Batched: true, DataFilters: [], Format: Parquet, Location: PrunedInMemoryFileIndex[file:/user/hive/warehouse/iteblog_tab1/k=0, file:/user/hive/warehouse/ite..., PartitionFilters: [isnotnull(k#1L), dynamicpruningexpression(k#1L IN dynamicpruning#20)], PushedFilters: [], ReadSchema: struct<id:bigint>
      :        +- SubqueryBroadcast dynamicpruning#20, 0, [k#3L], [id=#96]
      :           +- ReusedExchange [id#2L, k#3L], BroadcastExchange HashedRelationBroadcastMode(List(input[1, bigint, true])), [id=#72]
      +- BroadcastExchange HashedRelationBroadcastMode(List(input[1, bigint, true])), [id=#72]
         +- *(1) Project [id#2L, k#3L]
            +- *(1) Filter (isnotnull(id#2L) AND (id#2L < 2))
               +- *(1) ColumnarToRow
                  +- FileScan parquet default.iteblog_tab2[id#2L,k#3L] Batched: true, DataFilters: [isnotnull(id#2L), (id#2L < 2)], Format: Parquet, Location: PrunedInMemoryFileIndex[file:/user/hive/warehouse/iteblog_tab2/k=0, file:/user/hive/warehouse/ite..., PartitionFilters: [isnotnull(k#3L)], PushedFilters: [IsNotNull(id), LessThan(id,2)], ReadSchema: struct<id:bigint>

DAG 图如下：

如果想及时了解Spark、Hadoop或者Hbase相关的文章，欢迎关注微信公共帐号：iteblog_hadoop

可以看出，iteblog_tab1 表的扫描相比上面那个多了一个分区过滤（PartitionFilters），在一些情况下性能提升能达到 2 - 18 倍。