这也与 Spark 优化有关。看一个简短的例子:
HDFS 中包含结构和数据的大型 parquet 文件:
[hadoop@hdpnn ~]$ hadoop fs -ls /user/tickers/ticks.parquet
Found 27 items
drwxr-xr-x - root root 0 2019-01-16 12:55 /user/tickers/ticks.parquet/ticker_id=1
drwxr-xr-x - root root 0 2019-01-16 13:58 /user/tickers/ticks.parquet/ticker_id=10
drwxr-xr-x - root root 0 2019-01-16 14:04 /user/tickers/ticks.parquet/ticker_id=11
drwxr-xr-x - root root 0 2019-01-16 14:10 /user/tickers/ticks.parquet/ticker_id=12
...
每个分区内部都有分区(按日期)
[hadoop@hdpnn ~]$ hadoop fs -ls /user/tickers/ticks.parquet/ticker_id=1
Found 6 items
drwxr-xr-x - root root 0 2019-01-16 12:55 /user/tickers/ticks.parquet/ticker_id=1/ddate=2019-01-09
drwxr-xr-x - root root 0 2019-01-16 12:50 /user/tickers/ticks.parquet/ticker_id=1/ddate=2019-01-10
drwxr-xr-x - root root 0 2019-01-16 12:53 /user/tickers/ticks.parquet/ticker_id=1/ddate=2019-01-11
...
结构:
scala> spark.read.parquet("hdfs://hdpnn:9000/user/tickers/ticks.parquet").printSchema
root
|-- ticker_id: integer (nullable = true)
|-- ddate: date (nullable = true)
|-- db_tsunx: long (nullable = true)
|-- ask: double (nullable = true)
|-- bid: double (nullable = true)
例如,您有这样的 DS:
val maxTsunx = spark.read.parquet("hdfs://hdpnn:9000/user/tickers/ticks.parquet").select(col("ticker_id"),col("db_tsunx")).groupBy("ticker_id").agg(max("db_tsunx"))
包含每个ticker_id的max(db_tsunx)
F.E.:您只想从此 DS 中获取一个股票代码的数据
你有两种方法:
1) maxTsunx.filter(r => r.get(0) == 1)
2) maxTsunx.where(col("ticker_id")===1)
这是一个非常不同的“物理计划”
看看
1)
== Physical Plan ==
*(2) Filter <function1>.apply
+- *(2) HashAggregate(keys=[ticker_id#37], functions=[max(db_tsunx#39L)], output=[ticker_id#37, max(db_tsunx)#52L])
+- Exchange hashpartitioning(ticker_id#37, 200)
+- *(1) HashAggregate(keys=[ticker_id#37], functions=[partial_max(db_tsunx#39L)], output=[ticker_id#37, max#61L])
+- *(1) Project [ticker_id#37, db_tsunx#39L]
+- *(1) FileScan parquet [db_tsunx#39L,ticker_id#37,ddate#38] Batched: true, Format: Parquet,
Location: InMemoryFileIndex[hdfs://hdpnn:9000/user/tickers/ticks.parquet],
PartitionCount: 162,
PartitionFilters: [],
PushedFilters: [],
ReadSchema: struct<db_tsunx:bigint>
2)
== Physical Plan ==
*(2) HashAggregate(keys=[ticker_id#84], functions=[max(db_tsunx#86L)], output=[ticker_id#84, max(db_tsunx)#99L])
+- Exchange hashpartitioning(ticker_id#84, 200)
+- *(1) HashAggregate(keys=[ticker_id#84], functions=[partial_max(db_tsunx#86L)], output=[ticker_id#84, max#109L])
+- *(1) Project [ticker_id#84, db_tsunx#86L]
+- *(1) FileScan parquet [db_tsunx#86L,ticker_id#84,ddate#85] Batched: true, Format: Parquet,
Location: InMemoryFileIndex[hdfs://hdpnn:9000/user/tickers/ticks.parquet],
PartitionCount: 6,
PartitionFilters: [isnotnull(ticker_id#84), (ticker_id#84 = 1)],
PushedFilters: [],
ReadSchema: struct<db_tsunx:bigint>
比较 162 和 6 和
分区过滤器:[],
PartitionFilters: [isnotnull(ticker_id#84), (ticker_id#84 = 1)],
这意味着对来自 DS 的数据进行过滤操作,以及在 Spark 内部的何处用于优化。