39623481fc
This PR introduces several performance optimizations to `HadoopFsRelation` and `ParquetRelation2`:
1. Moving `FileStatus` listing from `DataSourceStrategy` into a cache within `HadoopFsRelation`.
This new cache generalizes and replaces the one used in `ParquetRelation2`.
This also introduces an interface change: to reuse cached `FileStatus` objects, `HadoopFsRelation.buildScan` methods now receive `Array[FileStatus]` instead of `Array[String]`.
1. When Parquet task side metadata reading is enabled, skip reading row group information when reading Parquet footers.
This is basically what PR #5334 does. Also, now we uses `ParquetFileReader.readAllFootersInParallel` to read footers in parallel.
Another optimization in question is, instead of asking `HadoopFsRelation.buildScan` to return an `RDD[Row]` for a single selected partition and then union them all, we ask it to return an `RDD[Row]` for all selected partitions. This optimization is based on the fact that Hadoop configuration broadcasting used in `NewHadoopRDD` takes 34% time in the following microbenchmark. However, this complicates data source user code because user code must merge partition values manually.
To check the cost of broadcasting in `NewHadoopRDD`, I also did microbenchmark after removing the `broadcast` call in `NewHadoopRDD`. All results are shown below.
### Microbenchmark
#### Preparation code
Generating a partitioned table with 50k partitions, 1k rows per partition:
```scala
import sqlContext._
import sqlContext.implicits._
for (n <- 0 until 500) {
val data = for {
p <- (n * 10) until ((n + 1) * 10)
i <- 0 until 1000
} yield (i, f"val_$i%04d", f"$p%04d")
data.
toDF("a", "b", "p").
write.
partitionBy("p").
mode("append").
parquet(path)
}
```
#### Benchmarking code
```scala
import sqlContext._
import sqlContext.implicits._
import org.apache.spark.sql.types._
import com.google.common.base.Stopwatch
val path = "hdfs://localhost:9000/user/lian/5k"
def benchmark(n: Int)(f: => Unit) {
val stopwatch = new Stopwatch()
def run() = {
stopwatch.reset()
stopwatch.start()
f
stopwatch.stop()
stopwatch.elapsedMillis()
}
val records = (0 until n).map(_ => run())
(0 until n).foreach(i => println(s"Round $i: ${records(i)} ms"))
println(s"Average: ${records.sum / n.toDouble} ms")
}
benchmark(3) { read.parquet(path).explain(extended = true) }
```
#### Results
Before:
```
Round 0: 72528 ms
Round 1: 68938 ms
Round 2: 65372 ms
Average: 68946.0 ms
```
After:
```
Round 0: 59499 ms
Round 1: 53645 ms
Round 2: 53844 ms
Round 3: 49093 ms
Round 4: 50555 ms
Average: 53327.2 ms
```
Also removing Hadoop configuration broadcasting:
(Note that I was testing on a local laptop, thus network cost is pretty low.)
```
Round 0: 15806 ms
Round 1: 14394 ms
Round 2: 14699 ms
Round 3: 15334 ms
Round 4: 14123 ms
Average: 14871.2 ms
```
Author: Cheng Lian <lian@databricks.com>
Closes #6225 from liancheng/spark-7673 and squashes the following commits:
2d58a2b [Cheng Lian] Skips reading row group information when using task side metadata reading
7aa3748 [Cheng Lian] Optimizes FileStatusCache by introducing a map from parent directories to child files
ba41250 [Cheng Lian] Reuses HadoopFsRelation FileStatusCache in ParquetRelation2
3d278f7 [Cheng Lian] Fixes a bug when reading a single Parquet data file
b84612a [Cheng Lian] Fixes Scala style issue
6a08b02 [Cheng Lian] WIP: Moves file status cache into HadoopFSRelation
(cherry picked from commit
|
||
---|---|---|
.. | ||
src | ||
pom.xml |