Hey Yong,
It seems that Hadoop `FileSystem` adds the size of a block to the
metrics even if you only touch a fraction of it (reading Parquet
metadata for example). This behavior can be verified by the following
snippet:
```scala
import org.apache.spark.sql.Row
import org.apache.spark.sql.SQLContext
val sqlContext = new SQLContext(sc)
import sc._
import sqlContext._
case class KeyValue(key: Int, value: String)
parallelize(1 to 1024 * 1024 * 20).
flatMap(i => Seq.fill(10)(KeyValue(i, i.toString))).
saveAsParquetFile("large.parquet")
hadoopConfiguration.set("parquet.task.side.metadata", "true")
sql("SET spark.sql.parquet.filterPushdown=true")
parquetFile("large.parquet").where('key ===
0).queryExecution.toRdd.mapPartitions { _ =>
new Iterator[Row] {
def hasNext = false
def next() = ???
}
}.collect()
```
Apparently we’re reading nothing here (except for Parquet metadata in
the footers), but the web UI still suggests that the input size of all
tasks equals to the file size.
Cheng
On 3/10/15 3:15 AM, java8964 wrote:
Hi, Currently most of the data in our production is using Avro +
Snappy. I want to test the benefits if we store the data in Parquet
format. I changed the our ETL to generate the Parquet format, instead
of Avor, and want to test a simple sql in Spark SQL, to verify the
benefits from Parquet.
I generated the same dataset in both Avro and Parquet in HDFS, and
load them both in Spark-SQL. Now I run the same query like
"select colum1 from src_table_avro/parqut where colum2=xxx", I can see
that for the parquet data format, the job runs much fast. The test
files size for both format are around 930M. So Avro job generated 8
tasks to read the data with 21s as the median duration, vs parquet job
generate 7 tasks to read the data with 0.4s as the median duration.
Since the dataset has more than 100 columns, I can see the parquet
file really coming with fast read. But my question is that from the
spark UI, both job show 900M as the input size, and 0 for rest, in
this case, how do I know column pruning really works? I think it is
due to that, so parquet file can be read so fast, but is there any
statistic can prove that to me on the Spark UI? Something like the
input total file size is 900M, but only 10M really read due to column
pruning? So in case that the columns pruning not work in parquet due
to what kind of SQL query, I can identify in the first place.
Thanks
Yong