1f50a5875b
## What changes were proposed in this pull request?
Defines a new subclass of UDF: `UserDefinedAggregator`. Also allows `Aggregator` to be registered as a udf. Under the hood, the implementation is based on the internal `TypedImperativeAggregate` class that spark's predefined aggregators make use of. The effect is that custom user defined aggregators are now serialized only on partition boundaries instead of being serialized and deserialized at each input row.
The two new modes of using `Aggregator` are as follows:
```scala
val agg: Aggregator[IN, BUF, OUT] = // typed aggregator
val udaf1 = UserDefinedAggregator(agg)
val udaf2 = spark.udf.register("agg", agg)
```
## How was this patch tested?
Unit testing has been added that corresponds to the testing suites for `UserDefinedAggregateFunction`. Additionally, unit tests explicitly count the number of aggregator ser/de cycles to ensure that it is governed only by the number of data partitions.
To evaluate the performance impact, I did two comparisons.
The code and REPL results are recorded on [this gist](https://gist.github.com/erikerlandson/b0e106a4dbaf7f80b4f4f3a21f05f892)
To characterize its behavior I benchmarked both a relatively simple aggregator and then an aggregator with a complex structure (a t-digest).
### performance
The following compares the new `Aggregator` based aggregation against UDAF. In this scenario, the new aggregation is about 100x faster. The difference in performance impact depends on the complexity of the aggregator. For very simple aggregators (e.g. implementing 'sum', etc), the performance impact is more like 25-30%.
```scala
scala> import scala.util.Random._, org.apache.spark.sql.Row, org.apache.spark.tdigest._
import scala.util.Random._
import org.apache.spark.sql.Row
import org.apache.spark.tdigest._
scala> val data = sc.parallelize(Vector.fill(50000){(nextInt(2), nextGaussian, nextGaussian.toFloat)}, 5).toDF("cat", "x1", "x2")
data: org.apache.spark.sql.DataFrame = [cat: int, x1: double ... 1 more field]
scala> val udaf = TDigestUDAF(0.5, 0)
udaf: org.apache.spark.tdigest.TDigestUDAF = TDigestUDAF(0.5,0)
scala> val bs = Benchmark.sample(10) { data.agg(udaf($"x1"), udaf($"x2")).first }
bs: Array[(Double, org.apache.spark.sql.Row)] = Array((16.523,[TDigestSQL(TDigest(0.5,0,130,TDigestMap(-4.9171836327285225 -> (1.0, 1.0), -3.9615949140987685 -> (1.0, 2.0), -3.792874086327091 -> (0.7500781537109753, 2.7500781537109753), -3.720534874164185 -> (1.796754196108008, 4.546832349818983), -3.702105588052377 -> (0.4531676501810167, 5.0), -3.665883591332569 -> (2.3434687534153142, 7.343468753415314), -3.649982231368131 -> (0.6565312465846858, 8.0), -3.5914188829817744 -> (4.0, 12.0), -3.530472305581248 -> (4.0, 16.0), -3.4060489584449467 -> (2.9372251939818383, 18.93722519398184), -3.3000694035428486 -> (8.12412890252889, 27.061354096510726), -3.2250016655261877 -> (8.30564453211017, 35.3669986286209), -3.180537395623448 -> (6.001782561137285, 41.3687811...
scala> bs.map(_._1)
res0: Array[Double] = Array(16.523, 17.138, 17.863, 17.801, 17.769, 17.786, 17.744, 17.8, 17.939, 17.854)
scala> val agg = TDigestAggregator(0.5, 0)
agg: org.apache.spark.tdigest.TDigestAggregator = TDigestAggregator(0.5,0)
scala> val udaa = spark.udf.register("tdigest", agg)
udaa: org.apache.spark.sql.expressions.UserDefinedAggregator[Double,org.apache.spark.tdigest.TDigestSQL,org.apache.spark.tdigest.TDigestSQL] = UserDefinedAggregator(TDigestAggregator(0.5,0),None,true,true)
scala> val bs = Benchmark.sample(10) { data.agg(udaa($"x1"), udaa($"x2")).first }
bs: Array[(Double, org.apache.spark.sql.Row)] = Array((0.313,[TDigestSQL(TDigest(0.5,0,130,TDigestMap(-4.9171836327285225 -> (1.0, 1.0), -3.9615949140987685 -> (1.0, 2.0), -3.792874086327091 -> (0.7500781537109753, 2.7500781537109753), -3.720534874164185 -> (1.796754196108008, 4.546832349818983), -3.702105588052377 -> (0.4531676501810167, 5.0), -3.665883591332569 -> (2.3434687534153142, 7.343468753415314), -3.649982231368131 -> (0.6565312465846858, 8.0), -3.5914188829817744 -> (4.0, 12.0), -3.530472305581248 -> (4.0, 16.0), -3.4060489584449467 -> (2.9372251939818383, 18.93722519398184), -3.3000694035428486 -> (8.12412890252889, 27.061354096510726), -3.2250016655261877 -> (8.30564453211017, 35.3669986286209), -3.180537395623448 -> (6.001782561137285, 41.36878118...
scala> bs.map(_._1)
res1: Array[Double] = Array(0.313, 0.193, 0.175, 0.185, 0.174, 0.176, 0.16, 0.186, 0.171, 0.179)
scala>
```
Closes #25024 from erikerlandson/spark-27296.
Authored-by: Erik Erlandson <eerlands@redhat.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
|
||
|---|---|---|
| .. | ||
| catalyst | ||
| core | ||
| hive | ||
| hive-thriftserver | ||
| create-docs.sh | ||
| gen-sql-markdown.py | ||
| mkdocs.yml | ||
| README.md | ||
Spark SQL
This module provides support for executing relational queries expressed in either SQL or the DataFrame/Dataset API.
Spark SQL is broken up into four subprojects:
- Catalyst (sql/catalyst) - An implementation-agnostic framework for manipulating trees of relational operators and expressions.
- Execution (sql/core) - A query planner / execution engine for translating Catalyst's logical query plans into Spark RDDs. This component also includes a new public interface, SQLContext, that allows users to execute SQL or LINQ statements against existing RDDs and Parquet files.
- Hive Support (sql/hive) - Includes extensions that allow users to write queries using a subset of HiveQL and access data from a Hive Metastore using Hive SerDes. There are also wrappers that allow users to run queries that include Hive UDFs, UDAFs, and UDTFs.
- HiveServer and CLI support (sql/hive-thriftserver) - Includes support for the SQL CLI (bin/spark-sql) and a HiveServer2 (for JDBC/ODBC) compatible server.
Running ./sql/create-docs.sh generates SQL documentation for built-in functions under sql/site.