spark-instrumented-optimizer/sql/catalyst
Hyukjin Kwon ab850c02f7 [SPARK-26901][SQL][R] Adds child's output into references to avoid column-pruning for vectorized gapply()
## What changes were proposed in this pull request?

Currently, looks column pruning is done to vectorized `gapply()`. Given R native function could use all referred fields so it shouldn't be pruned. To avoid this, it adds child's output into `references` like `OutputConsumer`.

```
$ ./bin/sparkR --conf spark.sql.execution.arrow.enabled=true
```

```r
df <- createDataFrame(mtcars)
explain(count(groupBy(gapply(df,
                             "gear",
                             function(key, group) {
                               data.frame(gear = key[[1]], disp = mean(group$disp))
                             },
                             structType("gear double, disp double")))), TRUE)
```

**Before:**

```
== Optimized Logical Plan ==
Aggregate [count(1) AS count#41L]
+- Project
   +- FlatMapGroupsInRWithArrow [...]
      +- Project [gear#9]
         +- LogicalRDD [mpg#0, cyl#1, disp#2, hp#3, drat#4, wt#5, qsec#6, vs#7, am#8, gear#9, carb#10], false

== Physical Plan ==
*(4) HashAggregate(keys=[], functions=[count(1)], output=[count#41L])
+- Exchange SinglePartition
   +- *(3) HashAggregate(keys=[], functions=[partial_count(1)], output=[count#44L])
      +- *(3) Project
         +- FlatMapGroupsInRWithArrow [...]
            +- *(2) Sort [gear#9 ASC NULLS FIRST], false, 0
               +- Exchange hashpartitioning(gear#9, 200)
                  +- *(1) Project [gear#9]
                     +- *(1) Scan ExistingRDD arrow[mpg#0,cyl#1,disp#2,hp#3,drat#4,wt#5,qsec#6,vs#7,am#8,gear#9,carb#10]
```

**After:**

```
== Optimized Logical Plan ==
Aggregate [count(1) AS count#91L]
+- Project
   +- FlatMapGroupsInRWithArrow [...]
      +- LogicalRDD [mpg#0, cyl#1, disp#2, hp#3, drat#4, wt#5, qsec#6, vs#7, am#8, gear#9, carb#10], false

== Physical Plan ==
*(4) HashAggregate(keys=[], functions=[count(1)], output=[count#91L])
+- Exchange SinglePartition
   +- *(3) HashAggregate(keys=[], functions=[partial_count(1)], output=[count#94L])
      +- *(3) Project
         +- FlatMapGroupsInRWithArrow [...]
            +- *(2) Sort [gear#9 ASC NULLS FIRST], false, 0
               +- Exchange hashpartitioning(gear#9, 200)
                  +- *(1) Scan ExistingRDD arrow[mpg#0,cyl#1,disp#2,hp#3,drat#4,wt#5,qsec#6,vs#7,am#8,gear#9,carb#10]
```

Currently, it adds corrupt values for missing columns (via pruned columnar batches to Arrow writers that requires non-pruned columns) such as:

```r
...
  c(7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 7.90505033345994e-323, 0, 0, 4.17777978645388e-314)
  c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1.04669129845114e+219)
  c(3.4482690635875e-313, 3.4482690635875e-313, 3.4482690635875e-313,
  c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2.47032822920623e-323)
...
```

which should be something like:

```r
...
  c(4, 4, 1, 2, 2, 4, 4, 1, 2, 1, 1, 2)
  c(26, 30.4, 15.8, 19.7, 15)
  c(4, 4, 8, 6, 8)
  c(120.3, 95.1, 351, 145, 301)
...
```

## How was this patch tested?

Manually tested, and unit tests were added.

The test code is basiaclly:

```r
df <- createDataFrame(mtcars)
count(gapply(df,
             c("gear"),
             function(key, group) {
                stopifnot(all(group$hp > 50))
                group
             },
             schema(df)))
```

`mtcars`'s hp is all more then 50.

```r
> mtcars$hp > 50
 [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[16] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[31] TRUE TRUE
```

However, due to corrpt value, (like 0 or 7.xxxxx), werid values were found. So, it's currently being failed as below in the master:

```
Error in handleErrors(returnStatus, conn) :
  org.apache.spark.SparkException: Job aborted due to stage failure: Task 82 in stage 1.0 failed 1 times, most recent failure: Lost task 82.0 in stage 1.0 (TID 198, localhost, executor driver): org.apache.spark.SparkException: R worker exited unexpectedly (crashed)
 Error in computeFunc(key, inputData) : all(group$hp > 50) is not TRUE
Error in computeFunc(key, inputData) : all(group$hp > 50) is not TRUE
Error in computeFunc(key, inputData) : all(group$hp > 50) is not TRUE
```

I also compared the total length while I am here. Regular `gapply` without Arrow has some holes .. so I had to compare the results with R data frame.

Closes #23810 from HyukjinKwon/SPARK-26901.

Lead-authored-by: Hyukjin Kwon <gurwls223@apache.org>
Co-authored-by: Hyukjin Kwon <gurwls223@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
2019-02-20 10:24:40 +08:00
..
benchmarks [SPARK-25657][SQL][TEST] Refactor HashBenchmark to use main method 2018-10-07 09:49:37 -07:00
src [SPARK-26901][SQL][R] Adds child's output into references to avoid column-pruning for vectorized gapply() 2019-02-20 10:24:40 +08:00
pom.xml [SPARK-25956] Make Scala 2.12 as default Scala version in Spark 3.0 2018-11-14 16:22:23 -08:00