对于循环Spark数据帧

Question

I have a Dataframe df that has, among others, a column of groupID ; that is, each observation belongs to a specific group. In total there are 8 groups. I would like to sample from each groupID a certain percent of observations (say, 20%). Here is my approach of doing this:

val sample_df = for ( i <- Array.range(0,7) ) yield {  
             val sel_df = df.filter($"groupID"===i)  
             sel_df.sample(false,0.2,seed1)  
             }  

The result of this code is:

Array[org.apache.spark.sql.DataFrame] = Array([text: string, groupID: int], [text: string, groupID: int])

I applied flatMap() on sample_df , but I got an error:

val flat_df = sample_df.flatMap(x => x)
         <console>:59: error: type mismatch;
         found: org.apache.spark.sql.DataFrame
         required: scala.collection.GenTraversableOnce[?]

How can I get a sampled dataframe?

Answer 1

As far as I understood, you are trying to get RDD of Row . For that you can simply call:

val rows: RDD[Row] = sample_df.rdd

To explain the error you get better, flatMap requires something traversable like Option but you supplied just a Row .

Also, to get all data to the driver, you can call:

val rows: Array[Row] = sample_df.collect

Answer 2

我猜你想要在每组上均匀采样。

sample_df.reduceLeft((result, df) => result.unionAll(df))

Answer 3

It seems to me you just want to take a 20% sample of the entire dataframe? If so, then there is no reason to create 8 different dataframes and then union them back.

df.sample(false, 0.2, seed)

will do the trick. If you want to do different fractions for each groupID then check out df.stat.sampleBy . If you want to be sure that there is exactly 20% of each class in the sample then you'll have to convert to a PairRDD and use stratified sampling like:

df.rdd.map(row => (row(groupIDIndex), row)).sampleByKeyExact(false, Map(0 -> 0.2, 1 -> 0.2, ..., 8 -> 0.2), seed)