Spark / Scala: forward fill with last observation
It is possible to do it only using Window function (without last function) and somehow clever partitionning. I personally really dislike having to use the combination of groupBy then further join.
So given:
date, currency, rate
20190101 JPY NULL
20190102 JPY 2
20190103 JPY NULL
20190104 JPY NULL
20190102 JPY 3
20190103 JPY 4
20190104 JPY NULL
We can use Window.unboundedPreceding and Window.unboundedFollowing to create a key for forward and backward fill.
The following code:
val w1 = Window.partitionBy("currency").orderBy(asc("date"))
df
.select("date", "currency", "rate")
// Equivalent of fill.na(0, Seq("rate")) but can be more generic here
// You may need an abs(col("rate")) if value col can be negative since it will not work with the following sums to build the forward and backward keys
.withColumn("rate_filled", when(col("rate").isNull, lit(0)).otherwise(col("rate")))
.withColumn("rate_backsum",
sum("rate_filled").over(w1.rowsBetween(Window.unboundedPreceding, Window.currentRow)))
.withColumn("rate_forwardsum",
sum("rate_filled").over(w1.rowsBetween(Window.currentRow, Window.unboundedFollowing)))
gives :
date, currency, rate, rate_filled, rate_backsum, rate_forwardsum
20190101 JPY NULL 0 0 9
20190102 JPY 2 2 2 9
20190103 JPY NULL 0 2 7
20190104 JPY NULL 0 2 7
20190102 JPY 3 3 5 7
20190103 JPY 4 4 9 4
20190104 JPY NULL 0 9 0
Therefore, we've built two keys (x_backsum and x_forwardsum) that can be used to ffill and bfill. With the two following spark lines :
val wb = Window.partitionBy("currency", "rate_backsum")
val wf = Window.partitionBy("currency", "rate_forwardsum")
...
.withColumn("rate_backfilled", avg("rate").over(wb))
.withColumn("rate_forwardfilled", avg("rate").over(wf))
Finally :
date, currency, rate, rate_backsum, rate_forwardsum, rate_ffilled
20190101 JPY NULL 0 9 2
20190102 JPY 2 2 9 2
20190103 JPY NULL 2 7 3
20190104 JPY NULL 2 7 3
20190102 JPY 3 5 7 3
20190103 JPY 4 9 4 4
20190104 JPY NULL 9 0 0
Initial answer (a single time series assumption):
First of all try avoid window functions if you cannot provide PARTITION BY
clause. It moves data to a single partition so most of the time it is simply not feasible.
What you can do is to fill gaps on RDD
using mapPartitionsWithIndex
. Since you didn't provide an example data or expected output consider this to be pseudocode not a real Scala program:
first lets order
DataFrame
by date and convert toRDD
import org.apache.spark.sql.Row import org.apache.spark.rdd.RDD val rows: RDD[Row] = df.orderBy($"Date").rdd
next lets find the last not null observation per partition
def notMissing(row: Row): Boolean = ??? val toCarry: scala.collection.Map[Int,Option[org.apache.spark.sql.Row]] = rows .mapPartitionsWithIndex{ case (i, iter) => Iterator((i, iter.filter(notMissing(_)).toSeq.lastOption)) } .collectAsMap
and convert this
Map
to broadcastval toCarryBd = sc.broadcast(toCarry)
finally map over partitions once again filling the gaps:
def fill(i: Int, iter: Iterator[Row]): Iterator[Row] = { // If it is the beginning of partition and value is missing // extract value to fill from toCarryBd.value // Remember to correct for empty / only missing partitions // otherwise take last not-null from the current partition } val imputed: RDD[Row] = rows .mapPartitionsWithIndex{ case (i, iter) => fill(i, iter) }
finally convert back to DataFrame
Edit (partitioned / time series per group data):
The devil is in the detail. If your data is partitioned after all then a whole problem can be solved using groupBy
. Lets assume you simply partition by column "v" of type T
and Date
is an integer timestamp:
def fill(iter: List[Row]): List[Row] = {
// Just go row by row and fill with last non-empty value
???
}
val groupedAndSorted = df.rdd
.groupBy(_.getAs[T]("k"))
.mapValues(_.toList.sortBy(_.getAs[Int]("Date")))
val rows: RDD[Row] = groupedAndSorted.mapValues(fill).values.flatMap(identity)
val dfFilled = sqlContext.createDataFrame(rows, df.schema)
This way you can fill all columns at the same time.
Can this be done with DataFrames instead of converting back and forth to RDD?
It depends, although it is unlikely to be efficient. If maximum gap is relatively small you can do something like this:
import org.apache.spark.sql.functions._
import org.apache.spark.sql.expressions.{WindowSpec, Window}
import org.apache.spark.sql.Column
val maxGap: Int = ??? // Maximum gap between observations
val columnsToFill: List[String] = ??? // List of columns to fill
val suffix: String = "_" // To disambiguate between original and imputed
// Take lag 1 to maxGap and coalesce
def makeCoalesce(w: WindowSpec)(magGap: Int)(suffix: String)(c: String) = {
// Generate lag values between 1 and maxGap
val lags = (1 to maxGap).map(lag(col(c), _)over(w))
// Add current, coalesce and set alias
coalesce(col(c) +: lags: _*).alias(s"$c$suffix")
}
// For each column you want to fill nulls apply makeCoalesce
val lags: List[Column] = columnsToFill.map(makeCoalesce(w)(maxGap)("_"))
// Finally select
val dfImputed = df.select($"*" :: lags: _*)
It can be easily adjusted to use different maximum gap per column.
A simpler way to achieve a similar result in the latest Spark version is to use last
with ignoreNulls
:
import org.apache.spark.sql.functions._
import org.apache.spark.sql.expressions.Window
val w = Window.partitionBy($"k").orderBy($"Date")
.rowsBetween(Window.unboundedPreceding, -1)
df.withColumn("value", coalesce($"value", last($"value", true).over(w)))
While it is possible to drop partitionBy
clause and apply this method globally, it would prohibitively expensive with large datasets.