I am using the spark 2.0.1 and want to fill nan values with the last good known value in the column.

The only reference for spark I could find Spark / Scala: forward fill with last observation or Fill in null with previously known good value with pyspark which seem to use RDD.

I would rather like to stay in the data frame / dataset world and possible handle multiple nan values. Is this possible?

My assumption is that the data (initially loaded from e.g. a CSV file is ordered by time and this order is preserved in the distributed setting e.g. filling by close / last good known value is correct. Maybe filling with the previous value is enough as for most records there are no 2 or more nan records in a row. Does this actually hold? The point is that a

myDf.sort("foo").show

Would destroy any order e.g. all null values will come first.

A small example:

import java.sql.{ Date, Timestamp }
case class FooBar(foo:Date, bar:String)
val myDf = Seq(("2016-01-01","first"),("2016-01-02","second"),("2016-wrongFormat","noValidFormat"), ("2016-01-04","lastAssumingSameDate"))
         .toDF("foo","bar")
         .withColumn("foo", 'foo.cast("Date"))
         .as[FooBar]

Results in

+----------+--------------------+
|       foo|                 bar|
+----------+--------------------+
|2016-01-01|               first|
|2016-01-02|              second|
|      null|       noValidFormat|
|2016-01-04|lastAssumingSameDate|
+----------+--------------------+

I would like to fix the value with the last good known value. How can I achieve this?

+----------+--------------------+
|       foo|                 bar|
+----------+--------------------+
|2016-01-01|               first|
|2016-01-02|              second|
|2016-01-02|       noValidFormat|
|2016-01-04|lastAssumingSameDate|
+----------+--------------------+

edit

in my case, it would be good enough to fill the value from the row above, as there are only very limited faulty values.

edit2

I try to add an index column

val myDf = Seq(("2016-01-01", "first"), ("2016-01-02", "second"), ("2016-wrongFormat", "noValidFormat"), ("2016-01-04", "lastAssumingSameDate"))
    .toDF("foo", "bar")
    .withColumn("foo", 'foo.cast("Date"))
    .as[FooBar]
    .withColumn("rowId", monotonically_increasing_id())

And then fill with the last value.

myDf.withColumn("fooLag", lag('foo, 1) over Window.orderBy('rowId)).show

But that reads the following warning: No Partition Defined for Window operation! Moving all data to a single partition, this can cause serious performance degradation. How could I introduce meaningful partitions?

+----------+--------------------+-----+----------+
|       foo|                 bar|rowId|    fooLag|
+----------+--------------------+-----+----------+
|2016-01-01|               first|    0|      null|
|2016-01-02|              second|    1|2016-01-01|
|      null|       noValidFormat|    2|2016-01-02|
|2016-01-04|lastAssumingSameDate|    3|      null|
+----------+--------------------+-----+----------+

This is an intermediate answer. However, it is not great as no partitions / only a single partition is used. I am still looking for a better way to solve the problem

df
    .withColumn("rowId", monotonically_increasing_id())
    .withColumn("replacement", lag('columnWithNull, 1) over Window.orderBy('rowId))
    .withColumn("columnWithNullReplaced",
      when($"columnWithNull" isNull, "replacement").otherwise($"columnWithNull")

    )

edit

I am working on building a better solution using mapPartitionsWithIndex https://gist.github.com/geoHeil/6a23d18ccec085d486165089f9f430f2 is not complete yet.

edit2

adding

if (i == 0) {
          lastNotNullRow = toCarryBd.value.get(i + 1).get
        } else {
          lastNotNullRow = toCarryBd.value.get(i - 1).get
        }

will lead to the desired result.