Spark find NULL value blocks in Series of values - apache-spark

Question

Assume this is my data:
date value
2016-01-01 1
2016-01-02 NULL
2016-01-03 NULL
2016-01-04 2
2016-01-05 3
2016-01-06 NULL
2016-01-07 NULL
2016-01-08 NULL
2016-01-09 1
I am trying to find the start and end dates that surround the NULL-value groups. An example output would be as follows:
start end
2016-01-01 2016-01-04
2016-01-05 2016-01-09
My first attempt at the problem produced the following:
df.filter($"value".isNull)\
.agg(to_date(date_add(max("date"), 1)) as "max",
to_date(date_sub(min("date"),1)) as "min"
)
but this only finds the total min and max value. I thought of using groupBy but don't know how to create a column for each of the null value blocks.

Answer 1

The tricky part is to get the borders of the groups, therefore you need several steps.
first to build groups of nulls/not-nulls (using window-functions)
then group by blocks to get the borders within the blocks
then again window-function to extend the borders
Here a working example:
import ss.implicits._
val df = Seq(
("2016-01-01", Some(1)),
("2016-01-02", None),
("2016-01-03", None),
("2016-01-04", Some(2)),
("2016-01-05", Some(3)),
("2016-01-06", None),
("2016-01-07", None),
("2016-01-08", None),
("2016-01-09", Some(1))
).toDF("date", "value")
df
// build blocks
.withColumn("isnull", when($"value".isNull, true).otherwise(false))
.withColumn("lag_isnull", lag($"isnull",1).over(Window.orderBy($"date")))
.withColumn("change", coalesce($"isnull"=!=$"lag_isnull",lit(false)))
.withColumn("block", sum($"change".cast("int")).over(Window.orderBy($"date")))
// now calculate min/max within groups
.groupBy($"block")
.agg(
min($"date").as("tmp_min"),
max($"date").as("tmp_max"),
(count($"value")===0).as("null_block")
)
// now extend groups to include borders
.withColumn("min", lag($"tmp_max", 1).over(Window.orderBy($"tmp_min")))
.withColumn("max", lead($"tmp_min", 1).over(Window.orderBy($"tmp_max")))
// only select null-groups
.where($"null_block")
.select($"min", $"max")
.orderBy($"min")
.show()
gives
+----------+----------+
| min| max|
+----------+----------+
|2016-01-01|2016-01-04|
|2016-01-05|2016-01-09|
+----------+----------+

Answer 2

I don't have a working solution but I do have a few recommendations.
Look at using a lag; you will also have to change that code a bit to produce a lead column.
Now assume you have your lag and lead column. Your resultant dataframe will now look like this:
date value lag_value lead_value
2016-01-01 1 NULL 1
2016-01-02 NULL NULL 1
2016-01-03 NULL 2 NULL
2016-01-04 2 3 NULL
2016-01-05 3 NULL 2
2016-01-06 NULL NULL 3
2016-01-07 NULL NULL NULL
2016-01-08 NULL 1 NULL
2016-01-09 1 1 NULL
Now what you want to do is just filter by the following conditions:
min date:
df.filter("value IS NOT NULL AND lag_value IS NULL")
max date:
df.filter("value IS NULL AND lead_value IS NOT NULL")
If you want to be a bit more advanced, you can also use a when command to create a new column which states if the date is a start or end date for a null group:
date value lag_value lead_value group_date_type
2016-01-01 1 NULL 1 start
2016-01-02 NULL NULL 1 NULL
2016-01-03 NULL 2 NULL NULL
2016-01-04 2 3 NULL end
2016-01-05 3 NULL 2 start
2016-01-06 NULL NULL 3 NULL
2016-01-07 NULL NULL NULL NULL
2016-01-08 NULL 1 NULL NULL
2016-01-09 1 1 NULL end
This can be created with something that looks like this:
from pyspark.sql import functions as F
df_2.withColumn('group_date_type',
F.when("value IS NOT NULL AND lag_value IS NULL", start)\
.when("value IS NULL AND lead_value IS NOT NULL", end)\
.otherwise(None)
)