I have json files of the following structure:
{"names":[{"name":"John","lastName":"Doe"},
{"name":"John","lastName":"Marcus"},
{"name":"David","lastName":"Luis"}
]}
I want to read several such json files and distinct them based on the "name" column inside names. I tried
df.dropDuplicates(Array("names.name"))
but it didn't do the magic.
This seems to be a regression that was added in spark 2.0. If you bring the nested column to the highest level you can drop the duplicates. If we create a new column based on the columns you want to dedup on. Then we drop the columns and finally drop the column. The following function will work for composite keys as well.
val columns = Seq("names.name")
df.withColumn("DEDUP_KEY", concat_ws(",", columns:_*))
.dropDuplicates("DEDUP_KEY")
.drop("DEDUP_KEY")
just for future reference, the solution looks like
val uniqueNams = allNames.withColumn("DEDUP_NAME_KEY",
org.apache.spark.sql.functions.explode(new Column("names.name")))
.cache()
.dropDuplicates(Array("DEDUP_NAME_KEY"))
.drop("DEDUP_NAME_KEY")
As an update to existing answer, similar thing can be achieved without explode. We can simply get value of each column and then do the concatenation for generating DEDUPE_KEY
val columns = Seq("names.name")
df.withColumn("DEDUPE_KEY", concat_ws("_", columns.map(att => col(att)):_*))
.dropDuplicates("DEDUPE_KEY")
.drop("DEDUPE_KEY")
