adding a new column using withColumn from a lookup table dynamically - apache-spark

Question

I am using spark-sql-2.4.1v with Java 8. I have a scenario where I need to dynamically add a column from a look up table.
I have data frame with columns
A, B, C , ..., X,Y, Z
When few (original) columns ( Ex: A,B,C) values are null , i need to take/substitute column( Ex: X,Y,Z) values else take the original column values.
I will get this mapping information as part of business logic.
If that is the case i will follow something like below hard-coded code
Dataset<Row> substitutedDs = ds
.withColumn("A",
when(col("A").isNull() , col("X").cast(DataTypes.StringType))
.otherwise(col("A").cast(DataTypes.StringType))
)
.withColumn("C",
when(col("C").isNull() , col("Z").cast(DataTypes.StringType))
.otherwise(col("C").cast(DataTypes.StringType))
Which is working fine. But I need to do this dynamically/configurable to avoid hard-coding.
I will get look up table with columns "code" and "code_substitutes" information as below
-------------------------
| Code | Code_Substitute |
-------------------------
A X
B Y
C Z
-------------------------
I need to dynamically construct above "substitutedDs" , how can this be done ?

Answer 1

With Java8, you can use this Stream.reduce() overload:
final Dataset<Row> dataframe = ...;
final Map<String, String> substitutes = ...;
final Dataset<Row> afterSubstitutions = codeSubstitutes.entrySet().stream()
.reduce(dataframe, (df, entry) ->
df.withColumn(entry.getKey(), when(/* replace with col(entry.getValue()) when null */)),
(left, right) -> { throw new IllegalStateException("Can't merge two dataframes. This stream should not be a parallel one!"); }
);
The combiner (last argument) is supposed to merge two dataframes processed in parallel (if the stream was a parallel() stream), but we'll simply not allow that, as we're only invoking this logic on a sequential() stream.
A more readable/maintainable version involves an extra-step for extracting the above logic into dedicated methods, such as:
// ...
Dataset<Row> nullSafeDf = codeSubstitutes.entrySet().stream()
.reduce(dataframe, this::replaceIfNull, this::throwingCombiner);
// ...
}
private Dataset<Row> replaceIfNull(Dataset<Row> df, Map.Entry<String, String> substitution) {
final String original = substitution.getKey();
final String replacement = substitution.getValue();
return df.withColumn(original, when(col(original).isNull(), col(replacement))
.otherwise(col(original)));
}
private <X> X throwingCombiner(X left, X right) {
throw new IllegalStateException("Combining not allowed");
}

Answer 2

In Scala, I would do like this
val substitueMapping: Map[String, String] = ??? //this is your substitute map, this is small as it contains columns and their null substitutes
val df = ??? //this is your main dataframe
val substitutedDf = substituteMapping.keys().foldLeft(df)((df, k) => {
df.withColumn(k, when(col(k).isNull, col(substituteMapping(k))).otherwise(col(k)))
//do approproate casting in above which you have done in post
})
I think foldLeft is not there in Java 8, you can emulate the same by modifying a variable repeatedly and doing iteration on substituteMapping.