[SPARK-47563][SQL] Add map normalization on creation #45721
Conversation
| case FloatType => NormalizeFloatingNumbers.FLOAT_NORMALIZER(value) | ||
| case DoubleType => NormalizeFloatingNumbers.DOUBLE_NORMALIZER(value) | ||
| case ArrayType(dt, _) => | ||
| new GenericArrayData(value.asInstanceOf[GenericArrayData].array.map { element => |
There was a problem hiding this comment.
if we have an array of 1 million strings we will go through each value even though we know we don't need to normalize strings
what about doing the same as in NormalizeFloatingNumbers and first check if we need to perform normalization
There was a problem hiding this comment.
Applied NormalizeFloatingNumbers.needNormalize here
| "mapKeyDedupPolicy" -> "\"spark.sql.mapKeyDedupPolicy\"") | ||
| ) | ||
|
|
||
| val builderStruct = new ArrayBasedMapBuilder(new StructType().add("i", "double"), IntegerType) |
There was a problem hiding this comment.
maybe add a case when array is inside of a struct
| ) | ||
| } | ||
|
|
||
| test("apply key normalization when creating") { |
There was a problem hiding this comment.
add another test for successful normalization
|
please add info in the description on why the change is needed, ie right now we can create a map with both keys -0 and 0 etc |
| case StructType(sf) => | ||
| new GenericInternalRow( | ||
| value.asInstanceOf[GenericInternalRow].values.zipWithIndex.map { element => | ||
| normalize(element._1, sf(element._2).dataType) |
There was a problem hiding this comment.
you could also check if you need to do normalization here right?
this way we would avoid normalization of all fields of a struct if only one actually needs it
There was a problem hiding this comment.
As noted by @cloud-fan below, complex types have been dropped.
|
|
||
| private lazy val keyNeedNormalize = NormalizeFloatingNumbers.needNormalize(keyType) | ||
|
|
||
| def normalize(value: Any, dataType: DataType): Any = dataType match { |
There was a problem hiding this comment.
we should return a lambda function to do normalization based on the data type, instead of matching the data type per row.
| def normalize(value: Any, dataType: DataType): Any = dataType match { | ||
| case FloatType => NormalizeFloatingNumbers.FLOAT_NORMALIZER(value) | ||
| case DoubleType => NormalizeFloatingNumbers.DOUBLE_NORMALIZER(value) | ||
| case ArrayType(dt, _) => |
There was a problem hiding this comment.
no need to handle complex types, as we use TreeMap for complex types which should handle floating points well.
| private lazy val keyNeedNormalize = | ||
| keyType.isInstanceOf[FloatType] || keyType.isInstanceOf[DoubleType] | ||
|
|
||
| def normalize(dataType: DataType): Any => Any = dataType match { |
There was a problem hiding this comment.
private lazy val keyNormalizer: Any => Any = keyType match {
case FloatType => NormalizeFloatingNumbers.FLOAT_NORMALIZER
case DoubleType => NormalizeFloatingNumbers.DOUBLE_NORMALIZER
case _ => identity
}
then the can just write
val keyNormalized = keyNormalizer(key)
|
thanks, merging to master! |
3 participants
What changes were proposed in this pull request?
Added normalization of map keys when they are put in
ArrayBasedMapBuilder.Why are the changes needed?
As map keys need to be unique, we need to add normalization on floating point numbers and prevent the following case when building a map:
Map(0.0, -0.0).This further unblocks GROUP BY statement for Map Types as per this discussion.
Does this PR introduce any user-facing change?
No
How was this patch tested?
New UTs in
ArrayBasedMapBuilderSuiteWas this patch authored or co-authored using generative AI tooling?
No