[
https://issues.apache.org/jira/browse/SPARK-39168?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Brian Schaefer updated SPARK-39168:
-----------------------------------
Description:
Schema inference fails on the following case:
{code:python}
>>> data = [{"a": [1, None], "b": [None, 2]}]
>>> spark.createDataFrame(data)
ValueError: Some of types cannot be determined after inferring
{code}
This is because only the first value in the array is used to infer the element
type for the array:
[https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/types.py#L1260].
The element type of the "b" array is inferred as {{NullType}} but I think it
makes sense to infer the element type as {{{}LongType{}}}.
One approach to address the above would be to infer the type from the first
non-null value in the array. However, consider a case with structs:
{code:python}
>>> spark.conf.set("spark.sql.pyspark.inferNestedDictAsStruct.enabled", True)
>>> data = [{"a": [{"b": 1}, {"c": 2}]}]
>>> spark.createDataFrame(data).schema
StructType([StructField('a', ArrayType(StructType([StructField('b', LongType(),
True)]), True), True)])
{code}
The element type of the "a" array is inferred as a struct with one field, "b".
However, it would be convenient to infer the element type as a struct with both
fields "b" and "c". Omitted fields from each dictionary would become null
values in each struct:
{code:java}
+----------------------+
| a|
+----------------------+
|[{1, null}, {null, 2}]|
+----------------------+
{code}
To support both of these cases, the type of each array element could be
inferred, and those types could be merged, similar to the approach
[here|https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/session.py#L574-L576].
was:
Schema inference fails on the following case:
{code:python}
>>> data = [{"a": [1, None], "b": [None, 2]}]
>>> spark.createDataFrame(data)
ValueError: Some of types cannot be determined after inferring
{code}
This is because only the first value in the array is used to infer the element
type for the array:
[https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/types.py#L1260].
The element type of the "b" array is inferred as {{NullType}} but I think it
makes sense to infer the element type as {{{}LongType{}}}.
One approach to address the above would be to infer the type from the first
non-null value in the array. However, consider a case with structs:
{code:python}
>>> spark.conf.set("spark.sql.pyspark.inferNestedDictAsStruct.enabled", True)
>>> data = [{"a": [{"b": 1}, {"c": 2}]}]
>>> spark.createDataFrame(data).schema
StructType([StructField('a', ArrayType(StructType([StructField('b', LongType(),
True)]), True), True)])
{code}
The element type of the "a" array is inferred as a struct with one field, "b".
However, it would be convenient to infer the element type as a struct with both
fields "b" and "c". Omitted fields from each dictionary would become null
values in each struct:
{code:java}
+----------------------+
| a|
+----------------------+
|[{1, null}, {null, 1}]|
+----------------------+
{code}
To support both of these cases, the type of each array element could be
inferred, and those types could be merged, similar to the approach
[here|https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/session.py#L574-L576].
> Consider all values in a python list when inferring schema
> ----------------------------------------------------------
>
> Key: SPARK-39168
> URL: https://issues.apache.org/jira/browse/SPARK-39168
> Project: Spark
> Issue Type: New Feature
> Components: PySpark
> Affects Versions: 3.2.1
> Reporter: Brian Schaefer
> Priority: Major
>
> Schema inference fails on the following case:
> {code:python}
> >>> data = [{"a": [1, None], "b": [None, 2]}]
> >>> spark.createDataFrame(data)
> ValueError: Some of types cannot be determined after inferring
> {code}
> This is because only the first value in the array is used to infer the
> element type for the array:
> [https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/types.py#L1260].
> The element type of the "b" array is inferred as {{NullType}} but I think it
> makes sense to infer the element type as {{{}LongType{}}}.
> One approach to address the above would be to infer the type from the first
> non-null value in the array. However, consider a case with structs:
> {code:python}
> >>> spark.conf.set("spark.sql.pyspark.inferNestedDictAsStruct.enabled", True)
> >>> data = [{"a": [{"b": 1}, {"c": 2}]}]
> >>> spark.createDataFrame(data).schema
> StructType([StructField('a', ArrayType(StructType([StructField('b',
> LongType(), True)]), True), True)])
> {code}
> The element type of the "a" array is inferred as a struct with one field,
> "b". However, it would be convenient to infer the element type as a struct
> with both fields "b" and "c". Omitted fields from each dictionary would
> become null values in each struct:
> {code:java}
> +----------------------+
> | a|
> +----------------------+
> |[{1, null}, {null, 2}]|
> +----------------------+
> {code}
> To support both of these cases, the type of each array element could be
> inferred, and those types could be merged, similar to the approach
> [here|https://github.com/apache/spark/blob/b63674ea5f746306a96ab8c39c23a230a6cb9566/python/pyspark/sql/session.py#L574-L576].
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
