[bug] CAST after expression causes weird exception

Bug report

select 
  playerID, 
  cast(hits as long) AS totalHits
from baseballStats 
order by totalHits DESC

or

select 
  playerID, 
  count(cast(hits as long)) AS totalHits
from baseballStats 
group by playerID
order by totalHits DESC

works fine.

but:

select 
  playerID, 
  cast(count(hits) as long) AS totalHits
from baseballStats 
group by playerID
order by totalHits DESC

gives

[
  {
    "message": "QueryExecutionError:
ProcessingException(errorCode:450, message:InternalError:
java.lang.NullPointerException
    at org.apache.pinot.core.operator.combine.GroupByOrderByCombineOperator.mergeResults(GroupByOrderByCombineOperator.java:236).
    at 
org.apache.pinot.core.operator.combine.BaseCombineOperator.getNextBlock(BaseCombineOperator.java:119)
    at org.apache.pinot.core.operator.combine.BaseCombineOperator.getNextBlock(BaseCombineOperator.java:50)",
    "errorCode": 200
  }
]

what’s more

select 
  playerID, 
  cast(count(hits) as long) AS totalHits
from baseballStats 
group by playerID

without order by gives a non-descriptive exception

ProcessingException(errorCode:450, message:InternalError:
java.io.IOException: Failed : HTTP error code : 500
	at org.apache.pinot.controller.api.resources.PinotQueryResource.sendPostRaw(PinotQueryResource.java:305)
	at org.apache.pinot.controller.api.resources.PinotQueryResource.sendRequestRaw(PinotQueryResource.java:343)
	at org.apache.pinot.controller.api.resources.PinotQueryResource.getQueryResponse(PinotQueryResource.java:225)
	at org.apache.pinot.controller.api.resources.PinotQueryResource.handlePostSql(PinotQueryResource.java:137))

[Jackie-Jiang]

Scalar function is used in the following places:

• Ingestion transform
• Compile time evaluation of literal only function
• Post aggregation transform
• If transform function is not implemented for the same name, scalar function can be used as transform function (poor performance)

For post aggregation transform and transform, if the result class for the scalar function is dynamic, we will treat it as STRING. In order to solve this problem, we need to add the support to derive the result type from the arguments’ type.

[walterddr]

I doubt the cause is the absence of a scalar function then. The motivation of that PR is to ensure casts in filters don’t prevent index use.

for the purpose of index use i think the PR works well. it doesn’t solve what’s describe in this issue though