[CALCITE-4204] Intermittent precision in Druid results when using aggregation functions over columns of type DOUBLE

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: druid-adapter

Environment

The results were obtained by using a dockerized Druid at 0.17.1: https://github.com/zabetak/calcite-druid-dataset 

• Assignee: Julian Hyde
• Reporter: Stamatis Zampetakis
• Created: 01/Sep/20 09:31
• Updated: 15/Oct/20 21:51

Description

Queries with aggregation functions on columns of type DOUBLE return results where the precision of the columns involved in the computation may differ from one execution to the other.

Consider the following query which can be found under DruidAdapterIT#testSingleAverageFunction.

select "store_state", sum("store_cost") / count(*) as a
from "foodmart" 
group by "store_state" 
order by a desc

The same query executed multiple times returns different results

Result 1

store_state=OR; A=2.6271402406293403
store_state=CA; A=2.599338206292706
store_state=WA; A=2.582870859286872

Result 2

store_state=OR; A=2.62714024062934
store_state=CA; A=2.599338206292706
store_state=WA; A=2.582870859286872

Result 3

store_state=OR; A=2.6271402406293394
store_state=CA; A=2.599338206292706
store_state=WA; A=2.582870859286872

Column "store_cost" in Druid is defined as shown below:

"metricsSpec" : [
...
{
  "name" : "store_sales",
  "type" : "doubleSum",
  "fieldName" : "store_sales"
}
...]

 

Issue Links

• is related to: CALCITE-4221 Update stale integration tests in Druid adapter