Volodymyr Tkach created DRILL-6154:

             Summary: NaN, Infinity issues
                 Key: DRILL-6154
                 URL: https://issues.apache.org/jira/browse/DRILL-6154
             Project: Apache Drill
          Issue Type: Bug
            Reporter: Volodymyr Tkach
            Assignee: Volodymyr Tkach



- *sqrt*
- *trunc*

*ISSUE_DESCRIPTION:* According to DRILL-5919/MD-2745, new json number literals 
were added: *NaN, Infinity, -Infinity*. The new data types must be processed 
properly by existing functions. There are a few issues:
*1. SQRT function*. Run the following test query: {code}select sqrt(Nan) NaN, 
sqrt(Positive_Infinity) POS_INF, sqrt(Negative_Infinity) NEG_INF from 
- EXPECTED_RESULT: it was expected to get the following result: _NaN, Infinity, 
NaN_ (expected result is based on java Math.sqrt() method)
- ACTUAL_RESULT: the test query returned: _NaN, Infinity, Infinity_

*2. TRUNC function*. According to DRILL docs 
(https://drill.apache.org/docs/math-and-trig/): _TRUNC(x, y) : Truncates x to y 
decimal places. *Specifying y is optional. Default is 1*_. So, the function 
must work properly without specifying *y*
However an error message appears. Run test_query: {code}select trunc(Nan) NaN, 
trunc(Positive_Infinity) POS_INF, trunc(Negative_Infinity) NEG_INF from 
- EXPECTED_RESULT: it was expected to get the following result *NaN, NaN, NaN*
- ACTUAL_RESULT: it appears the following error message: *Query Failed: An 
Error Occurred org.apache.drill.common.exceptions.UserRemoteException: SYSTEM 
ERROR: NumberFormatException Fragment 0:0 [Error Id: 
95e01fee-7433-4b0b-b913-32358b4a8f55 on node1:31010]*

Please investigate and fix, test file attached PN_Inf_NaN.json

2. Issue 


*ISSUE_DESCRIPTION:* There were added new Json data types in 
MD-2745/DRILL-5919: *NaN, Infinity, -Infinity*. 
During testing activities, it was detected a bit strange behavior of INNER JOIN 
operator - different query results in almost the same queries. 
*Query1* {code} select distinct t.name, tt.name from dfs.tmp.`ObjsX.json` t 
inner join dfs.tmp.`ObjsX.json` tt on t.attr4 = tt.attr4 {code}
*Query2* {code} select distinct t.name from dfs.tmp.`ObjsX.json` t inner join 
dfs.tmp.`ObjsX.json` tt on t.attr4 = tt.attr4 {code}

*Query1* differs from *Query2* by 1 columns only:
- In *Query1* - 2 columns are selected - t.name, tt.name
- In *Query2* - 1 column is selected - t.name

However *Query1*/*Query2* return completely different results:
- *Query1* returns
        name         name0
        object2         object2
        object2         object3
        object2         object4
        object3         object2
        object3         object3
        object3         object4
        object4         object2
        object4         object3
        object4         object4
This result seems to be correct.

- *Query2* returns _*No result found*_, not expected:
        *ACTUAL_RESULT*: {code}No result found{code}

*NB!:* the issue appears only if tables are _*JOINed by a column which contains 
newly-added data types (NaN, Infinity, -Infinity)*_. The issue is not 
reproducible is a user is JOINing tables by a column containing other data types

3. Issue



*ISSUE_DESCRIPTION:* 'ORDER BY/DESC' clause behaves in different ways when 
sorting columns containing NaN values. In one case it considers NaN to be the 
largest value, in another - the smallest one. 
- Select from the attached test file (orderBy.json, attached)
        {code}SELECT name, attr4 from dfs.tmp.`orderBy.json` order by name, 
- Check the attached screen shot (orderByIssue.jpg):
        *EXPECTED_RESULT:* It was expected the 'ORDER BY' clause to sort attr4 
columns data in the same way (most probably NaN should be the largest, see *NB*)
        *ACTUAL_RESULT:* attr4 column's values were sorted in different ways: 
for 'obj1'/'obj3' NaN is the largest, for 'obj2'/'obj4' NaN is the smallest.

*NB:* Postgres as well as Java's sorting (Collection.sort() / Arrays.sort() 
methods) treats NaN as the largest value

4. Issue

*AFFECTED_FUNCTIONALITY:* max(column_name), min(column_name),

*ISSUE_DESCRIPTION:* min/max aggregation functions return the same result if 
the selected column contains NaN value.
{code}SELECT name, max(attr4), min(attr4) from dfs.tmp.`minMax.json` group by 
name    Min     Max
obj1    NaN     NaN
obj2    NaN     NaN
obj3    NaN     NaN
obj4    NaN     NaN

As for me, this logic should be revised, current behavior is a bit confusing:
- Postgres considers NaN to be the largest value, so {{MAX(col_withNaN)}} will 
return NaN, {{MIN(col_withNaN)}} - will return other value, so DRILL's min/max 
logic can be adjusted to the Postgres' one, see postgres.jpg
- Or NAN can behave like NULL - DRILL's MIN/MAX functions ignore NULLs

This message was sent by Atlassian JIRA

Reply via email to