[
https://issues.apache.org/jira/browse/CALCITE-4467?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17265522#comment-17265522
]
Julian Hyde commented on CALCITE-4467:
--------------------------------------
A [discussion in
CockroachDB|https://github.com/cockroachdb/cockroach/issues/18860] noted that
Oracle behaves the same as PostgreSQL, noted problems with NaN values in unique
indexes. I also wonder whether NaN values should be grouped in GROUP BY, and
whether if we supported NaN we would also need to support -INF, +INF, signed
zero and denormal numbers.
I also found a tale of [data corruption in SQL
Server|https://www.red-gate.com/simple-talk/blogs/from-nan-to-infinity-and-beyond/]
that sounds very pertinent for systems such as Hive that can load data via
non-SQL paths.
{{RexSimplify}} is not the only place that simplification happens; it happens
in a hundred places, many of them in {{RelOptRule}} instances. I bet the {{NOT
IN}} rewrite is affected, for example.
> Incorrect simplification for 'NaN' value
> ----------------------------------------
>
> Key: CALCITE-4467
> URL: https://issues.apache.org/jira/browse/CALCITE-4467
> Project: Calcite
> Issue Type: Bug
> Components: core
> Reporter: Jesus Camacho Rodriguez
> Assignee: Jesus Camacho Rodriguez
> Priority: Major
> Labels: pull-request-available
> Time Spent: 10m
> Remaining Estimate: 0h
>
> {{RexSimplify}} simplifies {{x = x}} to {{null or x is not null}} (similarly
> <= and >=), and {{x != x}} to {{null and x is null}} (similarly < and >).
> https://github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/rex/RexSimplify.java#L363
> This may not be applicable in some cases. For instance, if the type of x is
> floating-point, x could be 'NaN'. While some RDBMS consider 'NaN' = 'NaN'
> (e.g., Postgres), some others consider 'NaN' != 'NaN' following the IEEE 754
> standard. For the latest, the rewriting above will result in incorrect
> results.
> I think we should simply ignore this simplification for floating-point type.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
