[IMPALA-6550] Support NAN in min-max runtime filters

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: Impala 2.12.0
• Component/s: Backend
• Labels: kudu runtime-filters
• Epic Color: ghx-label-2
• Assignee: Unassigned
• Reporter: Thomas Tauber-Marshall
• Created: 21/Feb/18 00:44
• Updated: 21/Feb/18 00:44

Description

When be/src/util/min-max-filter.* was written, we didn't think about the implications for float/double filters that have 'nan' inserted into them.

By a lucky coincidence, this isn't currently causing any problems, because:
1. Float/DoubleMinMaxFilter start out with min=std::max, max=std::min, and since comparing 'nan' to any number returns false the min/max will never be updated to 'nan'
2. MinMaxFilter is only used for joins on '=' (we don't currently support joins on 'is not distinct' because Kudu doesn't have a way of specifying a filter that allows nulls to pass), and 'nan' will never be output from a '=' join

But we should think about what the right behavior here is since MinMaxFilter will eventually be used more widely.