[HIVE-6614] Vectorized aggregates computed on map side diffe (hash mode) from values computed on reduce side (streaming mode)

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Assignee: Remus Rusanu
• Reporter: Remus Rusanu
• Created: 11/Mar/14 15:25
• Updated: 11/Mar/14 15:25

Description

HIVE-6222 allows vectorized aggregates to operate on streaming mode, ie. flush after each key change and let the shuffle+reduce side to compute the final aggregate values. An error in patch .2 for HIVE-6222 shows that when the queries run in streaming mode, there are rounding diffs for some agg functions (VAR and friends). These occurred for non-decimal types, like ctinyint:

select csmallint, VAR_POP(ctinyint) from alltypesorc where csmallint = -75 group by csmallint;

This produces 107.55555555555556 vs. 107.55555555555554.