[HIVE-20660] Group by statistics estimation could be improved by bounding the total number of rows to source table

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 4.0.0
• Fix Version/s: 4.0.0-alpha-1
• Component/s: Statistics
• Assignee: Vineet Garg
• Reporter: Vineet Garg
• Created: 30/Sep/18 23:58
• Updated: 17/Nov/22 08:54
• Resolved: 15/Oct/18 03:36

Description

Currently the stats for group by is estimated by taking product of NDVs of all the keys and bounding it by the number of rows of its input. This bound could be improved by using the source table instead of immediate input, the insight in this case is that cardinality/ndvs of a table can not go beyond the original (outer joins will only add NULLs thereby increasing the cardinality by 1).

Note that the assumption here is that group by keys all belong to the same source table/input.
This will improve the estimation in situations where group by is executed after joins wherein Hive could end up estimating the number of rows.

Reproducer

set hive.stats.fetch.column.stats=true;

create table t1(i int, j int);
alter table t1 update statistics set('numRows'='10000', 'rawDataSize'='18000');
alter table t1 update statistics for column i set('numDVs'='2500','numNulls'='50','highValue'='1000','lowValue'='0');
alter table t1 update statistics for column j set('numDVs'='500','numNulls'='30','highValue'='100','lowValue'='50');

create table t2(i2 int, j2 int);
alter table t2 update statistics set('numRows'='100000000', 'rawDataSize'='10000');
alter table t2 update statistics for column i2 set('numDVs'='10000000','numNulls'='0','highValue'='8000','lowValue'='0');
alter table t2 update statistics for column j2 set('numDVs'='10','numNulls'='0','highValue'='800','lowValue'='-1');

explain select count (1) from t1,t2
	where t1.j=t2.i2 
group by t1.i, t1.j;

Reducer 2
            Reduce Operator Tree:
              Merge Join Operator
                condition map:
                     Inner Join 0 to 1
                keys:
                  0 _col1 (type: int)
                  1 _col0 (type: int)
                outputColumnNames: _col0, _col1
                Statistics: Num rows: 99700 Data size: 797288 Basic stats: COMPLETE Column stats: COMPLETE
                Group By Operator
                  aggregations: count()
                  keys: _col0 (type: int), _col1 (type: int)
                  mode: hash
                  outputColumnNames: _col0, _col1, _col2
                  Statistics: Num rows: 49850 Data size: 797448 Basic stats: COMPLETE Column stats: COMPLETE <==========
                  Reduce Output Operator
                    key expressions: _col0 (type: int), _col1 (type: int)
                    sort order: ++
                    Map-reduce partition columns: _col0 (type: int), _col1 (type: int)
                    Statistics: Num rows: 49850 Data size: 797448 Basic stats: COMPLETE Column stats: COMPLETE
                    value expressions: _col2 (type: bigint)
  .....................

Issue Links

• causes: HIVE-20842 Fix logic introduced in HIVE-20660 to estimate statistics for group by