[DRILL-1977] Wrong result with aggregation on top of UNION ALL operator

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.8.0
• Fix Version/s: 0.8.0
• Component/s: Query Planning & Optimization
• Assignee: Sean Hsuan-Yi Chu
• Reporter: Victoria Markman
• Created: 10/Jan/15 01:30
• Updated: 28/Apr/15 00:10
• Resolved: 10/Mar/15 17:07

Description

Jan 10 01:17:13 root@~ ] # hadoop fs -cat /test/t.json
{ "a1": 0, "b1": 0, "c1": "true",   "d1" : "2015-01-02"}
{ "a1": 0, "b1": 0, "c1": "false" , "d1" : "2015-01-03"}
{ "a1": 0, "b1": 0, "c1": "false" , "d1" : "2015-01-04"}
{ "a1": 1, "b1": 1, "c1": "true" ,  "d1" : "2015-01-05"}
{ "a1": 1, "b1": 1, "c1": "true" ,  "d1" : "2015-01-06"}

Query below returns wrong result, we should have 4 groups, instead we get only two groups from right side of the union.
Notice lack of project step above union all (unclear if it has anything to to with the bug)

0: jdbc:drill:schema=dfs> select calc1, max(a1), min(b1), count(c1) from (select a1+10 as calc1, a1, b1, c1 from `t.json` union all select a1+100 as calc1, a1, b1, c1 from `t.json`) group by calc1;
+------------+------------+------------+------------+
|   calc1    |   EXPR$1   |   EXPR$2   |   EXPR$3   |
+------------+------------+------------+------------+
| 100        | 0          | 0          | 3          |
| 101        | 1          | 1          | 2          |
+------------+------------+------------+------------+
2 rows selected (0.181 seconds)

00-01      Project(calc1=[$0], EXPR$1=[$1], EXPR$2=[$2], EXPR$3=[$3])
00-02        HashAgg(group=[{0}], EXPR$1=[MAX($1)], EXPR$2=[MIN($2)], EXPR$3=[COUNT($3)])
00-03          UnionAll(all=[true])
00-05            Project(calc1=[+($2, 10)], a1=[$2], b1=[$1], c1=[$0])
00-07              Scan(groupscan=[EasyGroupScan [selectionRoot=/test/t.json, numFiles=1, columns=[`a1`, `b1`, `c1`], files=[maprfs:/test/t.json]]])
00-04            Project(calc1=[+($2, 100)], a1=[$2], b1=[$1], c1=[$0])
00-06              Scan(groupscan=[EasyGroupScan [selectionRoot=/test/t.json, numFiles=1, columns=[`a1`, `b1`, `c1`], files=[maprfs:/test/t.json]]])

Correct result, when order of the columns is different in aggregate functions: "min(b1), max(a1)" instead of "max(a1), min(b1)"

0: jdbc:drill:schema=dfs> select calc1, min(b1), max(a1), count(c1) from (select a1+10 as calc1, a1, b1, c1 from `t.json` union all select a1+100 as calc1, a1, b1, c1 from `t.json`) group by calc1;
+------------+------------+------------+------------+
|   calc1    |   EXPR$1   |   EXPR$2   |   EXPR$3   |
+------------+------------+------------+------------+
| 10         | 0          | 0          | 3          |
| 11         | 1          | 1          | 2          |
| 100        | 0          | 0          | 3          |
| 101        | 1          | 1          | 2          |
+------------+------------+------------+------------+
4 rows selected (0.256 seconds)

00-01      Project(calc1=[$0], EXPR$1=[$1], EXPR$2=[$2], EXPR$3=[$3])
00-02        HashAgg(group=[{0}], EXPR$1=[MIN($1)], EXPR$2=[MAX($2)], EXPR$3=[COUNT($3)])
00-03          Project(calc1=[$0], b1=[$2], a1=[$1], c1=[$3])
00-04            UnionAll(all=[true])
00-06              Project(calc1=[+($2, 10)], a1=[$2], b1=[$1], c1=[$0])
00-08                Scan(groupscan=[EasyGroupScan [selectionRoot=/test/t.json, numFiles=1, columns=[`a1`, `b1`, `c1`], files=[maprfs:/test/t.json]]])
00-05              Project(calc1=[+($2, 100)], a1=[$2], b1=[$1], c1=[$0])
00-07                Scan(groupscan=[EasyGroupScan [selectionRoot=/test/t.json, numFiles=1, columns=[`a1`, `b1`, `c1`], files=[maprfs:/test/t.json]]])

Non-union query works correctly:

0: jdbc:drill:schema=dfs> select a1+10 as calc1, min(a1), max(b1), count(c1) from `t.json` group by a1+10;
+------------+------------+------------+------------+
|   calc1    |   EXPR$1   |   EXPR$2   |   EXPR$3   |
+------------+------------+------------+------------+
| 10         | 0          | 0          | 3          |
| 11         | 1          | 1          | 2          |
+------------+------------+------------+------------+
2 rows selected (0.142 seconds)
0: jdbc:drill:schema=dfs> select a1+10 as calc1, min(b1), max(a1), count(c1) from `t.json` group by a1+10;
+------------+------------+------------+------------+
|   calc1    |   EXPR$1   |   EXPR$2   |   EXPR$3   |
+------------+------------+------------+------------+
| 10         | 0          | 0          | 3          |
| 11         | 1          | 1          | 2          |
+------------+------------+------------+------------+
2 rows selected (0.153 seconds)

If I switch legs of a UNION, we still get wrong result, but from the right side:

0: jdbc:drill:schema=dfs> select calc1, max(a1), min(b1), count(c1) from (select a1+100 as calc1, a1, b1, c1 from `t.json` union all select a1+10 as calc1, a1, b1, c1 from `t.json`) group by calc1;
+------------+------------+------------+------------+
|   calc1    |   EXPR$1   |   EXPR$2   |   EXPR$3   |
+------------+------------+------------+------------+
| 10         | 0          | 0          | 3          |
| 11         | 1          | 1          | 2          |
+------------+------------+------------+------------+
2 rows selected (0.19 seconds)

Issue Links

• depends upon: DRILL-2207 Fix Union All Operator
• is related to: DRILL-2376 UNION ALL on Aggregates with GROUP BY returns incomplete results