[IMPALA-5679] Count star optimization gives incorrect result for parquet table partitioned by STRING column

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.10.0
• Fix Version/s: Impala 2.10.0
• Component/s: Backend
• Labels: correctness parquet
• Target Version: Impala 2.10.0
• Assignee: Taras Bobrovytsky
• Reporter: Attila Jeges
• Created: 19/Jul/17 14:29
• Updated: 24/Jul/17 23:14
• Resolved: 24/Jul/17 23:14

Description

The issue can be reproduced as follows:

> create table part_tbl_parq (a integer) partitioned by (p STRING) stored as parquet;
> insert into part_tbl_parq partition(p="val100") values(100);
> select * from part_tbl_parq;
+-----+--------+
| a   | p      |
+-----+--------+
| 100 | val100 |
+-----+--------+

> select p, count(a) from part_tbl_parq group by p;
+--------+----------+
| p      | count(a) |
+--------+----------+
| val100 | 1        |
+--------+----------+

> select p, count(*) from part_tbl_parq group by p;
+---+----------+
| p | count(*) |
+---+----------+
|   | 0        |
+---+----------+

The result of the last select is obviously incorrect.

The problem does not happen if the table is partitioned by an INT column:

> create table part_tbl_parq2 (a integer) partitioned by (p integer) stored as parquet;
> insert into part_tbl_parq2 partition(p=100) values(100);
> select p, count(*) from part_tbl_parq2 group by p;
+-----+----------+
| p   | count(*) |
+-----+----------+
| 100 | 1        |
+-----+----------+

Issue Links

• is broken by: IMPALA-5036 Improve COUNT(*) performance of Parquet scans.