[DRILL-2803] Severe skew due to null values in columns even when other columns are non-null

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.8.0
• Fix Version/s: 0.9.0
• Component/s: Functions - Drill
• Assignee: Chun Chang
• Reporter: Aman Sinha
• Created: 16/Apr/15 02:44
• Updated: 28/Apr/15 21:03
• Resolved: 19/Apr/15 01:08

Description

If you have 2 columns that are hashed (either for distribution or for hash based operators) and one of those columns has lots of null values, it can result in substantial skew even if the other column has non-null values.

In the following query the combined hash value of 2 columns is 0 even when 1 column is non-null. The reason is that if the starting value is null (for cr_reason_sk all values are null in the above query), it does not matter what seed is passed in. The hash function treats the second parameter as a seed and not as a combiner, so it gets ignored.

select cr_call_center_sk, cr_reason_sk, hash64(cr_reason_sk, hash64(cr_call_center_sk)) as hash_value from catalog_returns  where cr_reason_sk is null and cr_call_center_sk is not null limit 10;
+-------------------+--------------+------------+
| cr_call_center_sk | cr_reason_sk | hash_value |
+-------------------+--------------+------------+
| 1                 | null         | 0          |
| 1                 | null         | 0          |
| 4                 | null         | 0          |
| 1                 | null         | 0          |
| 4                 | null         | 0          |
| 2                 | null         | 0          |
| 2                 | null         | 0          |
| 2                 | null         | 0          |
| 2                 | null         | 0          |
| 2                 | null         | 0          |
+-------------------+--------------+------------+