David Phillips created HIVE-21376:
-------------------------------------
Summary: Incompatible change in Hive bucket computation
Key: HIVE-21376
URL: https://issues.apache.org/jira/browse/HIVE-21376
Project: Hive
Issue Type: Bug
Affects Versions: 3.1.0
Reporter: David Phillips
HIVE-20007 seems to have inadvertently changed the bucket hash code computation
via {{ObjectInspectorUtils.getBucketHashCodeOld()}} for the {{DATE}} and
{{TIMESTAMP}} data type2.
{{DATE}} was previously computed using {{DateWritable}}, which uses
{{daysSinceEpoch}} as the hash code. It is now computed using
{{DateWritableV2}}, which uses the hash code of {{java.time.LocalDate}} (which
is not days since epoch).
{{TIMESTAMP}} was previous computed using {{TimestampWritable}} and now uses
{{TimestampWritableV2}}. They ostensibly use the same hash code computation,
but there are two important differences:
# {{TimestampWritable}} rounds the number of milliseconds into the seconds
portion of the computation, but {{TimestampWritableV2}} does not.
# {{TimestampWritable}} gets the epoch time from {{java.sql.Timestamp}}, which
returns it relative to the JVM time zone, not UTC. {{TimestampWritableV2}} uses
a {{LocalDateTime}} relative to UTC.
I was unable to get Hive 3.1 running in order to verify if this actually causes
data to be read or written incorrectly (there may be code above this library
method which makes things work correctly). However, if my understanding is
correct, this means Hive 3.1 is both forwards and backwards incompatible with
bucketed tables using either of these data types. It also indicates that Hive
needs tests to verify that the hash code does not change between releases.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
