[
https://issues.apache.org/jira/browse/HIVE-21290?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16797158#comment-16797158
]
Karen Coppage commented on HIVE-21290:
--------------------------------------
Patch 1 notes:
* Timestamps are converted from JVM time zone, not session ("set time zone...")
time zone, this is for backwards compatibility reasons.
* The writer time zone has to be passed through all the vectorized readers so
that
org.apache.hadoop.hive.ql.io.parquet.vector.ParquetDataColumnReaderFactory.TypesFromInt96PageReader#convert
can correctly convert int96 to Timestamp.
* ^ It might be a better idea to pass the entire reader metadata (Map<String,
String> with ~5 elements) instead of extracting skipConversion (boolean) and
writerTimezone (ZoneId) and passing these through all those constructors. Any
input is welcome.
> Restore historical way of handling timestamps in Parquet while keeping the
> new semantics at the same time
> ---------------------------------------------------------------------------------------------------------
>
> Key: HIVE-21290
> URL: https://issues.apache.org/jira/browse/HIVE-21290
> Project: Hive
> Issue Type: Sub-task
> Reporter: Zoltan Ivanfi
> Assignee: Karen Coppage
> Priority: Major
> Attachments: HIVE-21290.1.patch
>
>
> This sub-task is for implementing the Parquet-specific parts of the following
> plan:
> h1. Problem
> Historically, the semantics of the TIMESTAMP type in Hive depended on the
> file format. Timestamps in Avro, Parquet and RCFiles with a binary SerDe had
> _Instant_ semantics, while timestamps in ORC, textfiles and RCFiles with a
> text SerDe had _LocalDateTime_ semantics.
> The Hive community wanted to get rid of this inconsistency and have
> _LocalDateTime_ semantics in Avro, Parquet and RCFiles with a binary SerDe as
> well. *Hive 3.1 turned off normalization to UTC* to achieve this. While this
> leads to the desired new semantics, it also leads to incorrect results when
> new Hive versions read timestamps written by old Hive versions or when old
> Hive versions or any other component not aware of this change (including
> legacy Impala and Spark versions) read timestamps written by new Hive
> versions.
> h1. Solution
> To work around this issue, Hive *should restore the practice of normalizing
> to UTC* when writing timestamps to Avro, Parquet and RCFiles with a binary
> SerDe. In itself, this would restore the historical _Instant_ semantics,
> which is undesirable. In order to achieve the desired _LocalDateTime_
> semantics in spite of normalizing to UTC, newer Hive versions should record
> the session-local local time zone in the file metadata fields serving
> arbitrary key-value storage purposes.
> When reading back files with this time zone metadata, newer Hive versions (or
> any other new component aware of this extra metadata) can achieve
> _LocalDateTime_ semantics by *converting from UTC to the saved time zone
> (instead of to the local time zone)*. Legacy components that are unaware of
> the new metadata can read the files without any problem and the timestamps
> will show the historical Instant behaviour to them.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
