[
https://issues.apache.org/jira/browse/PARQUET-758?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17598151#comment-17598151
]
ASF GitHub Bot commented on PARQUET-758:
----------------------------------------
emkornfield commented on PR #184:
URL: https://github.com/apache/parquet-format/pull/184#issuecomment-1232420353
> t is not that trivial. For the half-precision floating point numbers we do
not have native support for either cpp or java so we can define the total
ordering as we want. But we shall do the same for the existing floating point
numbers that most languages have native support. Even though they are following
the same standard the total ordering either does not exist or have different
implementations. See
[PARQUET-1222](https://issues.apache.org/jira/browse/PARQUET-1222) for details.
I think these are orthogonal. I might be missing something but it seems
like it would not be to hard to case float16 to float in java/cpp and do the
comparison in that space and cast it back down. This might not be the most
efficient implementation but would be straightforward? I am probably missing
something. It would be nice to resolve
[PARQUET-1222](https://issues.apache.org/jira/browse/PARQUET-1222) so the same
semantics would apply to all floating point numbers.
> The tricky thing will be the implementations. Even though parquet-mr does
not really care about converting the values according to their logical types we
still need to care about the logical types at the ordering (min/max values in
the statistics).
It seems this would require parquet implementations to null out statistics
for logical types that they don't support, does parquet-mr do that today?
> [Format] HALF precision FLOAT Logical type
> ------------------------------------------
>
> Key: PARQUET-758
> URL: https://issues.apache.org/jira/browse/PARQUET-758
> Project: Parquet
> Issue Type: Improvement
> Components: parquet-format
> Reporter: Julien Le Dem
> Priority: Minor
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
