Correction: I looked deeper into the BigQuery and Trino implementations,
and both are using 2 separate integers as Felipe is proposing. I think it's
worth updating the proposal to reflect this layout. Thanks, folks!

*  •  **Tim Sweña (Swast)*
*  •  *Team Lead, BigQuery DataFrames
*  •  *Google Cloud Platform
*  •  *Chicago, IL, USA


On Fri, Nov 21, 2025 at 9:37 AM Tim Swena <[email protected]> wrote:

> > Would there be any reason to have (or not have) a canonical LogicalType
> for these in Parquet as well?
>
> I think it would be appropriate to add this to Parquet as well. I assume
> there's a different process / mailing list for that?
>
> > our goal here should be to standardize existing practice, not come up
> with a novel representation, IMHO.
>
> BigQuery is using 128-bits, which is why I went this proposal.
>
> Trino is using 96-bits (
> https://github.com/trinodb/trino/blob/eef66628759d7244c176f62be45f3d9f0e5a1a5d/core/trino-spi/src/main/java/io/trino/spi/type/LongTimestampType.java)
> but doesn't seem to me that would be much more efficient compared to 128.
>
> *  •  **Tim Sweña (Swast)*
> *  •  *Team Lead, BigQuery DataFrames
> *  •  *Google Cloud Platform
> *  •  *Chicago, IL, USA
>
>
> On Wed, Nov 19, 2025 at 3:35 AM Antoine Pitrou <[email protected]> wrote:
>
>>
>> I don't have a personal opinion on which representation is technical
>> better, but our goal here should be to standardize existing practice,
>> not come up with a novel representation, IMHO.
>>
>> Regards
>>
>> Antoine.
>>
>>
>> Le 18/11/2025 à 23:45, Felipe Oliveira Carvalho a écrit :
>> > One reason to avoid 128-bit integers is the requirement for 128-bit
>> > operations that it creates. Many high-resolution time representations
>> split
>> > the value in two integers in a way that is useful for many time-related
>> > operations.
>> >
>> > The picosecond resolution can be achieved by splitting into a (seconds:
>> > i64, picoseconds: i64) pair where the number of picoseconds in a day can
>> > fit in 53 bits and the number of seconds can represent much more than
>> 10K
>> > years in number of seconds.
>> >
>> > This removes the need for a128-bit division by 86400 to do anything
>> > interesting with the picoseconds timestamp. This layout could be a
>> > Canonical Extension Type proposal with the seconds timestamp fields
>> being
>> > one of the existing timestamp types allowing for very cheap casts from
>> the
>> > extension type to the timestamp with the precision in seconds.
>> >
>> > --
>> > Felipe
>> >
>> > On Tue, Nov 18, 2025 at 6:22 PM Curt Hagenlocher <[email protected]>
>> > wrote:
>> >
>> >> For both Duration and Timestamp, this would require adding a new field
>> >> to the FlatBuffers spec. That should be okay, right?
>> >>
>> >> A 128-bit timestamp would be useful at a nanosecond scale as well;
>> >> there are databases like Snowflake which support a precision and scale
>> >> for timestamps that force either truncation of precision or clipping
>> >> of range when representing as Arrow.
>> >>
>> >> Would there be any reason to have (or not have) a canonical
>> >> LogicalType for these in Parquet as well?
>> >>
>> >> On Fri, Nov 7, 2025 at 1:29 PM Tim Swena <[email protected]>
>> wrote:
>> >>>
>> >>> Hello,
>> >>>
>> >>> Per the process described at
>> >>>
>> >>
>> https://arrow.apache.org/docs/format/Changing.html#discussion-and-voting-process
>> >>> I am starting a discussion thread for the following spec change
>> proposal:
>> >>>
>> >>>
>> >>>     1.
>> >>>
>> >>>     Add a new time unit: PICOSECOND, which is unsupported in the
>> existing
>> >>>     64-bit timestamp-related types.
>> >>>     2.
>> >>>
>> >>>     Add support for bitWidth=128 to the timestamp data type, which
>> >> supports
>> >>>     all units, including PICOSECOND.
>> >>>     3.
>> >>>
>> >>>     Add support for bitWidth=128 to the duration data type, which
>> supports
>> >>>     all units, including PICOSECOND.
>> >>>
>> >>> This is motivated by some currently experimental changes in BigQuery
>> to
>> >>> support picosecond precision timestamps (source
>> >>> <
>> >>
>> https://docs.cloud.google.com/bigquery/docs/reference/storage/rpc/google.cloud.bigquery.storage.v1?content_ref=read%20api%20will%20return%20full%20precision%20picosecond%20value%20the%20value%20will%20be%20encoded%20as%20a%20string%20which%20conforms%20to%20iso%208601%20format#picostimestampprecision
>> >>> ),
>> >>> but from what I can tell such timestamps already have some support in
>> IBM
>> >>> Db2 (source
>> >>> <
>> >>
>> https://www.ibm.com/docs/en/db2-for-zos/13.0.0?topic=jdbc-dbtimestamp-class&content_ref=the+com+ibm+db2+jcc+dbtimestamp+class+can+be+used+to+create+timestamp+objects+with+a+precision+of+up+to+picoseconds+and+time+zone+information
>> >>> )
>> >>> and Trino (source
>> >>> <
>> >>
>> https://trino.io/docs/current/language/types.html?content_ref=heading+calendar+date+and+time+of+day+without+a+time+zone+with+pdigits+of+precision+for+the+fraction+of+seconds+a+precision+of+up+to+12+picoseconds+is+supported
>> >>> ).
>> >>> Note that reference implementation(s) are still very much a
>> >>> work-in-progress (https://github.com/apache/arrow/pull/48018 for a
>> >> start in
>> >>> C++), but I figured it would be useful to kick off the conversation
>> >> before
>> >>> diving in too much further into implementation.
>> >>>
>> >>> Inspired by other discussions, I've created a draft of a more formal
>> RFC
>> >>> document here: Arrow-RFC: timestamp128 and duration128 data types with
>> >>> support for picosecond units
>> >>> <
>> >>
>> https://docs.google.com/document/d/1-S0qvYTIEGlLnNkkgyWSHfnIvU4xpFqDQuMNTojaj9A/edit?tab=t.0#heading=h.as1aixu509k7
>> >>>
>> >>>
>> >>> *  •  **Tim Sweña (Swast)*
>> >>> *  •  *Team Lead, BigQuery DataFrames
>> >>> *  •  *Google Cloud Platform
>> >>> *  •  *Chicago, IL, USA
>> >>
>> >
>>
>>

Reply via email to