Hi,

After discussing with Micah, I've split out the spec change to describe how
unsupported logical/physical type
combinations should be handled by readers (TLDR the logical type and stats
should be dropped). I've opened
a parquet-format PR [1] and parquet-java PR [2] for the spec change and
reference implementation. The previous
spec change [3] and implementation are still open and ready for review.
Please take a look if interested.

Best,
Div

[1] https://github.com/apache/parquet-format/pull/606
[2] https://github.com/apache/parquet-java/pull/3711
[3] https://github.com/apache/parquet-format/pull/601
[4] https://github.com/apache/parquet-java/pull/3680

On Tue, Jul 21, 2026 at 7:42 PM Divjot Arora <[email protected]>
wrote:

> Hi folks,
>
> I've opened PRs for parquet-format [1] and parquet-java [2] to support the
> FLBA(12) physical type for the TimestampType logical annotation. Please
> take a look if interested and comment.
>
> Best,
> Div Arora
>
> [1] https://github.com/apache/parquet-format/pull/601
> [2] https://github.com/apache/parquet-java/pull/3680
>
> On Thu, Jul 16, 2026 at 4:16 PM Russell Spitzer <[email protected]>
> wrote:
>
>> I'd lean toward choosing a bit width and going forward for similar
>> reasons.
>> I don't think a lot
>> of engines/users are going to custom-define their bit-width ahead of time.
>> It seems more
>> likely if we wanted to benefit readers, we would just store the max
>> used-bit width as a stat.
>>
>> On Thu, Jul 16, 2026 at 12:30 AM Micah Kornfield <[email protected]>
>> wrote:
>>
>> > Being able to use compact data types is a good thing, no? :-)
>> >
>> >
>> > Generally yes, but I think there is a difference here between Decimal
>> and
>> > Timestamp. Decimal a user is specifying precision and scale up front,
>> so we
>> > can calculate the exact FLBA width. The closest thing we have with
>> > timestamp is for SQL purposes it has to cover a 10K year range. But in
>> > general, we won't know the exact min and max value up front.
>> >
>> > For existing granularities of milliseconds and microseconds there is not
>> > any benefit to using FLBA (users should use existing physical integer
>> > values).  Extra engineering effort for these granularities doesn't seem
>> > worthwhile.
>> >
>> > For nanoseconds, I believe we could use the FLBA<8> type. However, some
>> > systems are starting to move toward picoseconds [1][2] (and I've heard
>> > femtoseconds discussed) so we will eventually want to support all of
>> those
>> > ranges. FLBA<9> allows us to do it all with the same code.
>> >
>> > I think extending the spec to support different size types is certainly
>> > something we could do in the future, but doing it now seems like it is
>> just
>> > more code to maintain without a concrete use-case.
>> >
>> > Cheers,
>> > Micah
>> >
>> > [1]
>> >
>> >
>> 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
>> > [2] https://trino.io/docs/current/language/types.html#timestamp-p
>> >
>> > On Fri, Jul 10, 2026 at 2:38 AM Antoine Pitrou <[email protected]>
>> wrote:
>> >
>> > >
>> > > Le 10/07/2026 à 11:27, Divjot Arora via dev a écrit :
>> > > >> Yeah, well, regardless, this is derailing the discussion quite a
>> bit.
>> > > > This proposal is not about extending DBP to FLBA. Can we keep it on
>> > > topic?
>> > > >
>> > > > Agreed, let's keep this discussion focused to extended-precision
>> > > > timestamps. We can have separate discussions on encodings..
>> > > >
>> > > >> And conversely, I do not see what we get out of mandating FLBA<12>.
>> > :-)
>> > > >
>> > > > FLBA<N> works, but I would be concerned about added complexity on
>> the
>> > > > reader side as I assume engines will want to use compact data types
>> to
>> > > > represent timestamps as much as possible, so they'd have to manage
>> > > > different data types for different values of N.
>> > >
>> > > Being able to use compact data types is a good thing, no? :-)
>> > >
>> > > The good news is that some readers already have similar code for
>> reading
>> > > Decimals from arbitrary FLBA<N>. For example, Parquet C++ can read
>> > > Decimal-annotated FLBA<N> into Arrow Decimal128 or Decimal256,
>> depending
>> > > on the advertised precision.
>> > >
>> > > (and the code to do that is reasonably simple: you're really
>> > > sign-extending from one integer width to another)
>> > >
>> > > Regards
>> > >
>> > > Antoine.
>> > >
>> > >
>> > >
>> >
>>
>

Reply via email to