Thank you Russel. I updated the language to focus on reader semantics and
saying that a reader can choose to read either.

On Mon, Aug 24, 2026 at 4:37 PM Russell Spitzer <[email protected]>
wrote:

> +1, Although I'm not sure we should add language about the bytes in the
> location and inline *must* being the same. Since i'm not sure it's
> something we could actually control from within the parquet universe. If a
> writer knows the two are identical at the time of writing the row is that
> enough? I think it's fine to outline the fallback behavior and just say,
> "When a locator is present, it is implied that a reader can either use the
> locator or inline bytes interchangeably."
>
> On Mon, Aug 24, 2026 at 9:03 AM Alkis Evlogimenos via dev <
> [email protected]> wrote:
>
>> Hi folks,
>>
>> I would like to start a vote on adopting the format change in PR 603 [1]:
>> remove self-references from the FILE logical type. This came out of the
>> discussion on inheriting the inline codec [2].
>>
>> FILE as merged lets a value set offset and size with no uri, which
>> addresses a byte range inside the containing file. This byte range is
>> effectively owned by the Parquet writer. Specifying that range properly
>> means giving it a compression block, an encryption module, an AAD
>> identity,
>> and its own size accounting. That is a second page mechanism reachable
>> only
>> through FILE. The use case is mid sized values and it is better served by
>> non-contiguous pages, which we will propose separately. Removing
>> self-references keeps FILE to the forms it specifies cleanly, while the
>> format has not shipped in a release.
>>
>> The changes:
>>
>> 1. Remove self-references. A value that sets offset and size with no uri
>> is
>> invalid. offset may only be set together with uri.
>> 2. Resolve uri uniformly as an external reference, including when it names
>> the file that contains it. Parquet applies no compression or encryption of
>> its own to those bytes. The parquet writer does NOT own these bytes. An
>> earlier draft forbade referencing the current file, but neither a writer
>> nor a reader can check that, so the encoding is uniform instead.
>> 3. Drop the prohibition on modular encryption. It existed only because
>> self-referenced ranges were not encryption modules. The fields of a FILE
>> group are ordinary columns and are encoded, compressed, and encrypted like
>> any other column, inline included.
>> 4. Allow inline together with the locator fields. Both must denote the
>> same
>> bytes, so a reader may resolve from either and get the same result, and
>> inline is the cheaper path because it needs no external access. If the two
>> disagree the value is invalid, and a reader is not required to detect it.
>> This combination was underspecified before.
>>
>> The spec changes are in https://github.com/apache/parquet-format/pull/603
>> and touch LogicalTypes.md and a comment in parquet.thrift. Encryption.md
>> is
>> unchanged from master and no Thrift structures change. Inline and external
>> references behave as before for existing readers and writers.
>>
>> Please vote in the next 72 hours.
>>
>> [ ] +1 Approve the proposed format change
>> [ ] +0 No opinion
>> [ ] -1 Do not approve (please provide specific reasons)
>>
>> Cheers,
>>
>> [1] https://github.com/apache/parquet-format/pull/603
>> [2] https://lists.apache.org/thread/zrzc7t9fccg92rx3h4fw3ndw3bdo5xr7
>>
>

Reply via email to