Thanks for the explanation Dewey!

I've opened PR:
https://github.com/apache/parquet-format/pull/560

Let me know what you think!

P.S. - If you know any relevant party in geo community, lets involve them
explicitly as well.

cc: @Jia Yu <[email protected]>

Thanks,
Milan

On Sat, 28 Mar 2026 at 03:32, Dewey Dunnington <[email protected]>
wrote:

> Hi Milan,
>
> > Dewey, you mentioned current writers using inline strings - what are they
> > inlining ? are they inlining projjsons or authority:identifiers ?
>
> The writers are writing the CRS representation they receive, which for
> Arrow C++ and arrow-rs  comes from the geoarrow.wkb extension type
> metadata [1]. This is usually PROJJSON but is permitted to be a string
> (including authority:code). If you run
> pyarrow.parquet.write_table(pyarrow.table(geopandas_geof.to_arrow()))
> today, you will get a Parquet file with an inlined PROJJSON CRS,
> because that is how GeoPandas encodes CRSes when converting to Arrow.
>
> > reality of current implementations is such
> > that most implementations do write `authorithy:identifier`, spec should
> be
> > written so that at least it doesn't look like thats invalid.
>
> The reality of current implementations is that they are writing
> PROJJSON, although I would also happily support a rewording that adds
> authority:code to the recommended options list.
>
> > Arent EPSG:<number> also understood to map directly to
> > corresponding PROJJSON definition ?
>
> They can be mapped to a PROJJSON definition (or a number of other less
> friendly export formats) using a database with the licensing ambiguity
> Jia mentioned. Conversely, PROJJSON can be mapped to authority:code
> with some minimal JSON parsing (we do this in Arrow C++ and arrow-rs
> to canonically remove CRS definitions that correspond to lon/lat to
> produce more universally consumable Parquet files).
>
> Cheers,
>
> -dewey
>
> [1] https://geoarrow.org/extension-types.html#extension-metadata
>
> On Fri, Mar 27, 2026 at 3:52 PM Milan Stefanovic
> <[email protected]> wrote:
> >
> >  Thanks Jia and Dewey,
> >
> > Dewey, you mentioned current writers using inline strings - what are they
> > inlining ? are they inlining projjsons or authority:identifiers ?
> > Given that current implementations avoided using srid:<number> and
> > projjson:<field_ref> perhaps we should remove these examples from spec as
> > they seem to bring some confusion.
> >
> > @Jia Yu <[email protected]>, you mentioned that OGC:CRS84 are understood
> to
> > map directly to its corresponding PROJJSON definition.
> > Arent EPSG:<number> also understood to map directly to
> > corresponding PROJJSON definition ?
> >
> > Also I'm fine with not being explicit about `authorithy:identifier` if
> that
> > was the prior consensus, but if reality of current implementations is
> such
> > that most implementations do write `authorithy:identifier`, spec should
> be
> > written so that at least it doesn't look like thats invalid.
> >
> > What are your thoughts?
> >
> > Milan
> >
> > On Wed, 25 Mar 2026 at 15:53, Dewey Dunnington <
> [email protected]>
> > wrote:
> >
> > > Hi Milan,
> > >
> > > A short answer is that the current language of the spec does not
> > > forbid writing "OGC:CRS84" to the CRS field (which is "just a string"
> > > as far as thrift is concerned). All existing readers that I know about
> > > (DuckDB, arrow-rs, Arrow C++, GDAL) will accept that string and
> > > interpret it unambiguously on read (for example,
> > > `GeoPandas.from_arrow(pyarrow.parquet.read_table(...))` works). There
> > > is also an example file in parquet-testing that covers this case
> > > (arbitrary string that is neither of the recommended options) [1]. I
> > > put together a small example script to demonstrate the read path for
> > > the tools I mentioned [2].
> > >
> > > Jia is correct that the GeoParquet community will require writing an
> > > inline PROJJSON string in the forthcoming 2.0 version of the
> > > specification [3]. This was a pragmatic decision that reflects the
> > > needs of existing GeoParquet users because:
> > >
> > > - srid does not explicitly name the EPSG database, so any code written
> > > there does not have an unambiguous interpretation (even if it did it
> > > would place ambiguous licencing and/or dependency requirements on
> > > consumers)
> > > - projjson:some_field was not pragmatic to implement on the write side
> > > for either of the implementations I was involved in (C++ and Rust).
> > > Implementations just don't expose the global key/value metadata when
> > > converting types and doing so would have required breaking changes in
> > > the APIs. There are also ambiguities with respect to existing
> > > propagation of schema metadata (i.e., the projjson schema key is often
> > > propagated in unexpected ways into pyarrow and beyond, including being
> > > written into the key/value metadata of a resulting Parquet file).
> > >
> > > As a result, most of the tools that can write GEOMETRY and GEOGRAPHY
> > > (Arrow C++, GDAL, arrow-rs are currently writing inline strings
> > > (because inline strings are what is available in the representation
> > > passed to Arrow-based writers and this was better than omitting CRS
> > > information). For all the implementations I was involved in, we also
> > > try to explicitly omit the CRS when we detect that the string we were
> > > passed is lon/lat (i.e., if they see "OGC:CRS84", they write an
> > > omitted CRS to minimize the need for consumers to be CRS aware).
> > >
> > > I'll echo Jia's comment that none of us are keen to reopen a CRS
> > > discussion but I also agree that the current language of the spec is
> > > vague and doesn't reflect the reality of the ecosystem as it has
> > > evolved. I'm happy to review any PRs to improve the language or
> > > implementations :)
> > >
> > > Cheers,
> > >
> > > -dewey
> > >
> > > [1]
> > >
> https://github.com/apache/parquet-testing/tree/master/data/geospatial#geospatial-test-files
> > > [2]
> https://gist.github.com/paleolimbot/7759e58bf1f98ecf8f2c459367bbdeda
> > > [3]
> > >
> https://github.com/opengeospatial/geoparquet/blob/main/format-specs/geoparquet.md#crs-parquet-property
> > >
> > > On Wed, Mar 25, 2026 at 12:49 AM Jia Yu <[email protected]> wrote:
> > > >
> > > > Hi Milan,
> > > >
> > > > The authority:identifier pattern was explicitly rejected in prior
> > > > community discussions. The core concern is that it forces query
> > > > engines to rely on external registries to resolve CRS definitions,
> > > > which breaks the goal of self-contained data. More importantly, the
> > > > most widely used authority, the EPSG database, comes with licensing
> > > > terms that are not particularly open-source friendly:
> > > > https://epsg.org/terms-of-use.html
> > > >
> > > > As a result, the community has leaned toward requiring data writers
> to
> > > > use a fully self-contained CRS representation such as PROJJSON. In
> > > > that model, a reference like OGC:CRS84 is understood to map directly
> > > > to its corresponding PROJJSON definition, as outlined in the
> > > > GeoParquet specification:
> > > >
> > >
> https://github.com/opengeospatial/geoparquet/blob/main/format-specs/geoparquet.md#ogccrs84-details
> > > >
> > > > That said, this expectation is not clearly spelled out in the Parquet
> > > > and Iceberg specifications, which leaves some ambiguity in practice.
> > > >
> > > > I don’t have a strong stance either way. In fact, I can see the case
> > > > for allowing authority:identifier. But it’s worth noting that
> > > > introducing it now would likely reopen a fairly contentious
> discussion
> > > > in the community.
> > > >
> > > > Jia
> > > >
> > > > On Tue, Mar 24, 2026 at 10:09 AM Milan Stefanovic
> > > > <[email protected]> wrote:
> > > > >
> > > > > Hi everyone,
> > > > >
> > > > > I’m looking for some clarification (and potentially a small spec
> > > update)
> > > > > regarding the Geospatial Physical Types documentation -
> > > > > https://parquet.apache.org/docs/file-format/types/geospatial/,
> > > specifically
> > > > > the CRS Customization section.
> > > > >
> > > > > 1) The Confusion
> > > > >
> > > > > Currently, the spec states that custom CRS values should follow the
> > > > > `type:identifier` format, where type is either `srid` or
> `projjson` -
> > > > > (e.g., `srid:4326` or `projjson:property_name`). The spec also
> defines
> > > the
> > > > > default CRS as `OGC:CRS84`.
> > > > >
> > > > > Depending on how the specification is read, the reader may
> consider as
> > > > > valid CRS definition to be only strings of the form `srid:<some
> > > number>` or
> > > > > `projjson:<property name>`, which implies that `OGC:CRS84` does not
> > > adhere
> > > > > to the rules defined in the customization section. This creates
> > > confusion
> > > > > for implementers: should the type string always be parsed as a
> strict
> > > > > "custom" format which necessitates the srid: prefix?
> > > > >
> > > > > 2) The Suggestion
> > > > >
> > > > > I suggest we update the language to be explicit about allowed
> formats
> > > for
> > > > > CRS, and my suggestion is that we break it down like this:
> > > > >    - Standard CRS: Any string from a known authority in a format of
> > > > > `<authority>:<identifier>` (e.g., `EPSG:4326`, `OGC:CRS84`,
> > > `ESRI:102100`)
> > > > > is accepted.
> > > > >    - Custom CRS: in the format of `type:identifier`
> > > > >          - `srid:1234`: The definition resides in a local/database
> > > spatial
> > > > > reference table.
> > > > >          - `projjson:key`: The definition is stored in Parquet
> > > file/table
> > > > > metadata.
> > > > >
> > > > > This would validate `OGC:CRS84` as a first-class string while
> > > providing a
> > > > > clear "escape hatch" for custom definitions.
> > > > >
> > > > > What are your thoughts ?
> > > > >
> > > > > Kind regards,
> > > > > Milan
> > >
>

Reply via email to