Thanks Junbo, I think we're converging on a good direction.

I agree with keeping FIP-49 focused so we can deliver the REST write path
quickly, and I also agree that FIP-32 may be easier to review if we split
it into smaller protocol-specific FIPs over time.

>From an architectural perspective, I think it would be useful if we align
on a few high-level design principles first (as a suggestion):

   -

   FIP-49 should focus on the REST gateway MVP (writes/DDL, with any
   additional read functionality kept intentionally small if we decide to
   include it).
   -

   FIP-32 (and potential follow-up FIPs) should focus on the query
   architecture and multi-protocol support.
   -

   My understanding from the previous discussions is that DataFusion is
   primarily intended for the SQL/query path, while native Fluss operations
   continue to build on the fluss-rust foundation.

I'm also not suggesting introducing a large backend abstraction up front,
but I think both the REST and query paths should build on the same
underlying Fluss integration rather than evolving into separate
implementations over time (I could imagine a thin shared layer built on top
of fluss-rust). That could allow the REST and future protocol
implementations to share common infrastructure where it makes sense (e.g.
connection management, metadata handling and future authentication/doAs
support) without coupling all native Fluss operations to DataFusion.

If we can agree on those architectural boundaries first, I think we can
iterate on the individual protocol implementations much more independently.

WDYT? We should really make it clear, as there is much opinion and thoughts
currently in everybody's mind. I think a shared understanding can help to
reduce unnecessary efforts.

Am Sa., 1. Aug. 2026 um 03:28 Uhr schrieb Junbo Wang <[email protected]>:

> Thanks Keith for reviving this, and David & George for the FIP.
>
> +1 on narrowing scope.
> One point I'd like to check on Arrow Flight — my understanding is that the
> Arrow Flight SQL protocol in the FIP-32 <
> https://cwiki.apache.org/confluence/spaces/FLUSS/pages/406622520/FIP-32+Multiprotocol+Query+Gateway+for+Apache+Fluss>
> is a Flight SQL endpoint, which needs a SQL engine on the server side and
> thus depends on fluss-datafusion. If that's the case, since
> fluss-datafusion isn't ready yet (discussed in the fluss-rust roadmap),
> perhaps we could hold it off for 1.0 and narrow to just REST for the
> initial release, then add Flight SQL once fluss-datafusion lands. Please
> correct me if I'm misreading this.(My understanding is that plain Arrow
> Flight as a data-transport layer isn't part of this FIP — happy to discuss
> separately if there's demand.)
>
> Happy to pick up part of the implementation once the scope settles.
>
>
> Best regards,
> Junbo Wang
>
> > On 31 Jul 2026, at 18:02, Keith Lee <[email protected]> wrote:
> >
> > David, George
>
>

Reply via email to