Thanks Forward and Hongshun for the thoughtful input. On ForwardXu's question about the target release — I've been thinking a bit about the 0.3.0 scope, and would like to share my personal take, mostly as a starting point for discussion.
A possible 0.3.0 scope FIP-40: consolidate fluss-rust into apache/fluss. thanks to Anton for already opening the PR-3401 <https://github.com/apache/fluss/pull/3401> DataFusion integration adapter (Theme 1), as a fluss-rust submodule. Lake + log union read, Paimon-first (Theme 2), as a separate submodule so lake-side dependencies stay isolated from the core. S erver-side filter pushdown (Theme 3) — would be great to have if capacity allows, but I'd suggest keeping it optional rather than a hard requirement for 0.3.0. Gateway — perhaps outside 0.3.0 Once fluss-rust is consolidated into apache/fluss, it might make sense for the gateway to live as its own module depending on fluss-rust, rather than being tied to the 0.3.0 release train. That would also give the auth/authz questions Hongshun raised some room to be designed in a dedicated FIP. Just my personal thinking — happy to adjust based on what others feel is right. Best regards, Junbo Wang > 2026年7月2日 11:47，Forward Xu <[email protected]> 写道： > > Hi Anton, > > Thanks for putting this together, and thanks to everyone who landed the > previous roadmap. Framing the next phase around "fluss-rust as a > first-class analytical query surface" makes a lot of sense to me — it > builds naturally on the analytical primitives we just shipped. A few > thoughts, roughly following your structure. > > *Theme 1 — DataFusion integration adapter [MUST-HAVE]* +1. I strongly agree > it should be an *adapter over the Rust core*, not a bundled engine. Keeping > fluss-datafusion as a standalone crate exposing TableProvider + > CatalogProvider/SchemaProvider keeps the dependency direction clean and > lets any DataFusion-based consumer (or the gateway) opt in. Mapping > projection/filter/limit pushdown onto the existing access paths (PK > equality → lookup, bucket-key prefix → prefix lookup, LIMIT → bounded scan, > else log scan) is the right approach. One thing worth nailing down early is > the exactness contract we report back to DataFusion ( > TableProviderFilterPushDown::Exact vs Inexact), since getting that wrong > silently drops or double-applies filters. > > *Theme 2 — Lake + log union read (Paimon-first) [MUST-HAVE]* +1, and > Paimon-first is the right call — reusing paimon-rust by wrapping an > existing reader is low-risk, and the lake-snapshot-records-log-offset > design gives us a clean stitch point. Starting bounded-batch and deferring > streaming also de-risks correctness of the log-offset boundary before we > add continuous reads. The main correctness edge I'd like to see covered is > the boundary semantics for PK tables when applying newer log records over > the lake snapshot (dedup/ordering at the tiered offset). > > *Theme 3 — Foundations / server-side filter pushdown [STRETCH]* Agree this > is a stretch, but it's the item with the highest leverage since it prunes > both lake and log scans and unblocks the gateway's filter pushdown. > Cross-checking against the Java PredicateConverter semantics is important > so both clients behave identically. If capacity allows, I'd lean toward > pulling at least the simple value-predicate cases forward, since Themes 1 > and 4 both benefit. > > *Theme 4 — Gateway [STRETCH]* The "REST is the no-regret floor, SQL is the > ceiling deferred until Themes 1–2" framing is a good way to think about it. > Keeping it a thin, stateless HTTP frontend over the existing client > primitives (one endpoint per primitive) sounds right and keeps scope > contained. > > *Open questions* > > - *Themes 1–2 must-have, rest stretch — reasonable?* Yes, ordering looks > right to me. Nothing obviously missing. > - *Paimon first, Iceberg later?* Agree, provided the union-read path is > written against a lake abstraction so Iceberg/Lance slot in the same way > rather than requiring a rewrite. > - *Pin one arrow/DataFusion version across core, union-read/lake, and > DF?* Strong +1 — we should pin a single Arrow/DataFusion version > workspace-wide; version skew between these crates is a common and painful > source of breakage. > - *Separate crates: lake kernel, DataFusion adapter, gateway?* +1 to > separate crates. It matches the "adapter, not bundled engine" principle and > keeps the DataFusion dependency out of the core for consumers that don't > need it. > > One meta-question: do we have a rough sense of the target release (0.3.0?) > these must-haves land in, so we can size the Theme 3/4 stretch work > accordingly? > > Thanks again — happy to help on the DataFusion adapter side. > > Best, > > ForwardXu > > Junbo Wang <[email protected]> 于2026年7月1日周三 22:41写道： > >>> I'd also like us to focus on closing the gap with the Java client on >> predicate pushdown. >> Agreed — predicate pushdown is a great optimization, and the read path >> will benefit significantly from it. >> >> >> >> One thing that was missing in upstream paimon-rust was an >> object-store-safe existence check in the filesystem catalog. It checked >> whether database/table directories “exist”, which works on local >> filesystems but can fail on S3/OSS because those directories are just >> prefixes, not real objects. I forked it to fix that, so lake reads can >> reliably open Paimon tables from object storage. >> >> Thanks again, Anton, for putting this roadmap together! I think we could >> share it at the July 3rd community meeting. I've also been working on some >> designs around lake kernel read and the Fluss Gateway REST API — happy to >> discuss those as well if time allows. >> >> >> Best regards, >> Junbo Wang >> >>> 2026年6月25日 07:23，Anton Borisov <[email protected]> 写道： >>> >>> Hi Yuxia and Junbo, >>> >>> Yuxia, on PK changelog reads across the bindings: I have already >>> prepared record-mode CDC for Rust, Python and C++. It's small, and >>> the PK union read in Theme 2 needs it anyway. >>> >>> Junbo, I looked through your prototype branch. It is clear that a lot >>> of Themes 1, 2 and 4 already exist there as working code and it does >>> seem to answer two of the open questions: the three-way crate split >>> (lake kernel / adapter / gateway), and pinning a single Arrow/DataFusion >>> version at the workspace level. +1 to both. >>> >>> Two separate FIPs with detailed design for REST/SQL gateway parts - >> makes sense. >>> >>> One thing I noticed and want to discuss: >>> The "lake read" currently depends on a fork of paimon-rust, can you share >>> what was missing in paimon-rust? >>> >>> I'd also like us to focus on closing the gap with the Java client on >>> predicate pushdown. >>> In the interim we can push predicates into the lake (Paimon already >>> supports it) and apply them as a filter pass over the real-time log >>> tail. >>> >>> Good to hear you want to drive this forward, given the prototype, that >>> makes sense to me. I am happy to review the design and following PRs. >>> >>> Looking forward to the write-up with your findings. >>> >>> -- Anton >>> >>> вт, 23 июн. 2026 г. в 15:10, Junbo Wang <[email protected]>: >>>> >>>> Thanks Anton for kicking this off — funnily enough, I'd been arriving >> at very similar conclusions while prototyping on my side, so let me share a >> bit of what came out of those experiments. >>>> >>>>> THEME 1: DATAFUSION INTEGRATION ADAPTER [MUST-HAVE] >>>> >>>> Agree. Having the DataFusion integration live in fluss-rust would make >> it reusable across more downstream projects. One thing we might want to >> keep in mind: it would probably be worth aligning the DataFusion and Arrow >> versions with pg-datafusion so the whole stack stays on a consistent set of >> versions. >>>> >>>>> THEME 2: LAKE + LOG UNION READ (PAIMON-FIRST) [MUST-HAVE] >>>> >>>> Agree. I'd lean toward putting this in a separate fluss-lake module >> that exposes a client capable of doing the union read over the lake and the >> Fluss log. That way the lake-side dependencies stay isolated in their own >> module and don't bleed into the core fluss-rust code. One thing that might >> be worth flagging here: the lake side already supports predicate pushdown, >> while Fluss doesn't yet — so we'd probably think about how the union read >> handles that asymmetry in the interim. >>>> >>>> THEME 4: GATEWAY [STRETCH] — A thin HTTP frontend over the client >>>> >>>> I wonder if we could expand on this one a bit. My suggestion would be >> to split the SQL REST surface and the plain REST API into two separate >> FIPs. A read/write REST API on its own — somewhat in the spirit of >> kafka-rest, covering reads, writes, and metadata management — would already >> be a quick win for broadening Fluss's data ingestion/access reach, and it >> doesn't depend on anything else. The SQL REST surface, on the other hand, >> has to wait on Themes 1–2, so it might make sense to split them and ship >> the plain REST API first. >>>> >>>> >>>> >>>> Also, it might make sense to have the gateway as a separate crate >> inside the Fluss project itself — especially since fluss-rust will be >> moving into the Fluss repo anyway. >>>> >>>> On a personal note, this is something I'm really excited about and >> would love to help push forward. I'll share a write-up of what came out of >> my demo experiments with the community over the next few weeks. >>>> >>>> >>>> Best regards, >>>> Junbo Wang >>>> >>>>> 2026年6月23日 20:55，Yuxia Luo <[email protected]> 写道： >>>>> >>>>> Thanks for putting this together. +1 on the overall framing. My >> answers to the open questions, plus one addition: >>>>> >>>>> Open questions: >>>>> - Themes 1-2 as must-have, rest stretch: agree, the ordering makes >> sense to me. >>>>> - Paimon first, Iceberg later: +1. Reusing paimon-rust is the >> pragmatic path, and the union-read design generalizes to Iceberg the same >> way later. >>>>> - Pin one arrow/DataFusion version across core, union-read/lake, and >> DF: yes, we should pin a single version. Letting these drift will cause >> painful Arrow type/ABI mismatches across the crate boundary, so a shared >> pinned version (bumped deliberately, not per-crate) is worth the discipline. >>>>> - Separate crates (lake kernel, DataFusion adapter, gateway): +1 on >> separate crates. It keeps the dependency surface clean - the adapter and >> gateway are optional consumers, and the lake kernel shouldn't drag >> DataFusion into anyone who only needs the core read path. >>>>> >>>>> One addition I'd like to propose for the scope: >>>>> PK-table CHANGELOG read, and expose it across the bindings - >> specifically C++ and Python - so non-Rust CDC consumers can subscribe. >>>>> >>>>> On 2026/06/22 14:49:52 Anton Borisov wrote: >>>>>> Hi all, >>>>>> >>>>>> With the previous roadmap wrapping up - complex types >>>>>> (Array/Row/Map/nesting), limit and prefix scan, schema-aware KV >>>>>> decoding, the metrics framework, and write optimisations, thanks to >>>>>> everyone who contributed and reviewed, I'd like to open discussion on >>>>>> the roadmap. >>>>>> >>>>>> I propose framing around a single goal: make fluss-rust a first-class >>>>>> analytical query surface - a Rust-native SQL/DataFrame path over Fluss >>>>>> (DataFusion, and through it Polars/DuckDB/gateways), reading the lake >>>>>> tier at scale, built on the analytical primitives we just landed. >>>>>> >>>>>> As before, I grouped the items into themes with an initial must-have / >>>>>> stretch positioning. Please push back where you disagree. >>>>>> >>>>>> THEME 1: DATAFUSION INTEGRATION ADAPTER [MUST-HAVE] >>>>>> A standalone fluss-datafusion crate exposing TableProvider + Catalog >>>>>> over the existing client, so SELECT ... FROM . works from any >>>>>> DataFusion-based engine. It should be framed as an integration adapter >>>>>> over the Rust core, not a bundled engine. A gateway (FIP-32) or any >>>>>> analytical consumer can use it or not. >>>>>> >>>>>> Scope: TableProvider + CatalogProvider/SchemaProvider backed by the >>>>>> metadata path. >>>>>> >>>>>> Map DataFusion's projection/filter/limit pushdown onto the client's >>>>>> existing access paths: a full primary-key equality becomes a lookup, a >>>>>> bucket-key prefix becomes a prefix lookup, LIMIT becomes a bounded >>>>>> scan, otherwise a log scan. Filters an access path fully satisfies are >>>>>> reported exact, the rest are left for DataFusion to apply (server-side >>>>>> filter pushdown is Theme 3). >>>>>> >>>>>> THEME 2: LAKE + LOG UNION READ (PAIMON-FIRST) [MUST-HAVE] >>>>>> A client-side read that presents a lake-enabled table as one table. >>>>>> Its history is tiered to the lake (Paimon) and recent writes are still >>>>>> in the Fluss log, the lake snapshot records the log offset it covers, >>>>>> so we read the lake up to that offset and the Fluss log past it, then >>>>>> combine. >>>>>> Scope: >>>>>> - Per bucket: the lake snapshot plus the Fluss log past the tiered >> offset. >>>>>> - Log tables: concatenate the two. PK tables: apply the newer log >>>>>> records to get current state. >>>>>> - Paimon first, reusing paimon-rust, Iceberg/Lance later the same >>>>>> way. The DataFusion adapter reads lake-enabled tables through this >>>>>> path. >>>>>> - Bounded batch first, streaming later >>>>>> >>>>>> Reusing paimon-rust means wrapping an existing reader, so it's rather >>>>>> feasible and straightforward. >>>>>> >>>>>> THEME 3: FOUNDATIONS [STRETCH] >>>>>> Server-side value-predicate (filter) pushdown. The proto field exists, >>>>>> but the client currently sends none. This prunes lake/log scans and >>>>>> unblocks the gateway's filter pushdown (the open question I raised for >>>>>> 0.2.0), cross-checked against the Java PredicateConverter semantics. >>>>>> >>>>>> THEME 4: GATEWAY [STRETCH] >>>>>> A thin HTTP frontend over the client we already have independent of >>>>>> Themes 1–3. It's how non-Rust/non-SQL clients (TS/JS, microservices, >>>>>> curl) write to Fluss and do simple key access. >>>>>> - Write: upsert/append, delete, create/drop tables, metadata. The >>>>>> mature client write path essentially. >>>>>> - Read: point lookup, prefix lookup, bounded log / CDC poll - one >>>>>> endpoint per existing primitive, bounded and stateless. >>>>>> The tradeoff: no joins, filters, or aggregations - that's the SQL >>>>>> surface, and it needs the adapter and probably will be an ongoing >>>>>> effort to optimise later. >>>>>> - If Themes 1–2 ship, we unlock SQL frontends on top - PostgreSQL >>>>>> via datafusion-postgres, as one example >>>>>> In short: REST is the no-regret floor, SQL is the ceiling deferred >>>>>> until Themes 1–2. >>>>>> >>>>>> PARALLEL TRACK: ELIXIR + BINDING EXPOSURE >>>>>> Continue Elixir parity and expose newly-landed primitives (prefix >>>>>> lookup, etc.) across Python / C++ / Elixir. >>>>>> >>>>>> GOVERNANCE TRACK >>>>>> FIP-40: Consolidate apache/fluss-rust into apache/fluss - PR is ready, >>>>>> looks pretty good, waiting for a good moment to rebase/merge. >>>>>> >>>>>> OPEN QUESTIONS >>>>>> - Themes 1–2 must-have, the rest stretch. Reasonable? Anything >>>>>> missing or mis-ordered? >>>>>> - Paimon first, Iceberg later? >>>>>> - Pin one arrow/DataFusion version across core, union-read/lake, and >> DF? >>>>>> - Separate crates: lake kernel, DataFusion adapter, gateway? >>>>>> >>>>>> Looking forward to feedback. >>>>>> >>>>>> -- Anton >>>>>> >>>> >> >>
