Hi Prajwal, This is a great proposal. I have been creating end-to-end real-time profile use cases with Fluss features, and I had to create a few UDFs to better interact with BM. The FIP solves this structural problem as it makes BITMAP a first-class type, promotes those same operations into the catalog so no UDF jar is needed, and adds server-side pushdown so BITMAP_OR_AGG queries avoid materializing per-row data in Flink entirely.
The overall direction is solid, the three-layer decomposition is clean, and the backward compatibility story is well thought out. I have a few points I would like to discuss before moving forward, mostly around design decisions that I think are worth aligning on at the proposal stage. 1. DataTypeVisitor is @PublicStable but im wondering whether this can be a breaking change. Adding a new method to this interface would likely break any external implementors. It would be great if the proposal could settle on an approach: a default fallback method, an abstract base class users can extend, or a deferral to the next major version. Any of these would work. 2. Null semantics are a bit contradicted within the proposal itself. The public interface section states BITMAP_CARDINALITY(NULL) → NULL, while the reference implementation section returns 0 for null input. Could you clarify which behavior is the intended one? 3. BITMAP_BUILD_AGG accepts only INT, while the server already supports rbm64 for BIGINT. Given that most real-world entity IDs exceed Integer.MAX_VALUE and the 64-bit aggregator already exists server-side, it would be worth clarifying whether the initial scope is intentionally limited to 32-bit and, if so, the reasoning behind that choice. 4. BITMAP_AND_AGG is included in the public interface but has no server-side aggregator and no pushdown support. I'm reluctant that exposing this in the public API without a working server-side counterpart could lead to incorrect results when users combine it with the merge engine. Would it be possible to add a clear section defining exactly when it is safe to use? Some food for thought on my side is whether it makes sense (based on my examples) to consider adding a function that converts a bitmap to a string so it's human-readable for users debugging. This can be For example, take a bitmap and output something like (bitmap_to_string??): Output format: "count=3 [1001, 1002, 1007]." This can be out of scope of this FIP, and more of a future improvement, if it resonates. I just wanted to bring this to your attention, just in case you think it makes sense, because it's something that helped me a lot with my examples. Looking forward to this Best, Giannis On Mon, Apr 6, 2026 at 9:12 AM Prajwal Banakar <[email protected]> wrote: > Hi everyone, > > I have created the formal Confluence page for FIP-37 [1]. > > As this is my first time creating a FIP page, I would greatly appreciate > any feedback or suggestions for improvement. The discussion will continue > on this thread. > > Looking forward to your thoughts. > [1] > > https://cwiki.apache.org/confluence/display/FLUSS/FIP-37%3A+Native+RoaringBitmap+Integration+for+Apache+Fluss > > Best regards, > Prajwal Banakar > > On Fri, 3 Apr 2026 at 19:22, Giannis Polyzos <[email protected]> > wrote: > > > Hi Prajwal, > > It’s probably a mistake that needs fixing. > > > > Feel free to use FIP-37 > > > > Best, > > Giannis > > > > On Fri, 3 Apr 2026 at 4:49 PM, Prajwal Banakar < > [email protected] > > > > > wrote: > > > > > Hi dev's, > > > > > > I am currently creating the Confluence page for this proposal. > > > > > > I noticed that FIP-35 is not currently listed on the wiki. Could you > > please > > > confirm if this number is available for use, or if I should assigned a > > > different one? > > > > > > Best regards, > > > Prajwal Banakar > > > > > > On Wed, 11 Mar 2026 at 22:56, Prajwal Banakar < > > [email protected]> > > > wrote: > > > > > > > Hi Keith, > > > > > > > > Thank you for the follow-up. > > > > > > > > You are correct that FieldRoaringBitmap64Agg already exists in > > > > fluss-server. I have updated the proposal accordingly. To clarify, > the > > > > 32-bit scope is intended to keep the initial type system and SQL > > function > > > > surface focused and deliverable, rather than being a limitation of > the > > > > aggregator itself. Since the server-side aggregator is already in > > place, > > > > RBM64 will be a natural, low-risk follow-on once the type system and > > > > pushdown infrastructure are established. > > > > > > > > I have also removed the misleading motivation paragraph as you > > suggested. > > > > The updated document is available at the same link. Additionally, I > > would > > > > welcome Yang's input on the alignment with FIP-21. > > > > > > > > Best regards, > > > > Prajwal Banakar > > > > > > > > On Wed, 11 Mar 2026 at 17:37, Keith Lee <[email protected] > > > > > > wrote: > > > > > > > >> Hi Prajwal, > > > >> > > > >> Thank you for addressing / answering the questions. > > > >> > > > >> > This proposal adds the missing bridge: a proper BITMAP DDL type, > SQL > > > >> functions (BITMAP_BUILD, BITMAP_OR_AGG, BITMAP_CARDINALITY), and > > > pushdown > > > >> via applyAggregates(). The storage-side aggregation logic already > > > exists; > > > >> this proposal makes it accessible end-to-end > > > >> > > > >> 1. That makes sense. I think the motivation section should lead with > > > that > > > >> and remove the following as it can be misleading given that rbm is > > > >> supported by aggregation merge engine: “users requiring > > high-cardinality > > > >> unique counting (e.g., UV analytics) must execute Client-Side > > > Aggregation. > > > >> The TabletServer is forced to send massive amounts of raw > > LogRecordBatch > > > >> rows over the network to a Flink cluster for evaluation. This > results > > in > > > >> unnecessary network transfer and prevents efficient utilization of > the > > > >> existing aggregation merge engine.” > > > >> > > > >> 2. That makes sense. Thank you for the context. > > > >> > > > >> 3. > > > >> > > > >> > RBM64 requires a fundamentally different internal structure; a map > > of > > > >> RBM32 chunks which increases implementation and serialization > > complexity > > > >> significantly. > > > >> > > > >> My understanding is that the proposal wires existing > > > >> FieldRoaringBitmap32Agg to support rbm32. FieldRoaringBitmap64Agg > > should > > > >> already exist and handle the complexity that you mentioned? > > > >> > > > >> Additionally, it might be good for Yang to review / provide input on > > > this > > > >> given his work on FIP-21. > > > >> > > > >> Best regards > > > >> > > > >> Keith Lee > > > >> > > > >> > > > >> On Wed, 11 Mar 2026 at 05:49, Prajwal Banakar < > > > [email protected] > > > >> > > > > >> wrote: > > > >> > > > >> > Hi Keith, thank you for the detailed feedback. > > > >> > > > > >> > 1. On motivation vs existing aggregation merge engine: The > > aggregation > > > >> > merge engine in 0.9 supports rbm32/rbm64 at the storage level, but > > > >> BITMAP > > > >> > is not yet a first-class type in the DDL or type system. Users > today > > > >> must > > > >> > declare the column as BYTES (as shown in the 0.9 release example: > > > >> uv_bitmap > > > >> > BYTES), and there are no SQL functions to build, merge, or query > > > bitmaps > > > >> > from Flink SQL. This proposal adds the missing bridge: a proper > > BITMAP > > > >> DDL > > > >> > type, SQL functions (BITMAP_BUILD, BITMAP_OR_AGG, > > BITMAP_CARDINALITY), > > > >> and > > > >> > pushdown via applyAggregates(). The storage-side aggregation logic > > > >> already > > > >> > exists; this proposal makes it accessible end-to-end. > > > >> > > > > >> > 2. On NULL semantics: BITMAP_OR(bitmap, NULL) returns NULL > following > > > >> > standard SQL scalar function semantics where NULL inputs propagate > > to > > > >> NULL > > > >> > outputs. BITMAP_OR_AGG follows aggregate function convention > > > consistent > > > >> > with how SUM and AVG behave, where NULLs in individual rows are > > > skipped > > > >> and > > > >> > only a fully NULL input set returns NULL. This distinction follows > > > >> FLIP-556 > > > >> > and StarRocks semantics. > > > >> > > > > >> > 3. On 32-bit scope: The proposal is scoped to 32-bit initially > > because > > > >> > RoaringBitmap32 covers integer values up to 2^32 (~4 billion), > which > > > is > > > >> > sufficient for most user ID and session ID use cases. RBM64 > > requires a > > > >> > fundamentally different internal structure; a map of RBM32 chunks > > > which > > > >> > increases implementation and serialization complexity > significantly. > > > >> > Starting with 32-bit keeps the initial scope focused and > > deliverable. > > > >> RBM64 > > > >> > support is listed as a Could-Have in the MoSCoW deliverables and > can > > > >> follow > > > >> > in a subsequent iteration. > > > >> > > > > >> > Best regards, > > > >> > > > > >> > Prajwal Banakar > > > >> > > > > >> > > > > >> > On Wed, 11 Mar 2026 at 01:34, Keith Lee < > > [email protected]> > > > >> > wrote: > > > >> > > > > >> > > Hello Prajwal, > > > >> > > > > > >> > > Thank you for the detailed proposal. I enjoyed reading it and > > have a > > > >> few > > > >> > > questions/comments. > > > >> > > > > > >> > > 1. On motivation, can you provide context on how this differs > with > > > >> > > aggregation merge engine’s roaring bitmap implementation [1]? > > > >> > Specifically, > > > >> > > motivation part states that “users requiring high cardinality > > unique > > > >> > > counting … must execute client-side aggregation”. Aggregation > > merge > > > >> > engine > > > >> > > performs aggregation on server-side. The motivation section > should > > > >> > clarify > > > >> > > how the proposed changes improve or complement aggregation merge > > > >> engine, > > > >> > > which seems to have been considered as Section 2 references > FIP-21 > > > >> > > Aggregation Merge Engine. Adding this context will help readers > > > >> > understand > > > >> > > the motivation of the proposal better. > > > >> > > > > > >> > > 2. Can you clarify the NULL semantics section specifically on > the > > > >> > decision > > > >> > > on why BITMAP_OR(bitmap, NULL) returns NULL but BITMAP_OR_AGG > only > > > >> > returns > > > >> > > null when all rows are NULL? > > > >> > > > > > >> > > 3. Why is the scope limited to 32 bit bitmaps? Adding the > > rationale > > > >> > behind > > > >> > > these e.g. how (if any) support of 64bit bitmaps would increase > > > >> > > implementation complexity. Articulating these may help other > > > >> contributors > > > >> > > understand the complexity and perhaps come up with suggestions > on > > > how > > > >> to > > > >> > > address them. > > > >> > > > > > >> > > Best regards > > > >> > > > > > >> > > Keith Lee > > > >> > > > > > >> > > [1] > > > >> > > > > > >> > > > > > >> > > > > >> > > > > > > https://fluss.apache.org/blog/releases/0.9/#2-storage-level-processing--semantics > > > >> > > > > > >> > > > > > >> > > On Mon, 9 Mar 2026 at 05:31, Prajwal Banakar < > > > >> [email protected] > > > >> > > > > > >> > > wrote: > > > >> > > > > > >> > > > Hi Devs, > > > >> > > > > > > >> > > > I have pushed a working prototype to my public fork > > demonstrating > > > >> the > > > >> > > > BitmapType integrated with FieldRoaringBitmap32Agg. This > > includes > > > >> four > > > >> > > > passing unit tests. > > > >> > > > > > > >> > > > The link to the prototype is available in the Google Doc, and > > you > > > >> can > > > >> > > also > > > >> > > > find it here: > > > >> > > > > > > >> > https://github.com/Prajwal-banakar/fluss/tree/RoaringBitmap-prototype > > > >> > > > > > > >> > > > The Google Doc link remains the same. I look forward to your > > > >> feedback. > > > >> > > > > > > >> > > > Best regards, > > > >> > > > > > > >> > > > Prajwal Banakar > > > >> > > > > > > >> > > > > > > >> > > > On Sun, 1 Mar, 2026, 11:49 am Prajwal Banakar, < > > > >> > > [email protected] > > > >> > > > > > > > >> > > > wrote: > > > >> > > > > > > >> > > > > Hi everyone, > > > >> > > > > > > > >> > > > > I would like to start a discussion on the proposal for > Native > > > >> Bitmap > > > >> > > > > Integration & Stateless Pushdown Aggregation. > > > >> > > > > > > > >> > > > > This proposal enables end-to-end native support for the > BITMAP > > > >> type > > > >> > in > > > >> > > > > Fluss and integrates it with the existing aggregation merge > > > >> engine to > > > >> > > > > support server-side bitmap union pushdown. The goal is to > > reduce > > > >> > > network > > > >> > > > > transfer and offload DISTINCT-style aggregation from Flink > to > > > the > > > >> > > > > TabletServer. > > > >> > > > > > > > >> > > > > Key highlights of the proposal include: > > > >> > > > > > > > >> > > > > - Type System: Promoting BITMAP to a first-class logical > type. > > > >> > > > > - UDF Suite: Introducing BITMAP_BUILD, BITMAP_OR_AGG, and > > > >> > > > > BITMAP_CARDINALITY (aligned with FLIP-556 and StarRocks > > > >> semantics). > > > >> > > > > - Optimizer: Planner-based pushdown via applyAggregates in > the > > > >> Flink > > > >> > > > > connector. > > > >> > > > > - Safety: No changes to LogRecordBatch or WAL, making this > > > >> strictly > > > >> > > > > additive and migration-free. > > > >> > > > > > > > >> > > > > You can find the full proposal document here: > > > >> > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > >> > > > > > > https://docs.google.com/document/d/1sDhfkmo-w-UTvo2n3rsY1lytSSryswfkI83cSdka8s0/edit?usp=sharing > > > >> > > > > > > > >> > > > > I would appreciate feedback on the public interfaces, > pushdown > > > >> > > > > constraints, and overall scope. > > > >> > > > > > > > >> > > > > Best regards, > > > >> > > > > Prajwal Banakar > > > >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > >> > > > > > > > > > >
