Hi Yi, Thanks for this impressive proposal. Adding an RPC service deployment primitive can greatly expand Flink's capability scope and flexibility, with clear benefits for AI workloads and agent applications. It can also play a role in scenarios like state management or data lake compaction — relevant to but decoupled from the data pipeline. This is a very promising direction. That said, I still have some questions.
1. My biggest concern is heterogeneous resources, echoing Hongshun's Q4. Flink today uses homogeneous TM resources sliced for different task requirements, but decoupling GPU/CPU resources and enabling flexible independent scaling is one of the core values of this FLIP — and the current resource allocation approach cannot support this. We likely need new implementations for the resource allocation component. 2. Some components are described only at the idea level without implementation details. Hopefully these can be supplemented later or discussed in dedicated threads. A few worth discussing: - RpcTask: It's currently described as a simple single-threaded loop, with no plan to use Mailbox or OperatorChain. But if we want to support async, checkpoint, etc. later, introducing Mailbox would be valuable — handle the input queue as defaultAction, with async/CP/system messages as mails. This keeps semantics consistent with StreamTask and reduces maintenance burden. System events like FO-triggered request cancellation can also be prioritized via mailbox. OperatorChain is what SubtaskCheckpointCoordinator directly contacts, so having RpcOperator as the only node in the chain wouldn't add much complexity and lowers the cost when CP support comes. - Communication implementation: Will it reuse the Pekko + proxy mechanism? Pekko has risks with large message bodies (OOM, silently drop oversized frame, etc.). Since ROS is user-facing, we can't fully avoid this, and large-model scenarios often carry sizable context — needs careful handling. Also, the current design doesn't include table-function-style interfaces for unfolding data, so a large unfolded result has to be returned as a single response, inflating per-request payload size and amplifying the same risk. - Client-side implementation: Only the interface is described, no implementation details. Checkpoint is the prominent issue — since ROS is stateless by design, state recording falls to the client. Do we go with drain or record? Drain may take a long time; record on a keyed stream may cause request distribution confusion after parallelism changes. 3. ROS as container, not service: Is this allowed? If the system itself has service capabilities — especially non-Java stacks (agents, LLMs, also Python remote functions) — forwarding requests through the Flink Java layer adds unnecessary overhead. In that case RpcOperator would only handle lifecycle, while the framework manages resource scheduling and service discovery without carrying business traffic. nit: addRpcService is inconsistent with other places that use RpcOperator / the ros keyword. Worth unifying. Best, Gen On Mon, Jun 1, 2026 at 2:02 PM Yi Zhang <[email protected]> wrote: > Hi Guowei, > > > Thanks for the thoughtful review and valuable feedback! Your questions are > insightful > and have directly driven refinements to the proposal. Here are my > responses to each > point: > > > 1. Autoscaling > Supporting a (min, max) concurrency range is a great idea. However, > auto-scaling > involves much more than just the API surface — it requires discussion > around > scaling policies, coordination with the scheduler, metric-driven triggers, > and more. > To keep this FLIP focused, I'd like to defer that to a follow-up. Fixed > parallelism is > a common and necessary capability on its own, and the first version will > support > that. The interfaces can be extended later to include auto-scaling support > without > breaking changes. > > > 2. Async invocation & end-to-end examples > This is a good point. The `RosClient` already provides an async interface, > so using > it with `AsyncDataStream` is indeed more natural and idiomatic. I will add > an > asynchronous invocation example in the proposal. Regarding a complete > end-to-end > inference example: at the Java API level, it will still be primarily > conceptual, since > practical model inference examples really need Python API support (e.g., > loading a > PyTorch or ONNX model). I will refine the existing example to make the > overall > invocation chain as close to a real-world scenario as possible. > > > 3. GPU as a resource > This is indeed a key concern. The `extendedResource` API in the original > proposal > comes from Flink's existing resource model (introduced by FLIP-108). > However, I > agree that GPU deserves first-class treatment rather than being accessed > through a > generic extension mechanism. Promoting GPU to a first-class resource will > involve > changes to both the API and the scheduling layer, which we plan to address > in a > dedicated FLIP. For now, I will remove the `extendedResource` API from > `RosDescriptor` to keep this FLIP focused on the RPC service primitive > itself. The > resource declaration API can be extended once the GPU-specific FLIP is in > place. > > > 4. Serial method invocation > Agreed. I will make the serial invocation semantics more prominent — both > in the > FLIP documentation and in the code-level JavaDoc for `RpcOperator`. > > > 5. Why synchronous methods? > This is a key question. The initial design chose synchronous methods for > simplicity > — users don't need to worry about thread safety, and concurrency is > achieved by > scaling out instances (parallelism). However, I recognize that > asynchronous > processing is a critical requirement for certain scenarios (e.g., when the > RPC service > performs expensive operations that can be parallelized internally). I am > planning to > extend the current design to support asynchronous method execution: if a > user's > method returns a `CompletableFuture`, the framework will recognize it and > wait for > the future to complete before sending the response, without blocking the > invoke > thread. This allows the invoke thread to continue dispatching subsequent > requests, > while the user controls internal concurrency. Importantly, > `processRequest` calls > remain serial — the async behavior is purely opt-in. > > > 6. SQL UDF usage > The current design is primarily oriented toward the DataStream API, as it > provides > the foundational building blocks. I agree that SQL and Table API support > is an > important aspect to consider as well. On the invocation side, both Table > API and > SQL UDFs share the same runtime path through `FunctionContext`, so > enabling > RPC service access from UDFs is straightforward — I will add a > `getRosClient` > method to `FunctionContext` and include that change in this FLIP. > The real challenge for SQL lies on the definition side: how to define and > register > an RPC service using pure SQL syntax, which would require SQL grammar > extensions and planner changes. I think that part would be better > addressed in > a dedicated follow-up FLIP to give it the attention it deserves. > > > 7. Service sharing across jobs & lifecycle > In this FLIP, the RPC service is defined as part of the job (via > `env.addRpcService()` > before `env.execute()`). It cannot be accessed across jobs, and it is > released when > the job reaches a terminal state. I believe your question is more relevant > to the > "Shared RPC Service" mentioned in the Follow-up Tasks section. While it's > out of > scope for this FLIP, let me briefly describe the envisioned design: a > shared RPC > service would run as an independent job within the same Application, > accessible by > all jobs in that Application. It would terminate when the Application > ends, or could be > stopped manually via a user-facing API. > The embedded (per-job) service model discussed in this FLIP is simpler to > manage > and well-suited for many streaming scenarios, which is why we focus on it > first. > Importantly, both the embedded and shared modes will reuse the same set of > deployment primitives (`RpcOperator`, `RosClient`, `RosDescriptor`, etc.) > — the > difference is only in lifecycle management and visibility scope. > > > > > I hope the above addresses your questions. I will update the proposal to > reflect the > changes mentioned above. Please feel free to follow up if you have any > further > questions or suggestions. > > > Best, > Yi > > > > At 2026-05-28 16:39:49, "Guowei Ma" <[email protected]> wrote: > >HI，YI > > > >One more question I'd like to raise is around how the service behaves when > >it's referenced by more than one job. If two jobs are submitted within the > >same application and both reference the same RpcOperator service, is that > >service shared across them, and what is the expected behavior in that > case? > >I'm also curious how this plays out for bounded (batch) jobs in > particular: > >if one job completes while the other is still running, what does the > >service lifecycle look like — is the service torn down as soon as the > first > >job finishes, and if so, how does that affect the job that's still > running? > > > >Best, > >Guowei > > > > > >On Thu, May 28, 2026 at 2:55 PM Guowei Ma <[email protected]> wrote: > > > >> Hi,YI > >> > >> Thanks for putting together this FLIP — it's a well-thought-out > proposal, > >> and the RpcOperator Service is a really valuable addition for > AI/inference > >> workloads on Flink. I have a few questions and suggestions: > >> > >> > >> 1. Autoscaling > >> How is scale-up/scale-down expressed in this design? Could you consider > >> allowing concurrency to be configured as a (min, max) range? Currently > it > >> appears only min = max (i.e. fixed parallelism) is supported. > >> > >> > >> 2. Async invocation & end-to-end examples > >> The example shows a synchronous invocation style. I'd suggest also > >> providing an asynchronous invocation example — or even just a scaffold > that > >> exposes only the RpcOperator's handle and method. In addition, it would > be > >> best to include a complete end-to-end example of invoking a model (CPU > >> inference would be fine). > >> > >> > >> 3. GPU as a resource > >> GPU is surely a common/first-class resource. Accessing it here through > an > >> extend mechanism — wouldn't that mislead users? It would seem more > >> natural to declare GPU as a standard resource rather than going through > an > >> extension. > >> > >> > >> 4. Serial method invocation > >> Each method of an RpcOperator is invoked serially. I'd suggest stating > >> this explicitly in both the documentation and the code comments, so that > >> users — or an AI — can be clearly aware of it. > >> > >> > >> 5. Why synchronous methods? > >> Why is every method in RpcOperator a synchronous call? Does this imply > >> that no model inference supports concurrent invocation? > >> > >> > >> 6. SQL UDF usage > >> How would SQL UDF users invoke this capability? I'd suggest providing a > >> complete example for this as well. > >> > >> > >> Thanks again for the great work! > >> > >> Best, > >> Guowei > >> > >> > >> On Wed, May 27, 2026 at 8:36 PM Charles Zhang <[email protected]> > >> wrote: > >> > >>> Hi Yi, and Flink Community, > >>> > >>> Thanks for bringing up this excellent proposal. I am fully in favor of > >>> FLIP-582. > >>> > >>> In our production workloads, especially regarding large-scale AI > >>> inference, > >>> the tight coupling of the data plane and compute units has always been > a > >>> major pain point. If an inference subtask fails due to a GPU OOM or > driver > >>> issue, triggering a global rollback is incredibly expensive since model > >>> reloading takes minutes. > >>> > >>> The introduction of the RpcOperator Service as a first-class primitive > >>> masterfully decouples the heavy inference tasks from the mainstream > >>> topology. The fault isolation, independent scaling, and stateless > design > >>> perfectly match the requirements of modern AI-oriented data processing. > >>> > >>> This is a clean and robust architecture. Looking forward to seeing this > >>> merged! > >>> > >>> > >>> Best wishes, > >>> Charles Zhang > >>> from Apache InLong > >>> > >>> > >>> Yi Zhang <[email protected]> 于2026年5月27日周三 14:12写道： > >>> > >>> > Hi everyone, > >>> > > >>> > > >>> > > >>> > I would like to start a discussion on FLIP-582: Support RpcOperator > >>> > Service [1]. > >>> > > >>> > > >>> > AI-oriented workloads like multimodal data processing and model > >>> inference > >>> > are > >>> > growing rapidly in recent years. These workloads are characterized by > >>> > expensive > >>> > resources (GPUs) and high initialization costs (seconds to minutes > for > >>> > model > >>> > loading). In today's Flink, embedding them in the data plane couples > >>> their > >>> > parallelism and failover with surrounding operators; deploying them > as > >>> > external > >>> > services disconnects their lifecycle from the job and doubles > >>> operational > >>> > overhead. > >>> > > >>> > > >>> > This FLIP introduces RpcOperator Service — a framework-level > primitive > >>> > that runs > >>> > user-defined compute as RPC services in an independent Pipelined > Region > >>> > within > >>> > the Flink job. Because the service is isolated at the scheduling > level, > >>> it > >>> > can achieve > >>> > fault isolation, independent scaling, and dedicated resource > allocation. > >>> > As a native > >>> > Flink primitive, it also lays the foundation for automatic flow > control, > >>> > flexible load > >>> > balancing, and coordinated auto-scaling — all without introducing > >>> external > >>> > infrastructure or additional operational burden. > >>> > > >>> > > >>> > > >>> > > >>> > Looking forward to your feedback and suggestions! > >>> > > >>> > > >>> > > >>> > > >>> > [1] > >>> > > >>> > > >>> > https://cwiki.apache.org/confluence/display/FLINK/FLIP-582%3A+Support+RpcOperator+Service > >>> > > >>> > > >>> > > >>> > > >>> > > >>> > Best Regards, > >>> > Yi Zhang > >>> > >> >
