Hi team 👋 I would really appreciate your review on a new feature series under FLINK-38825 – AI-friendly Async Batch Operator for high-latency inference workloads. This is an incremental PR series introducing batch-oriented async processing into Flink’s DataStream API. The goal is to provide a minimal, extensible foundation for ML inference and other high-latency remote services. PR Overview 1️⃣ #27355 – Introduce AsyncBatchFunction and AsyncBatchWaitOperator Adds AsyncBatchFunction (PublicEvolving API) Introduces AsyncBatchWaitOperator (unordered, size-based batch triggering) Adds AsyncDataStream.unorderedWaitBatch(...) entry point Fully backward-compatible and minimal scope 2️⃣ #27356 – Add time-based batch triggering Supports timeout-based flushing of incomplete batches Prevents long-tail latency under low throughput scenarios 3️⃣ #27357 – Add ordered async batch support Adds ordered output semantics Introduces OrderedAsyncBatchWaitOperator with sequence tracking 4️⃣ #27358 – Add inference-oriented metrics Batch size and latency histograms Async invocation duration metrics Inflight batch counters and failure metrics 5️⃣ #27359 – Implement retry and timeout strategies Fixed-delay & exponential backoff retry Configurable timeout handling (fail / allow partial) 6️⃣ #27360 – Add SQL/Table API integration Introduces AsyncBatchLookupFunction Integrates with lookup join execution 7️⃣ #27361 – Add Python DataStream API integration Python AsyncBatchFunction async_invoke_batch() support Python-side DataStream wiring Why this matters Improves throughput for ML/GPU/remote inference workloads Keeps API surface small and incrementally extensible Fully additive and backward-compatible Aligns with Flink’s async processing model If you have bandwidth this week, I’d really value feedback on API design, operator semantics, and extensibility considerations. Thanks a lot 🙏