400Ping opened a new issue, #1016: URL: https://github.com/apache/mahout/issues/1016
## Summary This roadmap introduces an **optional Ray/KubeRay integration layer** for Mahout QDP to enable **multi-node, multi-GPU (distributed data-parallel)** encoding. The integration supports both **Parquet-first** (Ray Data) and **Tensor-first** (NumPy/Torch batch) entry points, with **externalized outputs** (write sharded artifacts + manifest; return paths/metadata) to avoid moving huge tensors through Ray’s object store. ## Motivation - QDP is GPU-accelerated and already supports file-based inputs (e.g., Parquet) via Python bindings. - Users often have GPUs across multiple hosts; Ray provides proven scheduling, GPU resource management, and fault tolerance. - A real workload (QDP encoding) also creates upstream engineering opportunities in Ray Core/Data and KubeRay. ## Design principles - Ray integration is **optional** (extra dependency), not required for core QDP usage. - Use **actor-per-GPU** to reuse CUDA context/native extension initialization. - Prefer **external outputs** (write shards + manifest) over returning large tensors to the driver/object store. - Use **bounded in-flight** submission for backpressure and memory safety. ## Phase 0: Define the integration contract Deliverables: - Input shard contract (paths/metadata) and output shard contract. - A stable `manifest` schema capturing params, commit hash, shard outputs, and timings. ## Phase 1: Ray Core actor-per-GPU pipeline (single-node) Deliverables: - `QdpWorkerActor` (`num_gpus=1`) that initializes `QdpEngine(device_id=0)` once and encodes assigned shards. - Driver that uses bounded in-flight scheduling. - External output writing + manifest generation. ## Phase 2: Multi-node (2 hosts / 2 GPUs) scaling + fault tolerance Deliverables: - Run end-to-end on two hosts (each 1 GPU) and publish baseline throughput and scaling behavior. - Define and implement idempotent shard output (retry-safe) and a commit protocol. ## Phase 3: Ray Data integration (Parquet-first + Tensor-first) Deliverables: - Parquet-first: `read_parquet -> partition/shard -> GPU actors -> write shards -> refs dataset`. - Tensor-first: `map_batches` / actor interface for NumPy/Torch batches using `encode_batch`. - Benchmarks comparing task-per-batch vs actor reuse. ## Phase 4: KubeRay operationalization Deliverables: - Reference deployment (RayJob/RayService) with GPUs and native extension packaging strategy. - Shared storage setup for shard outputs. - (Optional) E2E validation test. ## Phase 5: Connect to QDP feature roadmap Deliverables: - Integrate improvements such as IQP streaming from Parquet and additional readers into the Ray/KubeRay pipeline. - Re-run baselines to quantify improvements. ## Success criteria (initial) - 2-node / 2-GPU pipeline runs reliably and produces correct shard outputs + manifest. - Throughput improves vs 1-node baseline with reasonable scaling. - Retry does not corrupt outputs (idempotent / commit-safe). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
