Distributed Execution

About 170 wordsLess than 1 minute

RayOrch uses Ray rather than replacing it:

Ray cluster
  ├─ node placement and resource accounting
  ├─ actor lifecycle and RPC transport
  └─ runtime environments

RayOrch
  ├─ compile Pipeline dependencies
  ├─ track cardinality and lineage
  ├─ decide which Grain is READY
  ├─ form per-Call execution microbatches
  └─ reconstruct results and reports

Connect to an existing cluster with address="auto" in pipeline.run() or Executor. The equivalent functional form rayorch.run(pipeline, ...) remains available. Ray places each persistent actor according to the options declared on its Call.

Multi-node contract

For a workload to run correctly across nodes:

every eligible environment can import compatible RayOrch and workload code;
model, data, output, and artifact paths are visible consistently, normally through shared storage;
each stage declares honest CPU, GPU, memory, and custom-resource needs;
network services and credentials needed by UDFs are reachable from their actor nodes;
the Ray head and worker nodes use compatible Ray and Python versions.

RayOrch does not copy large datasets or model weights as part of graph execution.