System landscape#
Tributo is a Ray-native ML Framework/SDK. It owns portable contracts for job submission, bounded ingestion and writing, algorithm execution, model bundles, inference, explainability, serving, and Lance vector-index operations. Ray 2.55.1 remains the distributed execution runtime for Jobs, Data, Train, Tune, Serve, tasks, actors, and scheduling.
Reading the landscape#
The main flow runs from top to bottom:
Users and applications use the CLI or Python API. Persisted configuration uses strict JSON, while
TributoClientwraps submission, status, logs, and stop operations over the Ray Jobs API.Scenario workloads cover training and tuning, custom batch inference, online serving, and streaming input. Their stability is not uniform: legacy trainers remain Beta compatibility paths, portable algorithm execution is Alpha, vector indexing and explainability are Alpha, and Kafka is an Alpha
StreamSourcerather than a complete service loop.The portable contract layer separates bounded ingestion, algorithm dispatch, and Bundle-based model delivery. Providers, engine bindings, exporters, validators, model importers, flavors, hooks, runtime adapters, and sinks extend these contracts without changing the core orchestration path.
Ray executes the resolved work. Tributo does not replace Ray’s scheduler or distributed runtime.
External systems#
Data systems, Bundle stores, MLflow, Lance namespaces, and Kafka remain outside the framework boundary:
Data systems supply bounded inputs. Support varies by source and engine; adapter presence alone is not a support claim.
Bundle stores persist immutable artifacts and manifests. Current publication supports local paths,
file://, and S3; HDFS Bundle storage is not implemented.MLflow is an optional model-import, tracking, provenance, and registry integration. It is not the source of truth for Bundle readability.
Lance stores vector data and index metadata. Tributo validates requests and receipts; Lance-Ray executes distributed index and maintenance tasks.
Kafka provides a fail-closed microbatch input protocol. Tributo does not ship a built-in Kafka-to-inference long-running service loop.
The support matrix is the authoritative view of verified, Alpha, Beta, adapter-only, and unsupported paths. The diagram groups architectural responsibilities; it does not promote every extension contract to a supported product capability.
Framework boundary#
Tributo is intentionally not a multi-tenant ML platform. Kubernetes control planes, custom scheduling, tenant isolation, quota and RBAC, approval and canary workflows, and a centralized operations UI are outside the current framework scope. See Product Scope for the complete set of goals, non-goals, and re-evaluation triggers.