Skip to content

The local platform v0.2

Status: available in v0.2 (milestone M1)

v0.1 defined the documents that the platform accepts (ModelSpec, ScoreRequest). It also defined the error codes that the platform returns and the executor protocol that it controls. v0.2 puts an HTTP API, a job queue, a sandboxed runner, a local object store and the Expert web app around them. All of it runs on one machine. Start with the quickstart.

What changed for you

Each thing that you do in a notebook is also an API call. The work runs in a sandbox, not in your process:

In your process (v0.1) On the platform (v0.2)
validate_spec(spec, catalog, schema) POST /v1/specs:validate, all problems at one time with JSON pointers, and the live squiggles in the editor
fit_spec(spec, data, ...) POST /v1/runs with kind: fit. Progress comes as server-sent events.
score_bundle(bundle, request, data) POST /v1/models/{id}/versions/{v}:score with Prefer: wait
A bundle directory A model version in a registry, with lineage, metrics, a scorecard, plots and a signed manifest
InprocExecutor local-process and docker executors behind the same protocol 
flowchart LR
    C["Web app · CLI ·<br/>Python client · curl"] -->|"/v1 JSON"| API["API<br/>(validates from static JSON)"]
    API --> Q[("PostgreSQL<br/>runs · queue · events")]
    Q --> D["Dispatcher"]
    D --> X["Executor<br/>(local process / Docker)"]
    X --> R["Runner<br/>supervisor + worker"]
    R -->|"outputs + manifest"| S[("GCS shim")]
    R -->|"signed events"| API

Design commitments, kept

  • The API never imports plugin code and never decodes your arrays. It validates specs and score requests from the exported catalog JSON and stored metadata. An import-linter contract keeps numpy, xarray and all plugins out of the server. The build of the server image checks this contract against the installed packages.
  • Runs are idempotent and observable. Submissions have idempotency keys. The status from the executor is the source of truth. Events control progress bars. They never decide if a run is alive.
  • Errors are RFC 7807 problems with the same stable codes that you see in Python.
  • OpenAPI 3.1 describes each endpoint. CI checks the generated clients for drift.
  • Outputs are verified. The platform checks the outputs of a run against the checksums and generations of the object store. Then it copies the outputs to a location where runs cannot write, and signs them.

Where to go next