Skip to content

Project Anatomy — Storage, Pipelines & the Unified View

This guide covers the Project Anatomy layer (Increment 2 of Projects & Workspaces): the per-project MinIO storage location, the two pipeline surfaces (Prefect training + Ray Serve serving), and the assembled anatomy view that renders storage + connections + pipelines + models + members + quota + budget as one governable unit — in both exa project and the dashboard.

It builds on Projects & Workspaces (the project primitive, Named Connections, membership, and cost). Read that first if a project is new to you.

Mental model. The anatomy is a view over records, not a running system. A project's storage is a convention + a small binding record; its pipelines are an aggregation over existing Prefect deployments and Ray-served models. Nothing here launches containers on its own — the value is making a team's whole ML footprint named, visible, and consistent across the CLI and dashboard.

1. Per-project storage (P6)

Every project gets a stable MinIO location, so a team's pipeline artifacts, datasets, and cache no longer share one global bucket:

s3://examlops-projects/<project>/
    ├── artifacts/   # MLflow run artifacts (bound via a per-project MLflow experiment)
    ├── datasets/    # dataset snapshots for the project
    └── cache/       # scratch / intermediate
  • The bucket defaults to examlops-projects (override with EXAMLOPS_PROJECTS_BUCKET); the compose minio-init service creates it on startup.
  • The prefix is always <project>/, so the split is shared bucket, per-project prefix — no new IAM or bucket provisioning per project.
  • projects.storage_gb stays the quota; a small additive project_storage table records the binding and observed used_bytes.
  • Training routes into the project prefix through a per-project MLflow experiment (project/<name>) whose artifact_location points at .../artifacts/. When a model belongs to a project, exa pipeline run and the pipeline generator set that experiment automatically (fail-open: if project resolution or MinIO is unavailable, training falls back to the default experiment).

Show, bind, refresh

exa project storage climate-team                     # show the location, used GB, and any bound connection
exa project storage climate-team --refresh           # re-probe used bytes from MinIO (fail-open)
exa project storage climate-team --bind-connection raw-s3   # point storage at a P2 S3 connection

Binding a Named Connection sets the project's bucket from that connection's config and records connection_ref — the credentials stay in D7 secrets and are never copied into project_storage. Use this when a project should live in a dedicated bucket rather than the shared examlops-projects one.

exa project storage <name> --json (or the global -o json|yaml|csv) emits the raw record.

2. The two pipeline surfaces (P7)

A project exposes exactly two pipeline surfaces, each an aggregation over state that already exists — no heavyweight per-project pipeline objects are created:

Surface What it is Derived from
Prefect (training) The project's training deployments, schedule, last run, status Deployments tagged project:<name> (from deploy.py), plus the optional project_pipelines registry row
Ray Serve (serving) The project's served models and their traffic split Models where resolve_project(model) == <name>, joined with traffic_rules 
exa project pipelines climate-team

Prefect pipeline (training)
  Deployments   examlops-jpcp
  Schedule      0 2 * * *
  Last run      2026-07-15 02:00
  Status        healthy

Ray Serve pipeline (serving)
  Model   Traffic split
  JPCP    {'Production': 90, 'Canary': 10}

An empty project (no models assigned yet) shows both surfaces as absent — assign a model first:

exa project assign climate-team JPCP --kind model

The optional project_pipelines table is a light registry/cache keyed by (project, kind) where kind ∈ prefect | rayserve — so a project holds one Prefect surface and one Ray Serve surface, by primary key. The training surface writes to the project's P6 storage; the serving surface serves the project's promoted models.

3. The assembled anatomy view (P8)

exa project show <name> renders the whole workspace in one pane — storage, connections, pipelines, models, members, quota, and budget — reading the same get_project_full() model the dashboard uses, so the CLI and dashboard say the exact same words.

exa project show climate-team

The view adds three panels on top of the Increment-1 anatomy:

  • Storage — location, used GB against quota, bound connection (P6).
  • Connections — each project connection by name and kind, with a has_secret flag only — the secret value and its secret_ref are never transported (P2).
  • Pipelines — the Prefect and Ray Serve surfaces from §2 (P7).

Every source is fail-open: if MinIO, Prefect, Ray, or the connections layer is unavailable, the corresponding panel degrades to empty rather than failing the whole view.

In the dashboard

The Projects console → project detail page shows the same anatomy: a Storage card with a usage bar, a Pipelines card with the Prefect and Ray Serve surfaces, alongside the existing Models, Members, Quota, and Budget cards. GET /api/v1/projects/<name> returns the extended anatomy (storage, connections, pipelines), viewer-gated and secret-safe.

4. Isolation model

A project is a governable boundary, honest about enforcement depth on the Docker/MinIO/SQLite substrate:

Boundary Mechanism Enforcement
Network network_name = examlops-<name> (per-project docker bridge) advisory (compose)
Storage per-project MinIO prefix (P6) soft (shared bucket); hard via --bind-connection → dedicated bucket
Connections connections PK (project, name); credentials in D7 secrets scoped, secret-safe
People / RBAC authz_relations owner ⊇ editor ⊇ viewer on project:<name> checked (default-deny under EXAMLOPS_MULTITENANCY)
Compute quota project CPU/mem/GPU/storage limits; exa project compose advisory (compose)
Cost model_costs.project attributed
Audit audit_events with the project as target recorded

Hard enforcement (kernel-level network firewalling, per-bucket IAM, cluster-level queue admission) is a future live-runtime increment; the anatomy layer delivers the named, visible, consistent boundary.

5. Command summary

Command Purpose
exa project show <name> Full anatomy: storage · connections · pipelines · models · members · quota · budget
exa project storage <name> Show the MinIO storage location + usage
exa project storage <name> --bind-connection <conn> Point storage at a P2 S3 connection
exa project storage <name> --refresh Re-probe used bytes from MinIO
exa project pipelines <name> The Prefect (training) + Ray Serve (serving) surfaces
exa pipeline run --project <name> Train a project model into its storage prefix