Skip to content

Registry Dashboard

The registry dashboard is the UI for promoting models and approving changes right inside the notebook - so promotion doesn't have to be code. It surfaces the same names / aliases / tags / audit / lineage that the SDK and CLI drive, backed by the identical audited routes.

It describes the store your cells write to

The dashboard is available in both deployment modes. With STRATA_NOTEBOOK_REMOTE_STORE_URL set it describes that store rather than the local one - which is the point: a cell's strata.put(name=...) and every promotion land there, so reading the local store would show an empty registry on exactly the deployment where the registry matters. Everything the tab does goes to the same place: promotions, tags, the lineage view, and approvals - so a protected alias filed from one person's server can be approved from another's.

The forwarding happens on the server, not in the page, because the credentials do: notebook_remote_store_headers is the trusted-proxy identity the server holds. See Service Mode → shared research store.

This page is a step-by-step walkthrough. For the concepts (what a name, alias, tag, or approval gate is), see Artifacts & Model Registry.

What you'll build

Publish a model from a cell, promote it to champion, watch the move land in the registry with full audit + lineage - without leaving the notebook.

1. Open the notebook UI

Start the server and open it in a browser:

strata-notebook           # serves the bundled UI on http://localhost:8765

Open http://localhost:8765 and open (or create) a notebook.

2. Publish a model from a cell

The dashboard surfaces what your cells publish to the registry. Inside any Python cell, use the ambient strata client. It's already in the namespace, no import or setup:

# ... you've trained `model` in this or an upstream cell ...
art = strata.put(
    inputs=[],
    transform={"ref": "train-tip-model@v1"},
    data={"coef": model.coef_.tolist(), "intercept": [float(model.intercept_)]},
    name="taxi/tip-model",
)

data is written as Arrow, so pass a dict, a pa.Table, a pandas or polars DataFrame, or Arrow IPC bytes - the numbers that define the model, not the fitted estimator object.

Run the cell (Shift+Enter). strata.materialize(..., name="taxi/tip-model") works the same way. The artifact lands in the registry and is stamped with the cell that produced it.

3. The per-cell promote strip

A compact strip appears right below the cell that published:

⬡ taxi/tip-model  v1   [Promote ▾]   ⎘
  • ⬡ taxi/tip-model v1 - the name and version your cell put.
  • [Promote ▾] - promote this artifact (next step).
  • ⎘ - open its lineage.

Promote where you trained the model, without scrolling anywhere.

A cell does not have to publish with a name to share a result. With a team store configured (STRATA_NOTEBOOK_REMOTE_STORE_URL), any cell with stored outputs gets a ⇪ Promote to team… link on its strip. Pick the output, give it a team name, and Promote copies it and everything behind it to the team store, the same as strata artifact promote. The promoted result then shows on the strip like one the cell published itself.

A published cell with the promote strip along its bottom edge: the taxi/tip-model v1 chip, a champion alias chip, the Promote menu and the lineage button. A published cell with the promote strip along its bottom edge: the taxi/tip-model v1 chip, a champion alias chip, the Promote menu and the lineage button.

4. Promote to champion

Click [Promote ▾] and choose Set as champion (or Set as candidate). A toast confirms the result:

  • ✓ taxi/tip-model → champion - applied immediately (the normal case).
  • ⏳ champion change pending approval - the alias is protected (see step 6); the move is queued instead of applied.

champion / candidate are intent pointers - the post-stages model, not Staging/Production enums. A name can hold both at once (champion and a challenger candidate).

5. The Registry tab

Open the bottom drawer and click the Registry tab. It has three parts, top to bottom:

  1. Pending-approval banner - appears only when a protected-alias move is queued, with Approve / Reject buttons (the human gate, in the UI).
  2. Names table - every registry name, each row showing its alias chips (★champion=v1, candidate=v2), latest version, tags, the cells that read it (Read by), a [Promote ▾] menu, and a ⎘ lineage button. This is the same data on the per-cell strip, but for all names - not just what the current notebook published.
  3. Audit timeline (collapsible) - every name / alias / tag mutation, newest first, with who and from → to.

Read by lists the cells that read a name with # @dataset, each from a result it stored. Cells of the open notebook show by name as soon as they have run. A cell in another notebook shows as <notebook id>/<cell id> once one of its results reaches the store the tab describes, by promotion or publication; a result that stays in its own notebook is not visible from anywhere else.

The Registry tab in the bottom drawer: a names table with the champion alias chip, an rmse tag and a Promote menu, above an expanded audit timeline of every name, alias and tag change. The Registry tab in the bottom drawer: a names table with the champion alias chip, an rmse tag and a Promote menu, above an expanded audit timeline of every name, alias and tag change.

(No pending-approval banner here - nothing is queued. Step 6 covers what it looks like when a protected alias holds a move back.)

6. Approval gates (protected aliases)

To require a human for sensitive promotions, mark aliases protected when you start the server:

STRATA_REGISTRY_PROTECTED_ALIASES=champion,production strata-notebook

Now a promote to champion (step 4) returns ⏳ pending instead of applying. The pending banner appears at the top of the Registry tab; click Approve to apply it (the approver becomes the audit actor) or Reject to discard it. Unprotected aliases (candidate) still apply immediately.

7. View lineage

Click the ⎘ button on the strip or any names-table row. The lineage view renders the provenance chain:

model ← features ← scan ← table @ snapshot

This is the same chain strata artifact lineage prints on the CLI, as an interactive view. It answers "which snapshot trained this model?" in one click.

The lineage view for taxi/tip-model: the training transform above the feature set, the scan below it, and the Iceberg table at the root. The lineage view for taxi/tip-model: the training transform above the feature set, the scan below it, and the Iceberg table at the root.

Troubleshooting

Symptom Why
The registry looks empty on a machine that has published things Check STRATA_NOTEBOOK_REMOTE_STORE_URL. The dashboard describes the store the cells write to, so an unset URL on a server whose cells publish elsewhere - or a set one pointing at a store nobody publishes to - shows an empty registry rather than an error.
The tab reports a 502 The team store could not be reached. That is deliberately not an empty registry: "nobody has published anything" and "I could not ask" are different facts.
Registry tab is empty Nothing's been published with a name yet. Run a cell with strata.put(inputs=[], transform=…, data=…, name="…").
A cell ran but no per-cell strip The strip only shows artifacts a cell published via the ambient strata client (a named put / materialize). Artifacts created another way still appear in the names table, just not as a per-cell strip.
Promote said "pending" unexpectedly That alias is in STRATA_REGISTRY_PROTECTED_ALIASES - approve it from the pending banner.
Promote did nothing Setting an alias to the version it already points at is an idempotent no-op (the toast ends in (no change)).

See also