datalens-dataset-wizard

DataLens Dataset and Wizard

Work from the exact dashboard project/subproject. A description-only edit needs a narrow patch and saved readback. Field, formula, filter or source changes also need applicable field validation and bounded data preview; JSON alone cannot prove metric meaning. Use Dataset/Wizard contracts for typed drafts, field roles, preview and revision-safe updates. Use current field GUIDs; ambiguous labels cannot identify fields. Preserve dataset-global versus chart-local fields, calculation levels, field order and the existing object's technology. Choose technology and source using the project-first order; a JavaScript request alone does not exclude an existing Dataset.

For a new metric or upstream-dependent field, first follow current source readiness. A merged change or green CI does not prove that this connection exposes the field or data. Complete independent authorized edits while naming the dependent remainder.

New typed Wizard drafts resolve the same profile as recipes. Pass wizard.project_root for local overrides and wizard.presentation only for explicit exceptions; wizard.family retains a chosen family. Readback dataset_fields on the draft enables complete local SDK payload validation before create. Labels are actual measure roles, grids default off (including y2), and titles belong in Dashboard widgets. Supported grid overrides need no justification. Default palette collisions resolve deterministically; larger series sets receive an advisory. Use measure_colors keyed by GUID for meaningful series roles; preserve those identities across filters/periods. Default single-series bars use one color and category labels. Chart descriptions are not widget hints.

A new dataset draft keeps object_type, name and client_ref at its root; connection_id, source and fields belong inside dataset. Use typed public operations; the backend compiles against its pinned SDK. Measure Names and Measure Values are chart technical fields, not physical columns.

Follow authorized scope and delivery. Load visualization decisions only for the relevant semantic choice. Preview establishes bounded data evidence; provider acceptance and rendered results require their own checks.

Route placement to datalens-dashboard, JavaScript variants to datalens-editor, metadata or RLS identity lookup to datalens-inspect, and backup/cleanup to datalens-maintenance.

For unknown effects follow repeat-effect recovery: only proven non-application or separate informed repeat authorization permits another attempt. Failed reconciliation, absent inventory and a new operation_id do not make a retry safe.

Treat write_verified, dataset_validation, bounded data preview and Browser rendering as separate evidence. An invalid saved Dataset is not task success; metadata-only repairs may preserve pre-existing invalid state, but newly changed fields must pass prospective title/GUID validation before a batch writes. If a required upstream field is not exposed, finish independent metadata work and mark the dependent part blocked; do not wait for a merge or substitute a source.